chore(main): release 0.6.0 (#786)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

.gitignore

@@ -7,6 +7,7 @@ dist/
 .openclaude-profile.json
 reports/
 GEMINI.md
+CLAUDE.md
 package-lock.json
 /.claude
 coverage/

.release-please-manifest.json

@@ -1,3 +1,3 @@
 {
-  ".": "0.5.1"
+  ".": "0.6.0"
 }

CHANGELOG.md

@@ -1,5 +1,40 @@
 # Changelog
 
+## [0.6.0](https://github.com/Gitlawb/openclaude/compare/v0.5.2...v0.6.0) (2026-04-22)
+
+
+### Features
+
+* add model caching and benchmarking utilities ([#671](https://github.com/Gitlawb/openclaude/issues/671)) ([2b15e16](https://github.com/Gitlawb/openclaude/commit/2b15e16421f793f954a92c53933a07094544b29d))
+* add thinking token extraction ([#798](https://github.com/Gitlawb/openclaude/issues/798)) ([268c039](https://github.com/Gitlawb/openclaude/commit/268c0398e4bf1ab898069c61500a2b3c226a0322))
+* **api:** compress old tool_result content for small-context providers ([#801](https://github.com/Gitlawb/openclaude/issues/801)) ([a6a3de5](https://github.com/Gitlawb/openclaude/commit/a6a3de5ac155fe9d00befbfcab98d439314effd8))
+* **api:** improve local provider reliability with readiness and self-healing ([#738](https://github.com/Gitlawb/openclaude/issues/738)) ([4cb963e](https://github.com/Gitlawb/openclaude/commit/4cb963e660dbd6ee438c04042700db05a9d32c59))
+* **api:** smart model routing primitive (cheap-for-simple, strong-for-hard) ([#785](https://github.com/Gitlawb/openclaude/issues/785)) ([e908864](https://github.com/Gitlawb/openclaude/commit/e908864da7e7c987a98053ac5d18d702e192db2b))
+* enable 15 additional feature flags in open build ([#667](https://github.com/Gitlawb/openclaude/issues/667)) ([6a62e3f](https://github.com/Gitlawb/openclaude/commit/6a62e3ff76ba9ba446b8e20cf2bb139ee76a9387))
+* native Anthropic API mode for Claude models on GitHub Copilot ([#579](https://github.com/Gitlawb/openclaude/issues/579)) ([fdef4a1](https://github.com/Gitlawb/openclaude/commit/fdef4a1b4ce218ded4937ca83b30acce7c726472))
+* **provider:** expose Atomic Chat in /provider picker with autodetect ([#810](https://github.com/Gitlawb/openclaude/issues/810)) ([ee19159](https://github.com/Gitlawb/openclaude/commit/ee19159c17b3de3b4a8b4a4541a6569f4261d54e))
+* **provider:** zero-config autodetection primitive ([#784](https://github.com/Gitlawb/openclaude/issues/784)) ([a5bfcbb](https://github.com/Gitlawb/openclaude/commit/a5bfcbbadf8e9a1fd42f3e103d295524b8da64b0))
+
+
+### Bug Fixes
+
+* **api:** ensure strict role sequence and filter empty assistant messages after interruption ([#745](https://github.com/Gitlawb/openclaude/issues/745) regression) ([#794](https://github.com/Gitlawb/openclaude/issues/794)) ([06e7684](https://github.com/Gitlawb/openclaude/commit/06e7684eb56df8e694ac784575e163641931c44c))
+* Collapse all-text arrays to string for DeepSeek compatibility ([#806](https://github.com/Gitlawb/openclaude/issues/806)) ([761924d](https://github.com/Gitlawb/openclaude/commit/761924daa7e225fe8acf41651408c7cae639a511))
+* **model:** codex/nvidia-nim/minimax now read OPENAI_MODEL env ([#815](https://github.com/Gitlawb/openclaude/issues/815)) ([4581208](https://github.com/Gitlawb/openclaude/commit/458120889f6ce54cc9f0b287461d5e38eae48a20))
+* **provider:** saved profile ignored when stale CLAUDE_CODE_USE_* in shell ([#807](https://github.com/Gitlawb/openclaude/issues/807)) ([13de4e8](https://github.com/Gitlawb/openclaude/commit/13de4e85df7f5fadc8cd15a76076374dc112360b))
+* rename .claude.json to .openclaude.json with legacy fallback ([#582](https://github.com/Gitlawb/openclaude/issues/582)) ([4d4fb28](https://github.com/Gitlawb/openclaude/commit/4d4fb2880e4d0e3a62d8715e1ec13d932e736279))
+* replace discontinued gemini-2.5-pro-preview-03-25 with stable gemini-2.5-pro ([#802](https://github.com/Gitlawb/openclaude/issues/802)) ([64582c1](https://github.com/Gitlawb/openclaude/commit/64582c119d5d0278195271379da4a68d59a89c1f)), closes [#398](https://github.com/Gitlawb/openclaude/issues/398)
+* **security:** harden project settings trust boundary + MCP sanitization ([#789](https://github.com/Gitlawb/openclaude/issues/789)) ([ae3b723](https://github.com/Gitlawb/openclaude/commit/ae3b723f3b297b49925cada4728f3174aee8bf12))
+* **test:** autoCompact floor assertion is flag-sensitive ([#816](https://github.com/Gitlawb/openclaude/issues/816)) ([c13842e](https://github.com/Gitlawb/openclaude/commit/c13842e91c7227246520955de6ae0636b30def9a))
+* **ui:** prevent provider manager lag by deferring sync I/O ([#803](https://github.com/Gitlawb/openclaude/issues/803)) ([85eab27](https://github.com/Gitlawb/openclaude/commit/85eab2751e7d351bb0ed6a3fe0e15461d241c9cb))
+
+## [0.5.2](https://github.com/Gitlawb/openclaude/compare/v0.5.1...v0.5.2) (2026-04-20)
+
+
+### Bug Fixes
+
+* **api:** replace phrase-based reasoning sanitizer with tag-based filter ([#779](https://github.com/Gitlawb/openclaude/issues/779)) ([336ddcc](https://github.com/Gitlawb/openclaude/commit/336ddcc50d59d79ebff50993f2673652aecb0d7d))
+
 ## [0.5.1](https://github.com/Gitlawb/openclaude/compare/v0.5.0...v0.5.1) (2026-04-20)
 
 

README.md

@@ -125,7 +125,7 @@ Advanced and source-build guides:
 | Codex OAuth | `/provider` | Opens ChatGPT sign-in in your browser and stores Codex credentials securely |
 | Codex | `/provider` | Uses existing Codex CLI auth, OpenClaude secure storage, or env credentials |
 | Ollama | `/provider`, env vars, or `ollama launch` | Local inference with no API key |
-| Atomic Chat | advanced setup | Local Apple Silicon backend |
+| Atomic Chat | `/provider`, env vars, or `bun run dev:atomic-chat` | Local Model Provider; auto-detects loaded models |
 | Bedrock / Vertex / Foundry | env vars | Additional provider integrations for supported environments |
 
 ## What Works

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gitlawb/openclaude",
-  "version": "0.5.1",
+  "version": "0.6.0",
   "description": "Claude Code opened to any LLM — OpenAI, Gemini, DeepSeek, Ollama, and 200+ models",
   "type": "module",
   "bin": {

scripts/build.ts

@@ -19,30 +19,46 @@ const version = pkg.version
 // Most Anthropic-internal features stay off; open-build features can be
 // selectively enabled here when their full source exists in the mirror.
 const featureFlags: Record<string, boolean> = {
-  VOICE_MODE: false,
-  PROACTIVE: false,
-  KAIROS: false,
-  BRIDGE_MODE: false,
-  DAEMON: false,
-  AGENT_TRIGGERS: false,
-  MONITOR_TOOL: true,
-  ABLATION_BASELINE: false,
-  DUMP_SYSTEM_PROMPT: false,
-  CACHED_MICROCOMPACT: false,
-  COORDINATOR_MODE: true,
-  BUILTIN_EXPLORE_PLAN_AGENTS: true,
-  CONTEXT_COLLAPSE: false,
-  COMMIT_ATTRIBUTION: false,
-  TEAMMEM: true,
-  UDS_INBOX: false,
-  BG_SESSIONS: false,
-  AWAY_SUMMARY: false,
-  TRANSCRIPT_CLASSIFIER: false,
-  WEB_BROWSER_TOOL: false,
-  MESSAGE_ACTIONS: true,
-  BUDDY: true,
-  CHICAGO_MCP: false,
-  COWORKER_TYPE_TELEMETRY: false,
+  // ── Disabled: require Anthropic infrastructure or missing source ─────
+  VOICE_MODE: false,              // Push-to-talk STT via claude.ai OAuth endpoint
+  PROACTIVE: false,               // Autonomous agent mode (missing proactive/ module)
+  KAIROS: false,                  // Persistent assistant/session mode (cloud backend)
+  BRIDGE_MODE: false,             // Remote desktop bridge via CCR infrastructure
+  DAEMON: false,                  // Background daemon process (stubbed in open build)
+  AGENT_TRIGGERS: false,          // Scheduled remote agent triggers
+  ABLATION_BASELINE: false,       // A/B testing harness for eval experiments
+  CONTEXT_COLLAPSE: false,        // Context collapsing optimization (stubbed)
+  COMMIT_ATTRIBUTION: false,      // Co-Authored-By metadata in git commits
+  UDS_INBOX: false,               // Unix Domain Socket inter-session messaging
+  BG_SESSIONS: false,             // Background sessions via tmux (stubbed)
+  WEB_BROWSER_TOOL: false,        // Built-in browser automation (source not mirrored)
+  CHICAGO_MCP: false,             // Computer-use MCP (native Swift modules stubbed)
+  COWORKER_TYPE_TELEMETRY: false, // Telemetry for agent/coworker type classification
+
+  // ── Enabled: upstream defaults ──────────────────────────────────────
+  COORDINATOR_MODE: true,             // Multi-agent coordinator with worker delegation
+  BUILTIN_EXPLORE_PLAN_AGENTS: true,  // Built-in Explore/Plan specialized subagents
+  BUDDY: true,                        // Buddy mode for paired programming
+  MONITOR_TOOL: true,                 // MCP server monitoring/streaming tool
+  TEAMMEM: true,                      // Team memory management
+  MESSAGE_ACTIONS: true,              // Message action buttons in the UI
+
+  // ── Enabled: new activations ────────────────────────────────────────
+  DUMP_SYSTEM_PROMPT: true,           // --dump-system-prompt CLI flag for debugging
+  CACHED_MICROCOMPACT: true,          // Cache-aware tool result truncation optimization
+  AWAY_SUMMARY: true,                 // "While you were away" recap after 5min blur
+  TRANSCRIPT_CLASSIFIER: true,        // Auto-approval classifier for safe tool uses
+  ULTRATHINK: true,                   // Deep thinking mode — type "ultrathink" to boost reasoning
+  TOKEN_BUDGET: true,                 // Token budget tracking with usage warnings
+  HISTORY_PICKER: true,               // Enhanced interactive prompt history picker
+  QUICK_SEARCH: true,                 // Ctrl+G quick search across prompts
+  SHOT_STATS: true,                   // Shot distribution stats in session summary
+  EXTRACT_MEMORIES: true,             // Auto-extract durable memories from conversations
+  FORK_SUBAGENT: true,                // Implicit context-forking when omitting subagent_type
+  VERIFICATION_AGENT: true,           // Built-in read-only agent for test/verification
+  MCP_SKILLS: true,                   // Discover skills dynamically from MCP server resources
+  PROMPT_CACHE_BREAK_DETECTION: true, // Detect & log unexpected prompt cache invalidations
+  HOOK_PROMPTS: true,                 // Allow tools to request interactive user prompts
 }
 
 // ── Pre-process: replace feature() calls with boolean literals ──────

scripts/no-telemetry-growthbook-stub.test.ts

@@ -50,6 +50,23 @@ describe('growthbook stub — local feature flag overrides', () => {
     expect(stub.getAllGrowthBookFeatures()).toEqual({})
   })
 
+  // ── Open-build defaults (_openBuildDefaults) ────────────────────
+
+  test('returns open-build default when flags file is absent', () => {
+    // tengu_passport_quail is in _openBuildDefaults as true; without a
+    // flags file the stub should return the open-build override, not
+    // the call-site defaultValue.
+    expect(stub.getFeatureValue_CACHED_MAY_BE_STALE('tengu_passport_quail', false)).toBe(true)
+    expect(stub.getFeatureValue_CACHED_MAY_BE_STALE('tengu_coral_fern', false)).toBe(true)
+  })
+
+  test('flags file overrides open-build defaults', () => {
+    // User-provided feature-flags.json takes priority over _openBuildDefaults.
+    writeFileSync(flagsFile, JSON.stringify({ tengu_passport_quail: false }))
+
+    expect(stub.getFeatureValue_CACHED_MAY_BE_STALE('tengu_passport_quail', true)).toBe(false)
+  })
+
   // ── Valid JSON object ────────────────────────────────────────────
 
   test('loads and returns values from a valid JSON file', () => {

scripts/no-telemetry-plugin.ts

@@ -40,6 +40,151 @@ import _os from 'node:os';
 
 let _flags = undefined;
 
+// ── Open-build GrowthBook overrides ───────────────────────────────────
+// Override upstream defaultValue for runtime gates tied to build-time
+// features. Only keys that DIFFER from upstream belong here — the
+// catalog below is pure documentation and does NOT affect resolution.
+//
+// Priority: ~/.claude/feature-flags.json > _openBuildDefaults > defaultValue
+//
+// To override at runtime, create ~/.claude/feature-flags.json:
+//   { "tengu_some_flag": true }
+const _openBuildDefaults = {
+  'tengu_sedge_lantern': true,  // AWAY_SUMMARY — "while you were away" recap (upstream: false)
+  'tengu_hive_evidence': true,  // VERIFICATION_AGENT — read-only test/verification agent (upstream: false)
+  'tengu_passport_quail': true, // EXTRACT_MEMORIES — enable memory extraction (upstream: false)
+  'tengu_coral_fern': true,     // EXTRACT_MEMORIES — enable memory search in past context (upstream: false)
+};
+
+/* ── Known runtime feature keys (reference only) ───────────────────────
+ * This catalog does NOT participate in flag resolution. It documents
+ * the known GrowthBook keys and their upstream default values, scraped
+ * from src/ call sites. It is NOT exhaustive — new keys may be added
+ * upstream between catalog updates.
+ *
+ * Some keys have different defaults at different call sites — this is
+ * intentional upstream (the server unifies the value at runtime).
+ *
+ * To activate any of these, add them to ~/.claude/feature-flags.json
+ * or to _openBuildDefaults above.
+ *
+ * ── Reasoning & thinking ──────────────────────────────────────────────
+ *   tengu_turtle_carbon            = true       ULTRATHINK deep thinking runtime gate
+ *   tengu_thinkback                = gate       /thinkback replay command
+ *
+ * ── Agents & orchestration ────────────────────────────────────────────
+ *   tengu_amber_flint              = true       Agent swarms coordination
+ *   tengu_amber_stoat              = true       Built-in agent availability (Explore, Plan, etc.)
+ *   tengu_agent_list_attach        = true       Attach file context to agent list
+ *   tengu_auto_background_agents   = false      Auto-spawn background agents
+ *   tengu_slim_subagent_claudemd   = true       Lighter ClaudeMD for subagents
+ *   tengu_hive_evidence            = false      Verification agent / evidence tracking (4 call sites)
+ *   tengu_ultraplan_model          = model cfg  ULTRAPLAN model selection (dynamic config)
+ *
+ * ── Memory & context ──────────────────────────────────────────────────
+ *   tengu_passport_quail           = false      EXTRACT_MEMORIES main gate (isExtractModeActive)
+ *   tengu_coral_fern               = false      EXTRACT_MEMORIES search in past context
+ *   tengu_slate_thimble            = false      Memory dir paths (non-interactive sessions)
+ *   tengu_herring_clock            = true/false Team memory paths (varies by call site)
+ *   tengu_bramble_lintel           = null       Extract memories throttle (null → every turn)
+ *   tengu_sedge_lantern            = false      AWAY_SUMMARY "while you were away" recap
+ *   tengu_session_memory           = false      Session memory service
+ *   tengu_sm_config                = {}         Session memory config (dynamic)
+ *   tengu_sm_compact_config        = {}         Session memory compaction config (dynamic)
+ *   tengu_cobalt_raccoon           = false      Reactive compaction (suppress auto-compact)
+ *   tengu_pebble_leaf_prune        = false      Session storage pruning
+ *
+ * ── Kairos & cron ─────────────────────────────────────────────────────
+ *   tengu_kairos_brief             = false      Brief layout mode (KAIROS)
+ *   tengu_kairos_brief_config      = {}         Brief config (dynamic)
+ *   tengu_kairos_cron              = true       Cron scheduler enable
+ *   tengu_kairos_cron_durable      = true       Durable (disk-persistent) cron tasks
+ *   tengu_kairos_cron_config       = {}         Cron jitter config (dynamic)
+ *
+ * ── Bridge & remote (require Anthropic infra) ─────────────────────────
+ *   tengu_ccr_bridge               = false      CCR bridge connection
+ *   tengu_ccr_bridge_multi_session = gate       Multi-session spawn mode
+ *   tengu_ccr_mirror               = false      CCR session mirroring
+ *   tengu_ccr_bundle_seed_enabled  = gate       Git bundle seeding for CCR
+ *   tengu_ccr_bundle_max_bytes     = null       Bundle size limit (null → default)
+ *   tengu_bridge_repl_v2           = false      Environment-less REPL bridge v2
+ *   tengu_bridge_repl_v2_cse_shim_enabled = true CSE→Session tag retag shim
+ *   tengu_bridge_min_version       = {min:'0'}  Min CLI version for bridge (dynamic)
+ *   tengu_bridge_initial_history_cap = 200      Initial history cap for bridge
+ *   tengu_bridge_system_init       = false      Bridge system initialization
+ *   tengu_cobalt_harbor            = false      Auto-connect CCR at startup
+ *   tengu_cobalt_lantern           = false      Remote setup preconditions
+ *   tengu_remote_backend           = false      Remote TUI backend
+ *   tengu_surreal_dali             = false      Remote agent tasks / triggers
+ *
+ * ── Prompt & API ──────────────────────────────────────────────────────
+ *   tengu_attribution_header       = true       Attribution header in API requests
+ *   tengu_basalt_3kr               = true       MCP instructions delta
+ *   tengu_slate_prism              = true/false Message formatting (varies by call site)
+ *   tengu_amber_prism              = false      Message content formatting
+ *   tengu_amber_json_tools         = false      JSON format for tool schemas
+ *   tengu_fgts                     = false      API feature gates
+ *   tengu_otk_slot_v1              = false      One-time key slots for API auth
+ *   tengu_cicada_nap_ms            = 0          Background GrowthBook refresh throttle (ms)
+ *   tengu_miraculo_the_bard        = false      Service initialization gate
+ *   tengu_immediate_model_command  = false      Immediate /model command execution
+ *   tengu_chomp_inflection         = false      Prompt suggestions after responses
+ *   tengu_tool_pear                = gate       API betas for tool use
+ *   tengu-off-switch               = {act:false} Service kill switch (dynamic; uses dash)
+ *
+ * ── Permissions & security ────────────────────────────────────────────
+ *   tengu_birch_trellis            = true       Bash auto-mode permissions config
+ *   tengu_auto_mode_config         = {}         Auto-mode configuration (dynamic, many call sites)
+ *   tengu_iron_gate_closed         = true       Permission iron gate (with refresh)
+ *   tengu_destructive_command_warning = false    Warning for destructive bash commands
+ *   tengu_disable_bypass_permissions_mode = security Security killswitch (always false in open build)
+ *
+ * ── UI & UX ───────────────────────────────────────────────────────────
+ *   tengu_willow_mode              = 'off'      REPL rendering mode
+ *   tengu_terminal_panel           = false      Terminal panel keybinding
+ *   tengu_terminal_sidebar         = false      Terminal sidebar in REPL/config
+ *   tengu_marble_sandcastle        = false      Fast mode gate
+ *   tengu_jade_anvil_4             = false      Rate limit options UI ordering
+ *   tengu_collage_kaleidoscope     = true       Native clipboard image paste (macOS)
+ *   tengu_lapis_finch              = false      Plugin/hint recommendation
+ *   tengu_lodestone_enabled        = false      Deep links claude-cli:// protocol
+ *   tengu_copper_panda             = false      Skill improvement suggestions
+ *   tengu_desktop_upsell           = {}         Desktop app upsell config (dynamic)
+ *   tengu-top-of-feed-tip          = {}         Emergency tip of feed (dynamic; uses dash)
+ *
+ * ── File operations ───────────────────────────────────────────────────
+ *   tengu_quartz_lantern           = false      File read/write dedup optimization
+ *   tengu_moth_copse               = false      Attachments handling (variant A)
+ *   tengu_marble_fox               = false      Attachments handling (variant B)
+ *   tengu_scratch                  = gate       Scratchpad filesystem access / coordinator
+ *
+ * ── MCP & plugins ─────────────────────────────────────────────────────
+ *   tengu_harbor                   = false      MCP channel allowlist verification
+ *   tengu_harbor_permissions       = false      MCP channel permissions enforcement
+ *   tengu_copper_bridge            = false      Chrome MCP bridge
+ *   tengu_chrome_auto_enable       = false      Auto-enable Chrome MCP on startup
+ *   tengu_glacier_2xr              = false      Enhanced tool search / ToolSearchTool
+ *   tengu_malort_pedway            = {}         Computer-use (Chicago) config (dynamic)
+ *
+ * ── VSCode / IDE ──────────────────────────────────────────────────────
+ *   tengu_quiet_fern               = false      VSCode browser support
+ *   tengu_vscode_cc_auth           = false      VSCode in-band OAuth via claude_authenticate
+ *   tengu_vscode_review_upsell     = gate       VSCode review upsell
+ *   tengu_vscode_onboarding        = gate       VSCode onboarding experience
+ *
+ * ── Voice ─────────────────────────────────────────────────────────────
+ *   tengu_amber_quartz_disabled    = false      VOICE_MODE kill-switch (false = voice allowed)
+ *
+ * ── Auto-updater (stubbed in open build) ──────────────────────────────
+ *   tengu_version_config           = {min:'0'}  Min version enforcement (dynamic)
+ *   tengu_max_version_config       = {}         Max version / deprecation config (dynamic)
+ *
+ * ── Telemetry & tracing ───────────────────────────────────────────────
+ *   tengu_trace_lantern            = false      Beta session tracing
+ *   tengu_chair_sermon             = gate       Analytics / message formatting gate
+ *   tengu_strap_foyer              = false      Settings sync to cloud
+ */
+
 function _loadFlags() {
   if (_flags !== undefined) return;
   try {
@@ -55,6 +200,7 @@ function _loadFlags() {
 function _getFlagValue(key, defaultValue) {
   _loadFlags();
   if (_flags != null && Object.hasOwn(_flags, key)) return _flags[key];
+  if (Object.hasOwn(_openBuildDefaults, key)) return _openBuildDefaults[key];
   return defaultValue;
 }
 

scripts/system-check.test.ts

@@ -20,6 +20,23 @@ describe('formatReachabilityFailureDetail', () => {
     )
   })
 
+  test('redacts credentials and sensitive query parameters in endpoint details', () => {
+    const detail = formatReachabilityFailureDetail(
+      'http://user:pass@localhost:11434/v1/models?token=abc123&mode=test',
+      502,
+      'bad gateway',
+      {
+        transport: 'chat_completions',
+        requestedModel: 'llama3.1:8b',
+        resolvedModel: 'llama3.1:8b',
+      },
+    )
+
+    expect(detail).toBe(
+      'Unexpected status 502 from http://redacted:redacted@localhost:11434/v1/models?token=redacted&mode=test. Body: bad gateway',
+    )
+  })
+
   test('adds alias/entitlement hint for codex model support 400s', () => {
     const detail = formatReachabilityFailureDetail(
       'https://chatgpt.com/backend-api/codex/responses',

scripts/system-check.ts

@@ -7,6 +7,11 @@ import {
   resolveProviderRequest,
   isLocalProviderUrl as isProviderLocalUrl,
 } from '../src/services/api/providerConfig.js'
+import {
+  getLocalOpenAICompatibleProviderLabel,
+  probeOllamaGenerationReadiness,
+} from '../src/utils/providerDiscovery.js'
+import { redactUrlForDisplay } from '../src/utils/urlRedaction.js'
 
 type CheckResult = {
   ok: boolean
@@ -69,7 +74,7 @@ export function formatReachabilityFailureDetail(
   },
 ): string {
   const compactBody = responseBody.trim().replace(/\s+/g, ' ').slice(0, 240)
-  const base = `Unexpected status ${status} from ${endpoint}.`
+  const base = `Unexpected status ${status} from ${redactUrlForDisplay(endpoint)}.`
   const bodySuffix = compactBody ? ` Body: ${compactBody}` : ''
 
   if (request.transport !== 'codex_responses' || status !== 400) {
@@ -255,7 +260,7 @@ function checkOpenAIEnv(): CheckResult[] {
     results.push(pass('OPENAI_MODEL', process.env.OPENAI_MODEL))
   }
 
-  results.push(pass('OPENAI_BASE_URL', request.baseUrl))
+  results.push(pass('OPENAI_BASE_URL', redactUrlForDisplay(request.baseUrl)))
 
   if (request.transport === 'codex_responses') {
     const credentials = resolveCodexApiCredentials(process.env)
@@ -308,7 +313,7 @@ async function checkBaseUrlReachability(): Promise<CheckResult> {
     return pass('Provider reachability', 'Skipped (OpenAI-compatible mode disabled).')
   }
 
-  if (useGithub) {
+  if (useGithub && !useOpenAI) {
     return pass(
       'Provider reachability',
       'Skipped for GitHub Models (inference endpoint differs from OpenAI /models probe).',
@@ -326,6 +331,7 @@ async function checkBaseUrlReachability(): Promise<CheckResult> {
   const endpoint = request.transport === 'codex_responses'
     ? `${request.baseUrl}/responses`
     : `${request.baseUrl}/models`
+  const redactedEndpoint = redactUrlForDisplay(endpoint)
 
   const controller = new AbortController()
   const timeout = setTimeout(() => controller.abort(), 4000)
@@ -375,7 +381,10 @@ async function checkBaseUrlReachability(): Promise<CheckResult> {
     })
 
     if (response.status === 200 || response.status === 401 || response.status === 403) {
-      return pass('Provider reachability', `Reached ${endpoint} (status ${response.status}).`)
+      return pass(
+        'Provider reachability',
+        `Reached ${redactedEndpoint} (status ${response.status}).`,
+      )
     }
 
     const responseBody = await response.text().catch(() => '')
@@ -391,12 +400,100 @@ async function checkBaseUrlReachability(): Promise<CheckResult> {
     )
   } catch (error) {
     const message = error instanceof Error ? error.message : String(error)
-    return fail('Provider reachability', `Failed to reach ${endpoint}: ${message}`)
+    return fail(
+      'Provider reachability',
+      `Failed to reach ${redactedEndpoint}: ${message}`,
+    )
   } finally {
     clearTimeout(timeout)
   }
 }
 
+async function checkProviderGenerationReadiness(): Promise<CheckResult> {
+  const useGemini = isTruthy(process.env.CLAUDE_CODE_USE_GEMINI)
+  const useOpenAI = isTruthy(process.env.CLAUDE_CODE_USE_OPENAI)
+  const useGithub = isTruthy(process.env.CLAUDE_CODE_USE_GITHUB)
+  const useMistral = isTruthy(process.env.CLAUDE_CODE_USE_MISTRAL)
+
+  if (!useGemini && !useOpenAI && !useGithub && !useMistral) {
+    return pass('Provider generation readiness', 'Skipped (OpenAI-compatible mode disabled).')
+  }
+
+  if (useGithub && !useOpenAI) {
+    return pass(
+      'Provider generation readiness',
+      'Skipped for GitHub Models (runtime generation uses a different endpoint flow).',
+    )
+  }
+
+  if (useGemini || useMistral) {
+    return pass(
+      'Provider generation readiness',
+      'Skipped for managed provider mode.',
+    )
+  }
+
+  if (!useOpenAI) {
+    return pass('Provider generation readiness', 'Skipped (OpenAI-compatible mode disabled).')
+  }
+
+  const request = resolveProviderRequest({
+    model: process.env.OPENAI_MODEL,
+    baseUrl: process.env.OPENAI_BASE_URL,
+  })
+
+  if (request.transport === 'codex_responses') {
+    return pass(
+      'Provider generation readiness',
+      'Skipped for Codex responses (reachability probe already performs a lightweight generation request).',
+    )
+  }
+
+  if (!isLocalBaseUrl(request.baseUrl)) {
+    return pass('Provider generation readiness', 'Skipped for non-local provider URL.')
+  }
+
+  const localProviderLabel = getLocalOpenAICompatibleProviderLabel(request.baseUrl)
+  if (localProviderLabel !== 'Ollama') {
+    return pass(
+      'Provider generation readiness',
+      `Skipped for ${localProviderLabel} (no provider-specific generation probe).`,
+    )
+  }
+
+  const readiness = await probeOllamaGenerationReadiness({
+    baseUrl: request.baseUrl,
+    model: request.requestedModel,
+  })
+
+  if (readiness.state === 'ready') {
+    return pass(
+      'Provider generation readiness',
+      `Generated a test response with ${readiness.probeModel ?? request.requestedModel}.`,
+    )
+  }
+
+  if (readiness.state === 'unreachable') {
+    return fail(
+      'Provider generation readiness',
+      `Could not reach Ollama at ${redactUrlForDisplay(request.baseUrl)}.`,
+    )
+  }
+
+  if (readiness.state === 'no_models') {
+    return fail(
+      'Provider generation readiness',
+      'Ollama is reachable, but no installed models were found. Pull a model first (for example: ollama pull qwen2.5-coder:7b).',
+    )
+  }
+
+  const detailSuffix = readiness.detail ? ` Detail: ${readiness.detail}.` : ''
+  return fail(
+    'Provider generation readiness',
+    `Ollama is reachable, but generation failed for ${readiness.probeModel ?? request.requestedModel}.${detailSuffix}`,
+  )
+}
+
 function isAtomicChatUrl(baseUrl: string): boolean {
   try {
     const parsed = new URL(baseUrl)
@@ -567,6 +664,7 @@ async function main(): Promise<void> {
   results.push(checkBuildArtifacts())
   results.push(...checkOpenAIEnv())
   results.push(await checkBaseUrlReachability())
+  results.push(await checkProviderGenerationReadiness())
   results.push(checkOllamaProcessorMode())
 
   if (!options.json) {

src/__tests__/security-hardening.test.ts

@@ -0,0 +1,191 @@
+/**
+ * Security hardening regression tests.
+ *
+ * Covers:
+ * 1. MCP tool result Unicode sanitization
+ * 2. Sandbox settings source filtering (exclude projectSettings)
+ * 3. Plugin git clone/pull hooks disabled
+ * 4. ANTHROPIC_FOUNDRY_API_KEY removed from SAFE_ENV_VARS
+ * 5. WebFetch SSRF protection via ssrfGuardedLookup
+ */
+
+import { describe, test, expect } from 'bun:test'
+import { resolve } from 'path'
+
+const SRC = resolve(import.meta.dir, '..')
+const file = (relative: string) => Bun.file(resolve(SRC, relative))
+
+// ---------------------------------------------------------------------------
+// Fix 1: MCP tool result Unicode sanitization
+// ---------------------------------------------------------------------------
+describe('MCP tool result sanitization', () => {
+  test('transformResultContent sanitizes text content', async () => {
+    const content = await file('services/mcp/client.ts').text()
+    // Tool definitions are already sanitized (line ~1798)
+    expect(content).toContain('recursivelySanitizeUnicode(result.tools)')
+    // Tool results must also be sanitized
+    expect(content).toMatch(
+      /case 'text':[\s\S]*?recursivelySanitizeUnicode\(resultContent\.text\)/,
+    )
+  })
+
+  test('resource text content is also sanitized', async () => {
+    const content = await file('services/mcp/client.ts').text()
+    expect(content).toMatch(
+      /recursivelySanitizeUnicode\(\s*`\$\{prefix\}\$\{resource\.text\}`/,
+    )
+  })
+})
+
+// ---------------------------------------------------------------------------
+// Fix 2: Sandbox settings source filtering
+// ---------------------------------------------------------------------------
+describe('Sandbox settings trust boundary', () => {
+  test('getSandboxEnabledSetting does not use getSettings_DEPRECATED', async () => {
+    const content = await file('utils/sandbox/sandbox-adapter.ts').text()
+    // Extract the getSandboxEnabledSetting function body
+    const fnMatch = content.match(
+      /function getSandboxEnabledSetting\(\)[^{]*\{([\s\S]*?)\n\}/,
+    )
+    expect(fnMatch).not.toBeNull()
+    const fnBody = fnMatch![1]
+    // Must NOT use getSettings_DEPRECATED (reads all sources including project)
+    expect(fnBody).not.toContain('getSettings_DEPRECATED')
+    // Must use getSettingsForSource for individual trusted sources
+    expect(fnBody).toContain("getSettingsForSource('userSettings')")
+    expect(fnBody).toContain("getSettingsForSource('policySettings')")
+    // Must NOT read from projectSettings
+    expect(fnBody).not.toContain("'projectSettings'")
+  })
+})
+
+// ---------------------------------------------------------------------------
+// Fix 3: Plugin git hooks disabled
+// ---------------------------------------------------------------------------
+describe('Plugin git operations disable hooks', () => {
+  test('gitClone includes core.hooksPath=/dev/null', async () => {
+    const content = await file('utils/plugins/marketplaceManager.ts').text()
+    // The clone args must disable hooks
+    const cloneSection = content.slice(
+      content.indexOf('export async function gitClone('),
+      content.indexOf('export async function gitClone(') + 2000,
+    )
+    expect(cloneSection).toContain("'core.hooksPath=/dev/null'")
+  })
+
+  test('gitPull includes core.hooksPath=/dev/null', async () => {
+    const content = await file('utils/plugins/marketplaceManager.ts').text()
+    const pullSection = content.slice(
+      content.indexOf('export async function gitPull('),
+      content.indexOf('export async function gitPull(') + 2000,
+    )
+    expect(pullSection).toContain("'core.hooksPath=/dev/null'")
+  })
+
+  test('gitSubmoduleUpdate includes core.hooksPath=/dev/null', async () => {
+    const content = await file('utils/plugins/marketplaceManager.ts').text()
+    const subSection = content.slice(
+      content.indexOf('async function gitSubmoduleUpdate('),
+      content.indexOf('async function gitSubmoduleUpdate(') + 1000,
+    )
+    expect(subSection).toContain("'core.hooksPath=/dev/null'")
+  })
+})
+
+// ---------------------------------------------------------------------------
+// Fix 4: ANTHROPIC_FOUNDRY_API_KEY not in SAFE_ENV_VARS
+// ---------------------------------------------------------------------------
+describe('SAFE_ENV_VARS excludes credentials', () => {
+  test('ANTHROPIC_FOUNDRY_API_KEY is not in SAFE_ENV_VARS', async () => {
+    const content = await file('utils/managedEnvConstants.ts').text()
+    // Extract the SAFE_ENV_VARS set definition
+    const safeStart = content.indexOf('export const SAFE_ENV_VARS')
+    const safeEnd = content.indexOf('])', safeStart)
+    const safeSection = content.slice(safeStart, safeEnd)
+    expect(safeSection).not.toContain('ANTHROPIC_FOUNDRY_API_KEY')
+  })
+})
+
+// ---------------------------------------------------------------------------
+// Fix 5: WebFetch SSRF protection
+// ---------------------------------------------------------------------------
+describe('WebFetch SSRF guard', () => {
+  test('getWithPermittedRedirects uses ssrfGuardedLookup', async () => {
+    const content = await file('tools/WebFetchTool/utils.ts').text()
+    expect(content).toContain(
+      "import { ssrfGuardedLookup } from '../../utils/hooks/ssrfGuard.js'",
+    )
+    // The axios.get call in getWithPermittedRedirects must include lookup
+    const fnSection = content.slice(
+      content.indexOf('export async function getWithPermittedRedirects('),
+      content.indexOf('export async function getWithPermittedRedirects(') +
+        1000,
+    )
+    expect(fnSection).toContain('lookup: ssrfGuardedLookup')
+  })
+})
+
+// ---------------------------------------------------------------------------
+// Fix 6: Swarm permission file polling removed (security hardening)
+// ---------------------------------------------------------------------------
+describe('Swarm permission file polling removed', () => {
+  test('useSwarmPermissionPoller hook no longer exists', async () => {
+    const content = await file(
+      'hooks/useSwarmPermissionPoller.ts',
+    ).text()
+    // The file-based polling hook must not exist — it read from an
+    // unauthenticated resolved/ directory where any local process could
+    // forge approval files.
+    expect(content).not.toContain('function useSwarmPermissionPoller(')
+    // The file-based processResponse must not exist
+    expect(content).not.toContain('function processResponse(')
+  })
+
+  test('poller does not import from permissionSync', async () => {
+    const content = await file(
+      'hooks/useSwarmPermissionPoller.ts',
+    ).text()
+    // Must not import anything from permissionSync — all file-based
+    // functions have been removed from this module's dependencies
+    expect(content).not.toContain('permissionSync')
+  })
+
+  test('file-based permission functions are marked deprecated', async () => {
+    const content = await file(
+      'utils/swarm/permissionSync.ts',
+    ).text()
+    // All file-based functions must have @deprecated JSDoc
+    const deprecatedFns = [
+      'writePermissionRequest',
+      'readPendingPermissions',
+      'readResolvedPermission',
+      'resolvePermission',
+      'pollForResponse',
+      'removeWorkerResponse',
+    ]
+    for (const fn of deprecatedFns) {
+      // Find the function and check that @deprecated appears before it
+      const fnIndex = content.indexOf(`export async function ${fn}(`)
+      if (fnIndex === -1) continue // submitPermissionRequest is a const, not async function
+      const preceding = content.slice(Math.max(0, fnIndex - 500), fnIndex)
+      expect(preceding).toContain('@deprecated')
+    }
+  })
+
+  test('mailbox-based functions are NOT deprecated', async () => {
+    const content = await file(
+      'utils/swarm/permissionSync.ts',
+    ).text()
+    // These are the active path — must not be deprecated
+    const activeFns = [
+      'sendPermissionRequestViaMailbox',
+      'sendPermissionResponseViaMailbox',
+    ]
+    for (const fn of activeFns) {
+      const fnIndex = content.indexOf(`export async function ${fn}(`)
+      expect(fnIndex).not.toBe(-1)
+      const preceding = content.slice(Math.max(0, fnIndex - 300), fnIndex)
+      expect(preceding).not.toContain('@deprecated')
+    }
+  })
+})

src/commands/benchmark.ts

@@ -0,0 +1,56 @@
+import type { ToolUseContext } from '../Tool.js'
+import type { Command } from '../types/command.js'
+import {
+  benchmarkModel,
+  benchmarkMultipleModels,
+  formatBenchmarkResults,
+  isBenchmarkSupported,
+} from '../utils/model/benchmark.js'
+import { getOllamaModelOptions } from '../utils/model/ollamaModels.js'
+
+async function runBenchmark(
+  model?: string,
+  context?: ToolUseContext,
+): Promise<void> {
+  if (!isBenchmarkSupported()) {
+    context?.stdout?.write(
+      'Benchmark not supported for this provider.\n' +
+        'Supported: OpenAI-compatible endpoints (Ollama, NVIDIA NIM, MiniMax)\n',
+    )
+    return
+  }
+
+  let modelsToBenchmark: string[]
+
+  if (model) {
+    modelsToBenchmark = [model]
+  } else {
+    const ollamaModels = getOllamaModelOptions()
+    modelsToBenchmark = ollamaModels.slice(0, 3).map((m) => m.value)
+  }
+
+  context?.stdout?.write(`Benchmarking ${modelsToBenchmark.length} model(s)...\n`)
+
+  const results = await benchmarkMultipleModels(
+    modelsToBenchmark,
+    (completed, total, result) => {
+      context?.stdout?.write(
+        `[${completed}/${total}] ${result.model}: ` +
+          `${result.success ? result.tokensPerSecond.toFixed(1) + ' tps' : 'FAILED'}\n`,
+      )
+    },
+  )
+
+  context?.stdout?.write('\n' + formatBenchmarkResults(results) + '\n')
+}
+
+export const benchmark: Command = {
+  name: 'benchmark',
+
+  async onExecute(context: ToolUseContext): Promise<void> {
+    const args = context.args ?? {}
+    const model = args.model as string | undefined
+
+    await runBenchmark(model, context)
+  },
+}

src/commands/provider/provider.tsx

@@ -66,10 +66,44 @@ import {
 import {
   getOllamaChatBaseUrl,
   getLocalOpenAICompatibleProviderLabel,
-  hasLocalOllama,
-  listOllamaModels,
+  probeOllamaGenerationReadiness,
+  type OllamaGenerationReadiness,
 } from '../../utils/providerDiscovery.js'
 
+function describeOllamaReadinessIssue(
+  readiness: OllamaGenerationReadiness,
+  options?: {
+    baseUrl?: string
+    allowManualFallback?: boolean
+  },
+): string {
+  const endpoint = options?.baseUrl ?? 'http://localhost:11434'
+
+  if (readiness.state === 'unreachable') {
+    return `Could not reach Ollama at ${endpoint}. Start Ollama first, then run /provider again.`
+  }
+
+  if (readiness.state === 'no_models') {
+    const manualSuffix = options?.allowManualFallback
+      ? ', or enter details manually'
+      : ''
+    return `Ollama is running, but no installed models were found. Pull a chat model such as qwen2.5-coder:7b or llama3.1:8b first${manualSuffix}.`
+  }
+
+  if (readiness.state === 'generation_failed') {
+    const modelHint = readiness.probeModel ?? 'the selected model'
+    const detailSuffix = readiness.detail
+      ? ` Details: ${readiness.detail}.`
+      : ''
+    const manualSuffix = options?.allowManualFallback
+      ? ' You can also enter details manually.'
+      : ''
+    return `Ollama is reachable and models are installed, but a generation probe failed for ${modelHint}.${detailSuffix} Run "ollama run ${modelHint}" once and retry.${manualSuffix}`
+  }
+
+  return ''
+}
+
 type ProviderChoice = 'auto' | ProviderProfile | 'codex-oauth' | 'clear'
 
 type Step =
@@ -715,6 +749,7 @@ function AutoRecommendationStep({
     | {
         state: 'openai'
         defaultModel: string
+        reason: string
       }
     | {
         state: 'error'
@@ -728,19 +763,27 @@ function AutoRecommendationStep({
     void (async () => {
       const defaultModel = getGoalDefaultOpenAIModel(goal)
       try {
-        const ollamaAvailable = await hasLocalOllama()
-        if (!ollamaAvailable) {
+        const readiness = await probeOllamaGenerationReadiness()
+        if (readiness.state !== 'ready') {
           if (!cancelled) {
-            setStatus({ state: 'openai', defaultModel })
+            setStatus({
+              state: 'openai',
+              defaultModel,
+              reason: describeOllamaReadinessIssue(readiness),
+            })
           }
           return
         }
 
-        const models = await listOllamaModels()
-        const recommended = recommendOllamaModel(models, goal)
+        const recommended = recommendOllamaModel(readiness.models, goal)
         if (!recommended) {
           if (!cancelled) {
-            setStatus({ state: 'openai', defaultModel })
+            setStatus({
+              state: 'openai',
+              defaultModel,
+              reason:
+                'Ollama responded to a generation probe, but no recommended chat model matched this goal.',
+            })
           }
           return
         }
@@ -796,10 +839,10 @@ function AutoRecommendationStep({
       <Dialog title="Auto setup fallback" onCancel={onCancel}>
         <Box flexDirection="column" gap={1}>
           <Text>
-            No viable local Ollama chat model was detected. Auto setup can
-            continue into OpenAI-compatible setup with a default model of{' '}
+            Auto setup can continue into OpenAI-compatible setup with a default model of{' '}
             {status.defaultModel}.
           </Text>
+          <Text dimColor>{status.reason}</Text>
           <Select
             options={[
               { label: 'Continue to OpenAI-compatible setup', value: 'continue' },
@@ -883,32 +926,19 @@ function OllamaModelStep({
     let cancelled = false
 
     void (async () => {
-      const available = await hasLocalOllama()
-      if (!available) {
+      const readiness = await probeOllamaGenerationReadiness()
+      if (readiness.state !== 'ready') {
         if (!cancelled) {
           setStatus({
             state: 'unavailable',
-            message:
-              'Could not reach Ollama at http://localhost:11434. Start Ollama first, then run /provider again.',
+            message: describeOllamaReadinessIssue(readiness),
           })
         }
         return
       }
 
-      const models = await listOllamaModels()
-      if (models.length === 0) {
-        if (!cancelled) {
-          setStatus({
-            state: 'unavailable',
-            message:
-              'Ollama is running, but no installed models were found. Pull a chat model such as qwen2.5-coder:7b or llama3.1:8b first.',
-          })
-        }
-        return
-      }
-
-      const ranked = rankOllamaModels(models, 'balanced')
-      const recommended = recommendOllamaModel(models, 'balanced')
+      const ranked = rankOllamaModels(readiness.models, 'balanced')
+      const recommended = recommendOllamaModel(readiness.models, 'balanced')
       if (!cancelled) {
         setStatus({
           state: 'ready',

src/components/ConsoleOAuthFlow.test.tsx

@@ -112,8 +112,10 @@ test('third-party provider branch opens the first-run provider manager', async (
   )
 
   expect(output).toContain('Set up provider')
+  // Use alphabetically-early sentinels so they remain visible in the
+  // 13-row test frame after the provider list was sorted A→Z.
   expect(output).toContain('Anthropic')
-  expect(output).toContain('OpenAI')
-  expect(output).toContain('Ollama')
-  expect(output).toContain('LM Studio')
+  expect(output).toContain('Azure OpenAI')
+  expect(output).toContain('DeepSeek')
+  expect(output).toContain('Google Gemini')
 })

src/components/ProviderManager.test.tsx

@@ -97,6 +97,47 @@ async function waitForCondition(
   throw new Error('Timed out waiting for ProviderManager test condition')
 }
 
+// Provider list is sorted alphabetically by label in the preset picker, so
+// reaching a given provider takes more keypresses than it used to. Keep the
+// target-by-label indirection here so these tests survive future list edits
+// without further churn.
+//
+// Order matches ProviderManager.renderPresetSelection() when
+// canUseCodexOAuth === true (default in mocked tests).
+const PRESET_ORDER = [
+  'Alibaba Coding Plan',
+  'Alibaba Coding Plan (China)',
+  'Anthropic',
+  'Atomic Chat',
+  'Azure OpenAI',
+  'Codex OAuth',
+  'DeepSeek',
+  'Google Gemini',
+  'Groq',
+  'LM Studio',
+  'MiniMax',
+  'Mistral',
+  'Moonshot AI',
+  'NVIDIA NIM',
+  'Ollama',
+  'OpenAI',
+  'OpenRouter',
+  'Together AI',
+  'Custom',
+] as const
+
+async function navigateToPreset(
+  stdin: { write: (data: string) => void },
+  label: (typeof PRESET_ORDER)[number],
+): Promise<void> {
+  const index = PRESET_ORDER.indexOf(label)
+  if (index < 0) throw new Error(`Unknown preset label: ${label}`)
+  for (let i = 0; i < index; i++) {
+    stdin.write('j')
+    await Bun.sleep(25)
+  }
+}
+
 function createDeferred<T>(): {
   promise: Promise<T>
   resolve: (value: T) => void
@@ -149,17 +190,21 @@ function mockProviderManagerDependencies(
     applySavedProfileToCurrentSession?: (...args: unknown[]) => Promise<string | null>
     clearCodexCredentials?: () => { success: boolean; warning?: string }
     getProviderProfiles?: () => unknown[]
-    hasLocalOllama?: () => Promise<boolean>
-    listOllamaModels?: () => Promise<
-      Array<{
-        name: string
-        sizeBytes?: number | null
-        family?: string | null
-        families?: string[]
-        parameterSize?: string | null
-        quantizationLevel?: string | null
-      }>
-    >
+    probeOllamaGenerationReadiness?: () => Promise<{
+      state: 'ready' | 'unreachable' | 'no_models' | 'generation_failed'
+      models: Array<
+        {
+          name: string
+          sizeBytes?: number | null
+          family?: string | null
+          families?: string[]
+          parameterSize?: string | null
+          quantizationLevel?: string | null
+        }
+      >
+      probeModel?: string
+      detail?: string
+    }>
     codexSyncRead?: () => unknown
     codexAsyncRead?: () => Promise<unknown>
     updateProviderProfile?: (...args: unknown[]) => unknown
@@ -189,8 +234,12 @@ function mockProviderManagerDependencies(
   })
 
   mock.module('../utils/providerDiscovery.js', () => ({
-    hasLocalOllama: options?.hasLocalOllama ?? (async () => false),
-    listOllamaModels: options?.listOllamaModels ?? (async () => []),
+    probeOllamaGenerationReadiness:
+      options?.probeOllamaGenerationReadiness ??
+      (async () => ({
+        state: 'unreachable' as const,
+        models: [],
+      })),
   }))
 
   mock.module('../utils/githubModelsCredentials.js', () => ({
@@ -455,19 +504,22 @@ test('ProviderManager first-run Ollama preset auto-detects installed models', as
     async () => undefined,
     {
       addProviderProfile,
-      hasLocalOllama: async () => true,
-      listOllamaModels: async () => [
-        {
-          name: 'gemma4:31b-cloud',
-          family: 'gemma',
-          parameterSize: '31b',
-        },
-        {
-          name: 'kimi-k2.5:cloud',
-          family: 'kimi',
-          parameterSize: '2.5b',
-        },
-      ],
+      probeOllamaGenerationReadiness: async () => ({
+        state: 'ready',
+        models: [
+          {
+            name: 'gemma4:31b-cloud',
+            family: 'gemma',
+            parameterSize: '31b',
+          },
+          {
+            name: 'kimi-k2.5:cloud',
+            family: 'kimi',
+            parameterSize: '2.5b',
+          },
+        ],
+        probeModel: 'gemma4:31b-cloud',
+      }),
     },
   )
 
@@ -480,11 +532,10 @@ test('ProviderManager first-run Ollama preset auto-detects installed models', as
 
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Set up provider') && frame.includes('Ollama'),
+    frame => frame.includes('Set up provider'),
   )
 
-  mounted.stdin.write('j')
-  await Bun.sleep(50)
+  await navigateToPreset(mounted.stdin, 'Ollama')
   mounted.stdin.write('\r')
 
   const modelFrame = await waitForFrameOutput(
@@ -579,12 +630,7 @@ test('ProviderManager first-run Codex OAuth switches the current session after l
     frame => frame.includes('Set up provider') && frame.includes('Codex OAuth'),
   )
 
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
+  await navigateToPreset(mounted.stdin, 'Codex OAuth')
   mounted.stdin.write('\r')
 
   await waitForCondition(() => onDone.mock.calls.length > 0)
@@ -676,12 +722,7 @@ test('ProviderManager first-run Codex OAuth reports next-startup fallback when s
     frame => frame.includes('Set up provider') && frame.includes('Codex OAuth'),
   )
 
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
+  await navigateToPreset(mounted.stdin, 'Codex OAuth')
   mounted.stdin.write('\r')
 
   await waitForCondition(() => onDone.mock.calls.length > 0)
@@ -775,12 +816,7 @@ test('ProviderManager does not hijack a manual Codex profile when OAuth credenti
     frame => frame.includes('Set up provider') && frame.includes('Codex OAuth'),
   )
 
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
-  mounted.stdin.write('j')
-  await Bun.sleep(25)
+  await navigateToPreset(mounted.stdin, 'Codex OAuth')
   mounted.stdin.write('\r')
 
   await waitForCondition(() => onDone.mock.calls.length > 0)

src/components/ProviderManager.tsx

@@ -37,13 +37,16 @@ import {
   readGithubModelsTokenAsync,
 } from '../utils/githubModelsCredentials.js'
 import {
-  hasLocalOllama,
-  listOllamaModels,
+  probeAtomicChatReadiness,
+  probeOllamaGenerationReadiness,
+  type AtomicChatReadiness,
+  type OllamaGenerationReadiness,
 } from '../utils/providerDiscovery.js'
 import {
   rankOllamaModels,
   recommendOllamaModel,
 } from '../utils/providerRecommendation.js'
+import { redactUrlForDisplay } from '../utils/urlRedaction.js'
 import { updateSettingsForSource } from '../utils/settings/settings.js'
 import {
   type OptionWithDescription,
@@ -52,7 +55,6 @@ import {
 import { Pane } from './design-system/Pane.js'
 import TextInput from './TextInput.js'
 import { useCodexOAuthFlow } from './useCodexOAuthFlow.js'
-import { useSetAppState } from '../state/AppState.js'
 
 export type ProviderManagerResult = {
   action: 'saved' | 'cancelled'
@@ -69,6 +71,7 @@ type Screen =
   | 'menu'
   | 'select-preset'
   | 'select-ollama-model'
+  | 'select-atomic-chat-model'
   | 'codex-oauth'
   | 'form'
   | 'select-active'
@@ -89,6 +92,16 @@ type OllamaSelectionState =
     }
   | { state: 'unavailable'; message: string }
 
+type AtomicChatSelectionState =
+  | { state: 'idle' }
+  | { state: 'loading' }
+  | {
+      state: 'ready'
+      options: OptionWithDescription<string>[]
+      defaultValue?: string
+    }
+  | { state: 'unavailable'; message: string }
+
 const FORM_STEPS: Array<{
   key: DraftField
   label: string
@@ -222,6 +235,44 @@ function getGithubProviderSummary(
   return `github-models · ${GITHUB_PROVIDER_DEFAULT_BASE_URL} · ${getGithubProviderModel(processEnv)} · ${credentialSummary}${activeSuffix}`
 }
 
+function describeAtomicChatSelectionIssue(
+  readiness: AtomicChatReadiness,
+  baseUrl: string,
+): string {
+  if (readiness.state === 'unreachable') {
+    return `Could not reach Atomic Chat at ${redactUrlForDisplay(baseUrl)}. Start the Atomic Chat app first, or enter the endpoint manually.`
+  }
+
+  if (readiness.state === 'no_models') {
+    return 'Atomic Chat is running, but no models are loaded. Download and load a model inside the Atomic Chat app first, or enter details manually.'
+  }
+
+  return ''
+}
+
+function describeOllamaSelectionIssue(
+  readiness: OllamaGenerationReadiness,
+  baseUrl: string,
+): string {
+  if (readiness.state === 'unreachable') {
+    return `Could not reach Ollama at ${redactUrlForDisplay(baseUrl)}. Start Ollama first, or enter the endpoint manually.`
+  }
+
+  if (readiness.state === 'no_models') {
+    return 'Ollama is running, but no installed models were found. Pull a chat model such as qwen2.5-coder:7b or llama3.1:8b first, or enter details manually.'
+  }
+
+  if (readiness.state === 'generation_failed') {
+    const modelHint = readiness.probeModel ?? 'the selected model'
+    const detailSuffix = readiness.detail
+      ? ` Details: ${readiness.detail}.`
+      : ''
+    return `Ollama is reachable and models are installed, but a generation probe failed for ${modelHint}.${detailSuffix} Run "ollama run ${modelHint}" once and retry, or enter details manually.`
+  }
+
+  return ''
+}
+
 function findCodexOAuthProfile(
   profiles: ProviderProfile[],
   profileId?: string,
@@ -333,10 +384,12 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
   const initialIsGithubActive = isEnvTruthy(process.env.CLAUDE_CODE_USE_GITHUB)
   const initialHasGithubCredential = initialGithubCredentialSource !== 'none'
 
-  const [profiles, setProfiles] = React.useState(() => getProviderProfiles())
-  const [activeProfileId, setActiveProfileId] = React.useState(
-    () => getActiveProviderProfile()?.id,
-  )
+  // Deferred initialization: useState initializers run synchronously during
+  // render, so getProviderProfiles() and getActiveProviderProfile() would block
+  // the UI on first mount (sync file I/O). Use empty initial values and load
+  // asynchronously in useEffect with queueMicrotask to keep UI responsive.
+  const [profiles, setProfiles] = React.useState<ProviderProfile[]>([])
+  const [activeProfileId, setActiveProfileId] = React.useState<string | undefined>()
   const [githubProviderAvailable, setGithubProviderAvailable] = React.useState(
     () => isGithubProviderAvailable(initialGithubCredentialSource),
   )
@@ -370,11 +423,88 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
   const [ollamaSelection, setOllamaSelection] = React.useState<OllamaSelectionState>({
     state: 'idle',
   })
+  const [atomicChatSelection, setAtomicChatSelection] =
+    React.useState<AtomicChatSelectionState>({ state: 'idle' })
+  // Deferred initialization: useState initializers run synchronously during
+  // render, so getProviderProfiles() and getActiveProviderProfile() would block
+  // the UI (sync file I/O). Defer to queueMicrotask after first render.
+  // In test environment, skip defer to avoid timing issues with mocks.
+  const [isInitializing, setIsInitializing] = React.useState(
+    process.env.NODE_ENV !== 'test',
+  )
+  const [isActivating, setIsActivating] = React.useState(false)
+  const isRefreshingRef = React.useRef(false)
+
+  React.useEffect(() => {
+    // Skip deferred initialization in test environment (mocks are synchronous)
+    if (process.env.NODE_ENV === 'test') {
+      setProfiles(getProviderProfiles())
+      setActiveProfileId(getActiveProviderProfile()?.id)
+      setIsInitializing(false)
+      return
+    }
+
+    queueMicrotask(() => {
+      const profilesData = getProviderProfiles()
+      const activeId = getActiveProviderProfile()?.id
+      setProfiles(profilesData)
+      setActiveProfileId(activeId)
+      setIsInitializing(false)
+    })
+  }, [])
 
   const currentStep = FORM_STEPS[formStepIndex] ?? FORM_STEPS[0]
   const currentStepKey = currentStep.key
   const currentValue = draft[currentStepKey]
 
+  // Memoize menu options to prevent unnecessary re-renders when navigating
+  // the select menu. Without this, each arrow key press creates a new options
+  // array reference, causing Select to re-render and feel sluggish.
+  const hasProfiles = profiles.length > 0
+  const hasSelectableProviders = hasProfiles || githubProviderAvailable
+  const menuOptions = React.useMemo(
+    () => [
+      {
+        value: 'add',
+        label: 'Add provider',
+        description: 'Create a new provider profile',
+      },
+      {
+        value: 'activate',
+        label: 'Set active provider',
+        description: 'Switch the active provider profile',
+        disabled: !hasSelectableProviders,
+      },
+      {
+        value: 'edit',
+        label: 'Edit provider',
+        description: 'Update URL, model, or key',
+        disabled: !hasProfiles,
+      },
+      {
+        value: 'delete',
+        label: 'Delete provider',
+        description: 'Remove a provider profile',
+        disabled: !hasSelectableProviders,
+      },
+      ...(hasStoredCodexOAuthCredentials
+        ? [
+            {
+              value: 'logout-codex-oauth',
+              label: 'Log out Codex OAuth',
+              description: 'Clear securely stored Codex OAuth credentials',
+            },
+          ]
+        : []),
+      {
+        value: 'done',
+        label: 'Done',
+        description: 'Return to chat',
+      },
+    ],
+    [hasSelectableProviders, hasProfiles, hasStoredCodexOAuthCredentials],
+  )
+
   const refreshGithubProviderState = React.useCallback((): void => {
     const envCredentialSource = getGithubCredentialSourceFromEnv()
     const githubActive = isEnvTruthy(process.env.CLAUDE_CODE_USE_GITHUB)
@@ -450,32 +580,21 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
     setOllamaSelection({ state: 'loading' })
 
     void (async () => {
-      const available = await hasLocalOllama(draft.baseUrl)
-      if (!available) {
+      const readiness = await probeOllamaGenerationReadiness({
+        baseUrl: draft.baseUrl,
+      })
+      if (readiness.state !== 'ready') {
         if (!cancelled) {
           setOllamaSelection({
             state: 'unavailable',
-            message:
-              'Could not reach Ollama. Start Ollama first, or enter the endpoint manually.',
+            message: describeOllamaSelectionIssue(readiness, draft.baseUrl),
           })
         }
         return
       }
 
-      const models = await listOllamaModels(draft.baseUrl)
-      if (models.length === 0) {
-        if (!cancelled) {
-          setOllamaSelection({
-            state: 'unavailable',
-            message:
-              'Ollama is running, but no installed models were found. Pull a chat model such as qwen2.5-coder:7b or llama3.1:8b first, or enter details manually.',
-          })
-        }
-        return
-      }
-
-      const ranked = rankOllamaModels(models, 'balanced')
-      const recommended = recommendOllamaModel(models, 'balanced')
+      const ranked = rankOllamaModels(readiness.models, 'balanced')
+      const recommended = recommendOllamaModel(readiness.models, 'balanced')
       if (!cancelled) {
         setOllamaSelection({
           state: 'ready',
@@ -494,12 +613,61 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
     }
   }, [draft.baseUrl, screen])
 
+  React.useEffect(() => {
+    if (screen !== 'select-atomic-chat-model') {
+      return
+    }
+
+    let cancelled = false
+    setAtomicChatSelection({ state: 'loading' })
+
+    void (async () => {
+      const readiness = await probeAtomicChatReadiness({
+        baseUrl: draft.baseUrl,
+      })
+      if (readiness.state !== 'ready') {
+        if (!cancelled) {
+          setAtomicChatSelection({
+            state: 'unavailable',
+            message: describeAtomicChatSelectionIssue(readiness, draft.baseUrl),
+          })
+        }
+        return
+      }
+
+      if (!cancelled) {
+        setAtomicChatSelection({
+          state: 'ready',
+          defaultValue: readiness.models[0],
+          options: readiness.models.map(model => ({
+            label: model,
+            value: model,
+          })),
+        })
+      }
+    })()
+
+    return () => {
+      cancelled = true
+    }
+  }, [draft.baseUrl, screen])
+
   function refreshProfiles(): void {
-    const nextProfiles = getProviderProfiles()
-    setProfiles(nextProfiles)
-    setActiveProfileId(getActiveProviderProfile()?.id)
-    refreshGithubProviderState()
-    refreshCodexOAuthCredentialState()
+    // Defer sync I/O to next microtask to prevent UI freeze.
+    // getProviderProfiles() and getActiveProviderProfile() read config files
+    // synchronously, which can block the main thread on Windows (antivirus, disk cache).
+    // queueMicrotask ensures the current render completes first.
+    if (isRefreshingRef.current) return
+    isRefreshingRef.current = true
+
+    queueMicrotask(() => {
+      const nextProfiles = getProviderProfiles()
+      setProfiles(nextProfiles)
+      setActiveProfileId(getActiveProviderProfile()?.id)
+      refreshGithubProviderState()
+      refreshCodexOAuthCredentialState()
+      isRefreshingRef.current = false
+    })
   }
 
   function clearStartupProviderOverrideFromUserSettings(): string | null {
@@ -572,12 +740,24 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
   async function activateSelectedProvider(profileId: string): Promise<void> {
     let providerLabel = 'provider'
 
+    // Set loading state before sync I/O to keep UI responsive
+    setIsActivating(true)
+    setStatusMessage('Activating provider...')
+
     try {
+      // Defer sync I/O to next microtask - UI renders loading state first.
+      // setActiveProviderProfile(), activateGithubProvider(), and
+      // clearStartupProviderOverrideFromUserSettings() all perform sync file writes
+      // (saveGlobalConfig, saveProfileFile, updateSettingsForSource) which can
+      // block the main thread on Windows (antivirus, disk cache, NTFS metadata).
+      await new Promise<void>(resolve => queueMicrotask(resolve))
+
       if (profileId === GITHUB_PROVIDER_ID) {
         providerLabel = GITHUB_PROVIDER_LABEL
         const githubError = activateGithubProvider()
         if (githubError) {
           setErrorMessage(`Could not activate GitHub provider: ${githubError}`)
+          setIsActivating(false)
           returnToMenu()
           return
         }
@@ -593,6 +773,7 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
           mainLoopModel: GITHUB_PROVIDER_DEFAULT_MODEL,
         }))
         setStatusMessage(`Active provider: ${GITHUB_PROVIDER_LABEL}`)
+        setIsActivating(false)
         returnToMenu()
         return
       }
@@ -600,6 +781,7 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
       const active = setActiveProviderProfile(profileId)
       if (!active) {
         setErrorMessage('Could not change active provider.')
+        setIsActivating(false)
         returnToMenu()
         return
       }
@@ -647,10 +829,12 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
             ? `Active provider: ${active.name}. Warning: could not clear startup provider override (${settingsOverrideError}).`
             : `Active provider: ${active.name}`,
       )
+      setIsActivating(false)
       returnToMenu()
     } catch (error) {
       refreshProfiles()
       setStatusMessage(undefined)
+      setIsActivating(false)
       const detail = error instanceof Error ? error.message : String(error)
       setErrorMessage(`Could not finish activating ${providerLabel}: ${detail}`)
       returnToMenu()
@@ -774,6 +958,12 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
       return
     }
 
+    if (preset === 'atomic-chat') {
+      setAtomicChatSelection({ state: 'loading' })
+      setScreen('select-atomic-chat-model')
+      return
+    }
+
     setScreen('form')
   }
 
@@ -849,6 +1039,86 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
     returnToMenu()
   }
 
+  function renderAtomicChatSelection(): React.ReactNode {
+    if (
+      atomicChatSelection.state === 'loading' ||
+      atomicChatSelection.state === 'idle'
+    ) {
+      return (
+        <Box flexDirection="column" gap={1}>
+          <Text color="remember" bold>
+            Checking Atomic Chat
+          </Text>
+          <Text dimColor>Looking for loaded Atomic Chat models...</Text>
+        </Box>
+      )
+    }
+
+    if (atomicChatSelection.state === 'unavailable') {
+      return (
+        <Box flexDirection="column" gap={1}>
+          <Text color="remember" bold>
+            Atomic Chat setup
+          </Text>
+          <Text dimColor>{atomicChatSelection.message}</Text>
+          <Select
+            options={[
+              {
+                value: 'manual',
+                label: 'Enter manually',
+                description: 'Fill in the base URL and model yourself',
+              },
+              {
+                value: 'back',
+                label: 'Back',
+                description: 'Choose another provider preset',
+              },
+            ]}
+            onChange={(value: string) => {
+              if (value === 'manual') {
+                setFormStepIndex(0)
+                setCursorOffset(draft.name.length)
+                setScreen('form')
+                return
+              }
+              setScreen('select-preset')
+            }}
+            onCancel={() => setScreen('select-preset')}
+            visibleOptionCount={2}
+          />
+        </Box>
+      )
+    }
+
+    return (
+      <Box flexDirection="column" gap={1}>
+        <Text color="remember" bold>
+          Choose an Atomic Chat model
+        </Text>
+        <Text dimColor>
+          Pick one of the models loaded in Atomic Chat to save into a local
+          provider profile.
+        </Text>
+        <Select
+          options={atomicChatSelection.options}
+          defaultValue={atomicChatSelection.defaultValue}
+          defaultFocusValue={atomicChatSelection.defaultValue}
+          inlineDescriptions
+          visibleOptionCount={Math.min(8, atomicChatSelection.options.length)}
+          onChange={(value: string) => {
+            const nextDraft = {
+              ...draft,
+              model: value,
+            }
+            setDraft(nextDraft)
+            persistDraft(nextDraft)
+          }}
+          onCancel={() => setScreen('select-preset')}
+        />
+      </Box>
+    )
+  }
+
   function renderOllamaSelection(): React.ReactNode {
     if (ollamaSelection.state === 'loading' || ollamaSelection.state === 'idle') {
       return (
@@ -979,21 +1249,35 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
 
   function renderPresetSelection(): React.ReactNode {
     const canUseCodexOAuth = !isBareMode()
+    // Providers sorted alphabetically by label. `Custom` is pinned to the end
+    // because it's the catch-all / escape hatch — users scanning the list
+    // should always find known providers first. `Skip for now` (first-run
+    // only) comes last, after Custom.
     const options = [
+      {
+        value: 'dashscope-intl',
+        label: 'Alibaba Coding Plan',
+        description: 'Alibaba DashScope International endpoint',
+      },
+      {
+        value: 'dashscope-cn',
+        label: 'Alibaba Coding Plan (China)',
+        description: 'Alibaba DashScope China endpoint',
+      },
       {
         value: 'anthropic',
         label: 'Anthropic',
         description: 'Native Claude API (x-api-key auth)',
       },
       {
-        value: 'ollama',
-        label: 'Ollama',
-        description: 'Local or remote Ollama endpoint',
+        value: 'atomic-chat',
+        label: 'Atomic Chat',
+        description: 'Local Model Provider',
       },
       {
-        value: 'openai',
-        label: 'OpenAI',
-        description: 'OpenAI API with API key',
+        value: 'azure-openai',
+        label: 'Azure OpenAI',
+        description: 'Azure OpenAI endpoint (model=deployment name)',
       },
       ...(canUseCodexOAuth
         ? [
@@ -1005,11 +1289,6 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
             },
           ]
         : []),
-      {
-        value: 'moonshotai',
-        label: 'Moonshot AI',
-        description: 'Kimi OpenAI-compatible endpoint',
-      },
       {
         value: 'deepseek',
         label: 'DeepSeek',
@@ -1020,50 +1299,30 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
         label: 'Google Gemini',
         description: 'Gemini OpenAI-compatible endpoint',
       },
-      {
-        value: 'together',
-        label: 'Together AI',
-        description: 'Together chat/completions endpoint',
-      },
       {
         value: 'groq',
         label: 'Groq',
         description: 'Groq OpenAI-compatible endpoint',
       },
-      {
-        value: 'mistral',
-        label: 'Mistral',
-        description: 'Mistral OpenAI-compatible endpoint',
-      },
-      {
-        value: 'azure-openai',
-        label: 'Azure OpenAI',
-        description: 'Azure OpenAI endpoint (model=deployment name)',
-      },
-      {
-        value: 'openrouter',
-        label: 'OpenRouter',
-        description: 'OpenRouter OpenAI-compatible endpoint',
-      },
       {
         value: 'lmstudio',
         label: 'LM Studio',
         description: 'Local LM Studio endpoint',
       },
       {
-        value: 'dashscope-cn',
-        label: 'Alibaba Coding Plan (China)',
-        description: 'Alibaba DashScope China endpoint',
+        value: 'minimax',
+        label: 'MiniMax',
+        description: 'MiniMax API endpoint',
       },
       {
-        value: 'dashscope-intl',
-        label: 'Alibaba Coding Plan',
-        description: 'Alibaba DashScope International endpoint',
+        value: 'mistral',
+        label: 'Mistral',
+        description: 'Mistral OpenAI-compatible endpoint',
       },
       {
-        value: 'custom',
-        label: 'Custom',
-        description: 'Any OpenAI-compatible provider',
+        value: 'moonshotai',
+        label: 'Moonshot AI',
+        description: 'Kimi OpenAI-compatible endpoint',
       },
       {
         value: 'nvidia-nim',
@@ -1071,9 +1330,29 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
         description: 'NVIDIA NIM endpoint',
       },
       {
-        value: 'minimax',
-        label: 'MiniMax',
-        description: 'MiniMax API endpoint',
+        value: 'ollama',
+        label: 'Ollama',
+        description: 'Local or remote Ollama endpoint',
+      },
+      {
+        value: 'openai',
+        label: 'OpenAI',
+        description: 'OpenAI API with API key',
+      },
+      {
+        value: 'openrouter',
+        label: 'OpenRouter',
+        description: 'OpenRouter OpenAI-compatible endpoint',
+      },
+      {
+        value: 'together',
+        label: 'Together AI',
+        description: 'Together chat/completions endpoint',
+      },
+      {
+        value: 'custom',
+        label: 'Custom',
+        description: 'Any OpenAI-compatible provider',
       },
       ...(mode === 'first-run'
         ? [
@@ -1165,49 +1444,10 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
   }
 
   function renderMenu(): React.ReactNode {
+    // Use memoized menuOptions from component scope
     const hasProfiles = profiles.length > 0
     const hasSelectableProviders = hasProfiles || githubProviderAvailable
 
-    const options = [
-      {
-        value: 'add',
-        label: 'Add provider',
-        description: 'Create a new provider profile',
-      },
-      {
-        value: 'activate',
-        label: 'Set active provider',
-        description: 'Switch the active provider profile',
-        disabled: !hasSelectableProviders,
-      },
-      {
-        value: 'edit',
-        label: 'Edit provider',
-        description: 'Update URL, model, or key',
-        disabled: !hasProfiles,
-      },
-      {
-        value: 'delete',
-        label: 'Delete provider',
-        description: 'Remove a provider profile',
-        disabled: !hasSelectableProviders,
-      },
-      ...(hasStoredCodexOAuthCredentials
-        ? [
-            {
-              value: 'logout-codex-oauth',
-              label: 'Log out Codex OAuth',
-              description: 'Clear securely stored Codex OAuth credentials',
-            },
-          ]
-        : []),
-      {
-        value: 'done',
-        label: 'Done',
-        description: 'Return to chat',
-      },
-    ]
-
     return (
       <Box flexDirection="column" gap={1}>
         <Text color="remember" bold>
@@ -1244,7 +1484,7 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
           )}
         </Box>
         <Select
-          options={options}
+          options={menuOptions}
           onChange={(value: string) => {
             setErrorMessage(undefined)
             switch (value) {
@@ -1257,7 +1497,7 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
                 }
                 break
               case 'edit':
-                if (profiles.length > 0) {
+                if (hasProfiles) {
                   setScreen('select-edit')
                 }
                 break
@@ -1314,7 +1554,7 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
           }}
           onCancel={() => closeWithCancelled('Provider manager closed')}
           defaultFocusValue={menuFocusValue}
-          visibleOptionCount={options.length}
+          visibleOptionCount={menuOptions.length}
         />
       </Box>
     )
@@ -1393,6 +1633,9 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
     case 'select-ollama-model':
       content = renderOllamaSelection()
       break
+    case 'select-atomic-chat-model':
+      content = renderAtomicChatSelection()
+      break
     case 'codex-oauth':
       content = (
         <CodexOAuthSetup
@@ -1550,5 +1793,21 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
       break
   }
 
-  return <Pane color="permission">{content}</Pane>
+  return (
+    <Pane color="permission">
+      {isInitializing ? (
+        <Box flexDirection="column" gap={1}>
+          <Text color="remember" bold>Loading providers...</Text>
+          <Text dimColor>Reading provider profiles from disk.</Text>
+        </Box>
+      ) : isActivating ? (
+        <Box flexDirection="column" gap={1}>
+          <Text color="remember" bold>Activating provider...</Text>
+          <Text dimColor>Please wait while the provider is being configured.</Text>
+        </Box>
+      ) : (
+        content
+      )}
+    </Pane>
+  )
 }

src/components/Settings/Config.tsx

@@ -281,6 +281,24 @@ export function Config({
         enabled: autoCompactEnabled
       });
     }
+  }, {
+    id: 'toolHistoryCompressionEnabled',
+    label: 'Tool history compression',
+    value: globalConfig.toolHistoryCompressionEnabled,
+    type: 'boolean' as const,
+    onChange(toolHistoryCompressionEnabled: boolean) {
+      saveGlobalConfig(current => ({
+        ...current,
+        toolHistoryCompressionEnabled
+      }));
+      setGlobalConfig({
+        ...getGlobalConfig(),
+        toolHistoryCompressionEnabled
+      });
+      logEvent('tengu_tool_history_compression_setting_changed', {
+        enabled: toolHistoryCompressionEnabled
+      });
+    }
   }, {
     id: 'spinnerTipsEnabled',
     label: 'Show tips',
@@ -1158,6 +1176,9 @@ export function Config({
     if (globalConfig.autoCompactEnabled !== initialConfig.current.autoCompactEnabled) {
       formattedChanges.push(`${globalConfig.autoCompactEnabled ? 'Enabled' : 'Disabled'} auto-compact`);
     }
+    if (globalConfig.toolHistoryCompressionEnabled !== initialConfig.current.toolHistoryCompressionEnabled) {
+      formattedChanges.push(`${globalConfig.toolHistoryCompressionEnabled ? 'Enabled' : 'Disabled'} tool history compression`);
+    }
     if (globalConfig.respectGitignore !== initialConfig.current.respectGitignore) {
       formattedChanges.push(`${globalConfig.respectGitignore ? 'Enabled' : 'Disabled'} respect .gitignore in file picker`);
     }

src/components/StartupScreen.ts

@@ -123,6 +123,8 @@ function detectProvider(): { name: string; model: string; baseUrl: string; isLoc
       name = 'MiniMax'
     else if (resolvedRequest.transport === 'codex_responses' || baseUrl.includes('chatgpt.com/backend-api/codex'))
       name = 'Codex'
+    else if (/moonshot/i.test(baseUrl) || /kimi/i.test(rawModel))
+      name = 'Moonshot (Kimi)'
     else if (/deepseek/i.test(baseUrl) || /deepseek/i.test(rawModel))
       name = 'DeepSeek'
     else if (/openrouter/i.test(baseUrl))

src/components/memory/memoryFileSelectorPaths.test.ts

@@ -53,17 +53,20 @@ describe('getProjectMemoryPathForSelector', () => {
   })
 
   test('defaults to a new AGENTS.md in the current cwd when no project file is loaded', () => {
-    expect(getProjectMemoryPathForSelector([], '/repo/packages/app')).toBe(
-      '/repo/packages/app/AGENTS.md',
+    const cwd = join('/repo', 'packages', 'app')
+    expect(getProjectMemoryPathForSelector([], cwd)).toBe(
+      join(cwd, 'AGENTS.md'),
     )
   })
 
   test('ignores loaded project instruction files outside the current cwd ancestry', () => {
+    const outsideRepoPath = join('/other-worktree', 'AGENTS.md')
+    const cwd = join('/repo', 'packages', 'app')
     expect(
       getProjectMemoryPathForSelector(
-        [projectFile('/other-worktree/AGENTS.md')],
-        '/repo/packages/app',
+        [projectFile(outsideRepoPath)],
+        cwd,
       ),
-    ).toBe('/repo/packages/app/AGENTS.md')
+    ).toBe(join(cwd, 'AGENTS.md'))
   })
 })

src/constants/prompts.ts

@@ -823,6 +823,11 @@ function getFunctionResultClearingSection(model: string): string | null {
     return null
   }
   const config = getCachedMCConfigForFRC()
+  if (!config) {
+    // External/stub builds return null from getCachedMCConfig — abort the
+    // section rather than trying to read .supportedModels off null.
+    return null
+  }
   const isModelSupported = config.supportedModels?.some(pattern =>
     model.includes(pattern),
   )

src/hooks/useOfficialMarketplaceNotification.tsx

@@ -19,7 +19,7 @@ async function _temp() {
     logForDebugging("Showing marketplace config save failure notification");
     notifs.push({
       key: "marketplace-config-save-failed",
-      jsx: <Text color="error">Failed to save marketplace retry info · Check ~/.claude.json permissions</Text>,
+      jsx: <Text color="error">Failed to save marketplace retry info · Check ~/.openclaude.json permissions</Text>,
       priority: "immediate",
       timeoutMs: 10000
     });

src/hooks/useSwarmPermissionPoller.ts

@@ -1,34 +1,23 @@
 /**
- * Swarm Permission Poller Hook
+ * Swarm Permission Callback Registry
  *
- * This hook polls for permission responses from the team leader when running
- * as a worker agent in a swarm. When a response is received, it calls the
- * appropriate callback (onAllow/onReject) to continue execution.
+ * Manages callback registrations for permission requests and responses
+ * in agent swarms. Responses are delivered exclusively via the mailbox
+ * system (useInboxPoller → processMailboxPermissionResponse).
  *
- * This hook should be used in conjunction with the worker-side integration
- * in useCanUseTool.ts, which creates pending requests that this hook monitors.
+ * The legacy file-based polling (resolved/ directory) has been removed
+ * because it created an unauthenticated attack surface — any local process
+ * could forge approval files. The mailbox path is the sole active channel.
  */
 
-import { useCallback, useEffect, useRef } from 'react'
-import { useInterval } from 'usehooks-ts'
 import { logForDebugging } from '../utils/debug.js'
-import { errorMessage } from '../utils/errors.js'
 import {
   type PermissionUpdate,
   permissionUpdateSchema,
 } from '../utils/permissions/PermissionUpdateSchema.js'
-import {
-  isSwarmWorker,
-  type PermissionResponse,
-  pollForResponse,
-  removeWorkerResponse,
-} from '../utils/swarm/permissionSync.js'
-import { getAgentName, getTeamName } from '../utils/teammate.js'
-
-const POLL_INTERVAL_MS = 500
 
 /**
- * Validate permissionUpdates from external sources (mailbox IPC, disk polling).
+ * Validate permissionUpdates from external sources (mailbox IPC).
  * Malformed entries from buggy/old teammate processes are filtered out rather
  * than propagated unchecked into callback.onAllow().
  */
@@ -225,106 +214,9 @@ export function processSandboxPermissionResponse(params: {
   return true
 }
 
-/**
- * Process a permission response by invoking the registered callback
- */
-function processResponse(response: PermissionResponse): boolean {
-  const callback = pendingCallbacks.get(response.requestId)
-
-  if (!callback) {
-    logForDebugging(
-      `[SwarmPermissionPoller] No callback registered for request ${response.requestId}`,
-    )
-    return false
-  }
-
-  logForDebugging(
-    `[SwarmPermissionPoller] Processing response for request ${response.requestId}: ${response.decision}`,
-  )
-
-  // Remove from registry before invoking callback
-  pendingCallbacks.delete(response.requestId)
-
-  if (response.decision === 'approved') {
-    const permissionUpdates = parsePermissionUpdates(response.permissionUpdates)
-    const updatedInput = response.updatedInput
-    callback.onAllow(updatedInput, permissionUpdates)
-  } else {
-    callback.onReject(response.feedback)
-  }
-
-  return true
-}
-
-/**
- * Hook that polls for permission responses when running as a swarm worker.
- *
- * This hook:
- * 1. Only activates when isSwarmWorker() returns true
- * 2. Polls every 500ms for responses
- * 3. When a response is found, invokes the registered callback
- * 4. Cleans up the response file after processing
- */
-export function useSwarmPermissionPoller(): void {
-  const isProcessingRef = useRef(false)
-
-  const poll = useCallback(async () => {
-    // Don't poll if not a swarm worker
-    if (!isSwarmWorker()) {
-      return
-    }
-
-    // Prevent concurrent polling
-    if (isProcessingRef.current) {
-      return
-    }
-
-    // Don't poll if no callbacks are registered
-    if (pendingCallbacks.size === 0) {
-      return
-    }
-
-    isProcessingRef.current = true
-
-    try {
-      const agentName = getAgentName()
-      const teamName = getTeamName()
-
-      if (!agentName || !teamName) {
-        return
-      }
-
-      // Check each pending request for a response
-      for (const [requestId, _callback] of pendingCallbacks) {
-        const response = await pollForResponse(requestId, agentName, teamName)
-
-        if (response) {
-          // Process the response
-          const processed = processResponse(response)
-
-          if (processed) {
-            // Clean up the response from the worker's inbox
-            await removeWorkerResponse(requestId, agentName, teamName)
-          }
-        }
-      }
-    } catch (error) {
-      logForDebugging(
-        `[SwarmPermissionPoller] Error during poll: ${errorMessage(error)}`,
-      )
-    } finally {
-      isProcessingRef.current = false
-    }
-  }, [])
-
-  // Only poll if we're a swarm worker
-  const shouldPoll = isSwarmWorker()
-  useInterval(() => void poll(), shouldPoll ? POLL_INTERVAL_MS : null)
-
-  // Initial poll on mount
-  useEffect(() => {
-    if (isSwarmWorker()) {
-      void poll()
-    }
-  }, [poll])
-}
+// Legacy file-based polling (useSwarmPermissionPoller, processResponse)
+// has been removed. Permission responses are now delivered exclusively
+// via the mailbox system:
+//   Leader: sendPermissionResponseViaMailbox() → writeToMailbox()
+//   Worker: useInboxPoller → processMailboxPermissionResponse()
+// See: fix(security) — remove unauthenticated file-based permission channel

src/ink/termio/osc.test.ts

@@ -11,14 +11,16 @@ const execFileNoThrowMock = mock(
   async () => ({ code: 0, stdout: '', stderr: '' }),
 )
 
-mock.module('../../utils/execFileNoThrow.js', () => ({
-  execFileNoThrow: execFileNoThrowMock,
-  execFileNoThrowWithCwd: execFileNoThrowMock,
-}))
+function installOscMocks(): void {
+  mock.module('../../utils/execFileNoThrow.js', () => ({
+    execFileNoThrow: execFileNoThrowMock,
+    execFileNoThrowWithCwd: execFileNoThrowMock,
+  }))
 
-mock.module('../../utils/tempfile.js', () => ({
-  generateTempFilePath: generateTempFilePathMock,
-}))
+  mock.module('../../utils/tempfile.js', () => ({
+    generateTempFilePath: generateTempFilePathMock,
+  }))
+}
 
 async function importFreshOscModule() {
   return import(`./osc.ts?ts=${Date.now()}-${Math.random()}`)
@@ -45,6 +47,7 @@ async function waitForExecCall(
 
 describe('Windows clipboard fallback', () => {
   beforeEach(() => {
+    installOscMocks()
     execFileNoThrowMock.mockClear()
     generateTempFilePathMock.mockClear()
     process.env = { ...originalEnv }
@@ -62,14 +65,12 @@ describe('Windows clipboard fallback', () => {
     const { setClipboard } = await importFreshOscModule()
 
     await setClipboard('Привет мир')
-    await flushClipboardCopy()
+    const windowsCall = await waitForExecCall('powershell')
 
     expect(execFileNoThrowMock.mock.calls.some(([cmd]) => cmd === 'clip')).toBe(
       false,
     )
-    expect(
-      execFileNoThrowMock.mock.calls.some(([cmd]) => cmd === 'powershell'),
-    ).toBe(true)
+    expect(windowsCall).toBeDefined()
   })
 
   test('passes Windows clipboard text through a UTF-8 temp file instead of stdin', async () => {
@@ -97,6 +98,7 @@ describe('Windows clipboard fallback', () => {
 
 describe('clipboard path behavior remains stable', () => {
   beforeEach(() => {
+    installOscMocks()
     execFileNoThrowMock.mockClear()
     process.env = { ...originalEnv }
     delete process.env['SSH_CONNECTION']

src/migrations/resetAutoModeOptInForDefaultOffer.ts

@@ -12,7 +12,7 @@ import {
  * One-shot migration: clear skipAutoPermissionPrompt for users who accepted
  * the old 2-option AutoModeOptInDialog but don't have auto as their default.
  * Re-surfaces the dialog so they see the new "make it my default mode" option.
- * Guard lives in GlobalConfig (~/.claude.json), not settings.json, so it
+ * Guard lives in GlobalConfig (~/.openclaude.json), not settings.json, so it
  * survives settings resets and doesn't re-arm itself.
  *
  * Only runs when tengu_auto_mode_config.enabled === 'enabled'. For 'opt-in'

src/screens/REPL.tsx

@@ -3873,7 +3873,7 @@ export function REPL({
   // empty to non-empty, not on every length change -- otherwise a render loop
   // (concurrent onQuery thrashing, etc.) spams saveGlobalConfig, which hits
   // ELOCKED under concurrent sessions and falls back to unlocked writes.
-  // That write storm is the primary trigger for ~/.claude.json corruption
+  // That write storm is the primary trigger for ~/.openclaude.json corruption
   // (GH #3117).
   const hasCountedQueueUseRef = useRef(false);
   useEffect(() => {

src/services/analytics/growthbook.ts

@@ -334,7 +334,7 @@ async function processRemoteEvalPayload(
   // Empty object is truthy — without the length check, `{features: {}}`
   // (transient server bug, truncated response) would pass, clear the maps
   // below, return true, and syncRemoteEvalToDisk would wholesale-write `{}`
-  // to disk: total flag blackout for every process sharing ~/.claude.json.
+  // to disk: total flag blackout for every process sharing ~/.openclaude.json.
   if (!payload?.features || Object.keys(payload.features).length === 0) {
     return false
   }

src/services/api/claude.ts

@@ -23,6 +23,7 @@ import { randomUUID } from 'crypto'
 import {
   getAPIProvider,
   isFirstPartyAnthropicBaseUrl,
+  isGithubNativeAnthropicMode,
 } from 'src/utils/model/providers.js'
 import {
   getAttributionHeader,
@@ -334,8 +335,13 @@ export function getPromptCachingEnabled(model: string): boolean {
   // Prompt caching is an Anthropic-specific feature. Third-party providers
   // do not understand cache_control blocks and strict backends (e.g. Azure
   // Foundry) reject or flag requests that contain them.
+  //
+  // Exception: when the GitHub provider is configured in native Anthropic API
+  // mode (CLAUDE_CODE_GITHUB_ANTHROPIC_API=1), requests are sent in Anthropic
+  // format, so cache_control blocks are supported.
   const provider = getAPIProvider()
-  if (provider !== 'firstParty' && provider !== 'bedrock' && provider !== 'vertex') {
+  const isNativeGithub = isGithubNativeAnthropicMode(model)
+  if (provider !== 'firstParty' && provider !== 'bedrock' && provider !== 'vertex' && !isNativeGithub) {
     return false
   }
 
@@ -1211,7 +1217,7 @@ async function* queryModel(
     cachedMCEnabled = featureEnabled && modelSupported
     const config = getCachedMCConfig()
     logForDebugging(
-      `Cached MC gate: enabled=${featureEnabled} modelSupported=${modelSupported} model=${options.model} supportedModels=${jsonStringify(config.supportedModels)}`,
+      `Cached MC gate: enabled=${featureEnabled} modelSupported=${modelSupported} model=${options.model} supportedModels=${jsonStringify(config?.supportedModels)}`,
     )
   }
 

src/services/api/client.ts

@@ -14,6 +14,7 @@ import { getSmallFastModel } from 'src/utils/model/model.js'
 import {
   getAPIProvider,
   isFirstPartyAnthropicBaseUrl,
+  isGithubNativeAnthropicMode,
 } from 'src/utils/model/providers.js'
 import { getProxyFetchOptions } from 'src/utils/proxy.js'
 import {
@@ -174,6 +175,25 @@ export async function getAnthropicClient({
       providerOverride,
     }) as unknown as Anthropic
   }
+  // GitHub provider in native Anthropic API mode: send requests in Anthropic
+  // format so cache_control blocks are honoured and prompt caching works.
+  // Requires the GitHub endpoint (OPENAI_BASE_URL) to support Anthropic's
+  // messages API — set CLAUDE_CODE_GITHUB_ANTHROPIC_API=1 to opt in.
+  if (isGithubNativeAnthropicMode(model)) {
+    const githubBaseUrl =
+      process.env.OPENAI_BASE_URL?.replace(/\/$/, '') ??
+      'https://api.githubcopilot.com'
+    const githubToken =
+      process.env.GITHUB_TOKEN ?? process.env.GH_TOKEN ?? ''
+    const nativeArgs: ConstructorParameters<typeof Anthropic>[0] = {
+      ...ARGS,
+      baseURL: githubBaseUrl,
+      authToken: githubToken,
+      // No apiKey — we authenticate via Bearer token (authToken)
+      apiKey: null,
+    }
+    return new Anthropic(nativeArgs)
+  }
   if (
     isEnvTruthy(process.env.CLAUDE_CODE_USE_OPENAI) ||
     isEnvTruthy(process.env.CLAUDE_CODE_USE_GITHUB) ||

src/services/api/codexShim.test.ts

@@ -547,7 +547,7 @@ describe('Codex request translation', () => {
     ])
   })
 
-  test('strips leaked reasoning preamble from completed Codex text responses', () => {
+  test('strips <think> tag block from completed Codex text responses', () => {
     const message = convertCodexResponseToAnthropicMessage(
       {
         id: 'resp_1',
@@ -560,7 +560,7 @@ describe('Codex request translation', () => {
               {
                 type: 'output_text',
                 text:
-                  'The user just said "hey" - a simple greeting. I should respond briefly and friendly.\n\nHey! How can I help you today?',
+                  '<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?',
               },
             ],
           },
@@ -578,6 +578,37 @@ describe('Codex request translation', () => {
     ])
   })
 
+  test('strips unterminated <think> tag at block boundary in Codex completed response', () => {
+    const message = convertCodexResponseToAnthropicMessage(
+      {
+        id: 'resp_1',
+        model: 'gpt-5.4',
+        output: [
+          {
+            type: 'message',
+            role: 'assistant',
+            content: [
+              {
+                type: 'output_text',
+                text:
+                  'Here is the answer.\n<think>wait, let me reconsider the user request',
+              },
+            ],
+          },
+        ],
+        usage: { input_tokens: 12, output_tokens: 4 },
+      },
+      'gpt-5.4',
+    )
+
+    expect(message.content).toEqual([
+      {
+        type: 'text',
+        text: 'Here is the answer.',
+      },
+    ])
+  })
+
   test('translates Codex SSE text stream into Anthropic events', async () => {
     const responseText = [
       'event: response.output_item.added',
@@ -609,7 +640,7 @@ describe('Codex request translation', () => {
     ])
   })
 
-  test('strips leaked reasoning preamble from Codex SSE text stream', async () => {
+  test('strips <think> tag block from Codex SSE text stream', async () => {
     const responseText = [
       'event: response.output_item.added',
       'data: {"type":"response.output_item.added","item":{"id":"msg_1","type":"message","status":"in_progress","content":[],"role":"assistant"},"output_index":0,"sequence_number":0}',
@@ -618,13 +649,13 @@ describe('Codex request translation', () => {
       'data: {"type":"response.content_part.added","content_index":0,"item_id":"msg_1","output_index":0,"part":{"type":"output_text","text":""},"sequence_number":1}',
       '',
       'event: response.output_text.delta',
-      'data: {"type":"response.output_text.delta","content_index":0,"delta":"The user just said \\"hey\\" - a simple greeting. I should respond briefly and friendly.\\n\\nHey! How can I help you today?","item_id":"msg_1","output_index":0,"sequence_number":2}',
+      'data: {"type":"response.output_text.delta","content_index":0,"delta":"<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?","item_id":"msg_1","output_index":0,"sequence_number":2}',
       '',
       'event: response.output_item.done',
-      'data: {"type":"response.output_item.done","item":{"id":"msg_1","type":"message","status":"completed","content":[{"type":"output_text","text":"The user just said \\"hey\\" - a simple greeting. I should respond briefly and friendly.\\n\\nHey! How can I help you today?"}],"role":"assistant"},"output_index":0,"sequence_number":3}',
+      'data: {"type":"response.output_item.done","item":{"id":"msg_1","type":"message","status":"completed","content":[{"type":"output_text","text":"<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?"}],"role":"assistant"},"output_index":0,"sequence_number":3}',
       '',
       'event: response.completed',
-      'data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","model":"gpt-5.4","output":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"The user just said \\"hey\\" - a simple greeting. I should respond briefly and friendly.\\n\\nHey! How can I help you today?"}]}],"usage":{"input_tokens":2,"output_tokens":1}},"sequence_number":4}',
+      'data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","model":"gpt-5.4","output":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?"}]}],"usage":{"input_tokens":2,"output_tokens":1}},"sequence_number":4}',
       '',
     ].join('\n')
 
@@ -646,6 +677,50 @@ describe('Codex request translation', () => {
       }
     }
 
-    expect(textDeltas).toEqual(['Hey! How can I help you today?'])
+    expect(textDeltas.join('')).toBe('Hey! How can I help you today?')
+  })
+
+  test('preserves prose without tags (no phrase-based false positive)', async () => {
+    // Regression test: older phrase-based sanitizer would incorrectly strip text
+    // starting with "I should" or "The user". The tag-based approach leaves it alone.
+    const responseText = [
+      'event: response.output_item.added',
+      'data: {"type":"response.output_item.added","item":{"id":"msg_1","type":"message","status":"in_progress","content":[],"role":"assistant"},"output_index":0,"sequence_number":0}',
+      '',
+      'event: response.content_part.added',
+      'data: {"type":"response.content_part.added","content_index":0,"item_id":"msg_1","output_index":0,"part":{"type":"output_text","text":""},"sequence_number":1}',
+      '',
+      'event: response.output_text.delta',
+      'data: {"type":"response.output_text.delta","content_index":0,"delta":"I should note that the user role requires a briefly concise friendly response format.","item_id":"msg_1","output_index":0,"sequence_number":2}',
+      '',
+      'event: response.output_item.done',
+      'data: {"type":"response.output_item.done","item":{"id":"msg_1","type":"message","status":"completed","content":[{"type":"output_text","text":"I should note that the user role requires a briefly concise friendly response format."}],"role":"assistant"},"output_index":0,"sequence_number":3}',
+      '',
+      'event: response.completed',
+      'data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","model":"gpt-5.4","output":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"I should note that the user role requires a briefly concise friendly response format."}]}],"usage":{"input_tokens":2,"output_tokens":1}},"sequence_number":4}',
+      '',
+    ].join('\n')
+
+    const stream = new ReadableStream({
+      start(controller) {
+        controller.enqueue(new TextEncoder().encode(responseText))
+        controller.close()
+      },
+    })
+
+    const textDeltas: string[] = []
+    for await (const event of codexStreamToAnthropic(
+      new Response(stream),
+      'gpt-5.4',
+    )) {
+      const delta = (event as { delta?: { type?: string; text?: string } }).delta
+      if (delta?.type === 'text_delta' && typeof delta.text === 'string') {
+        textDeltas.push(delta.text)
+      }
+    }
+
+    expect(textDeltas.join('')).toBe(
+      'I should note that the user role requires a briefly concise friendly response format.',
+    )
   })
 })

src/services/api/codexShim.ts

@@ -1,4 +1,5 @@
 import { APIError } from '@anthropic-ai/sdk'
+import { compressToolHistory } from './compressToolHistory.js'
 import { fetchWithProxyRetry } from './fetchWithProxyRetry.js'
 import type {
   ResolvedCodexCredentials,
@@ -6,10 +7,9 @@ import type {
 } from './providerConfig.js'
 import { sanitizeSchemaForOpenAICompat } from './openaiSchemaSanitizer.js'
 import {
-  looksLikeLeakedReasoningPrefix,
-  shouldBufferPotentialReasoningPrefix,
-  stripLeakedReasoningPreamble,
-} from './reasoningLeakSanitizer.js'
+  createThinkTagFilter,
+  stripThinkTags,
+} from './thinkTagSanitizer.js'
 
 export interface AnthropicUsage {
   input_tokens: number
@@ -485,13 +485,15 @@ export async function performCodexRequest(options: {
   defaultHeaders: Record<string, string>
   signal?: AbortSignal
 }): Promise<Response> {
-  const input = convertAnthropicMessagesToResponsesInput(
+  const compressedMessages = compressToolHistory(
     options.params.messages as Array<{
       role?: string
       message?: { role?: string; content?: unknown }
       content?: unknown
     }>,
+    options.request.resolvedModel,
   )
+  const input = convertAnthropicMessagesToResponsesInput(compressedMessages)
   const body: Record<string, unknown> = {
     model: options.request.resolvedModel,
     input: input.length > 0
@@ -734,25 +736,22 @@ export async function* codexStreamToAnthropic(
     { index: number; toolUseId: string }
   >()
   let activeTextBlockIndex: number | null = null
-  let activeTextBuffer = ''
-  let textBufferMode: 'none' | 'pending' | 'strip' = 'none'
+  const thinkFilter = createThinkTagFilter()
   let nextContentBlockIndex = 0
   let sawToolUse = false
   let finalResponse: Record<string, any> | undefined
 
   const closeActiveTextBlock = async function* () {
     if (activeTextBlockIndex === null) return
-    if (textBufferMode !== 'none') {
-      const sanitized = stripLeakedReasoningPreamble(activeTextBuffer)
-      if (sanitized) {
-        yield {
-          type: 'content_block_delta',
-          index: activeTextBlockIndex,
-          delta: {
-            type: 'text_delta',
-            text: sanitized,
-          },
-        }
+    const tail = thinkFilter.flush()
+    if (tail) {
+      yield {
+        type: 'content_block_delta',
+        index: activeTextBlockIndex,
+        delta: {
+          type: 'text_delta',
+          text: tail,
+        },
       }
     }
     yield {
@@ -760,8 +759,6 @@ export async function* codexStreamToAnthropic(
       index: activeTextBlockIndex,
     }
     activeTextBlockIndex = null
-    activeTextBuffer = ''
-    textBufferMode = 'none'
   }
 
   const startTextBlockIfNeeded = async function* () {
@@ -837,43 +834,17 @@ export async function* codexStreamToAnthropic(
 
     if (event.event === 'response.output_text.delta') {
       yield* startTextBlockIfNeeded()
-      activeTextBuffer += payload.delta ?? ''
       if (activeTextBlockIndex !== null) {
-        if (
-          textBufferMode === 'strip' ||
-          looksLikeLeakedReasoningPrefix(activeTextBuffer)
-        ) {
-          textBufferMode = 'strip'
-          continue
-        }
-
-        if (textBufferMode === 'pending') {
-          if (shouldBufferPotentialReasoningPrefix(activeTextBuffer)) {
-            continue
-          }
+        const visible = thinkFilter.feed(payload.delta ?? '')
+        if (visible) {
           yield {
             type: 'content_block_delta',
             index: activeTextBlockIndex,
             delta: {
               type: 'text_delta',
-              text: activeTextBuffer,
+              text: visible,
             },
           }
-          textBufferMode = 'none'
-          continue
-        }
-
-        if (shouldBufferPotentialReasoningPrefix(activeTextBuffer)) {
-          textBufferMode = 'pending'
-          continue
-        }
-        yield {
-          type: 'content_block_delta',
-          index: activeTextBlockIndex,
-          delta: {
-            type: 'text_delta',
-            text: payload.delta ?? '',
-          },
         }
       }
       continue
@@ -969,7 +940,7 @@ export function convertCodexResponseToAnthropicMessage(
         if (part?.type === 'output_text') {
           content.push({
             type: 'text',
-            text: stripLeakedReasoningPreamble(part.text ?? ''),
+            text: stripThinkTags(part.text ?? ''),
           })
         }
       }

src/services/api/compressToolHistory.test.ts

@@ -0,0 +1,572 @@
+import { afterEach, beforeEach, expect, mock, test } from 'bun:test'
+import { compressToolHistory, getTiers } from './compressToolHistory.js'
+
+// Mock the two dependencies so tests are deterministic and don't read disk config.
+const mockState = {
+  enabled: true,
+  effectiveWindow: 100_000,
+}
+
+mock.module('../../utils/config.js', () => ({
+  getGlobalConfig: () => ({
+    toolHistoryCompressionEnabled: mockState.enabled,
+  }),
+}))
+
+mock.module('../compact/autoCompact.js', () => ({
+  getEffectiveContextWindowSize: () => mockState.effectiveWindow,
+}))
+
+beforeEach(() => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 100_000
+})
+
+afterEach(() => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 100_000
+})
+
+type Block = Record<string, unknown>
+type Msg = { role: string; content: Block[] | string }
+
+function bigText(n: number): string {
+  return 'x'.repeat(n)
+}
+
+function buildToolExchange(id: number, resultLength: number): Msg[] {
+  return [
+    {
+      role: 'assistant',
+      content: [
+        {
+          type: 'tool_use',
+          id: `toolu_${id}`,
+          name: 'Read',
+          input: { file_path: `/path/to/file${id}.ts` },
+        },
+      ],
+    },
+    {
+      role: 'user',
+      content: [
+        {
+          type: 'tool_result',
+          tool_use_id: `toolu_${id}`,
+          content: bigText(resultLength),
+        },
+      ],
+    },
+  ]
+}
+
+function buildConversation(numToolExchanges: number, resultLength = 5_000): Msg[] {
+  const out: Msg[] = [{ role: 'user', content: 'Initial request' }]
+  for (let i = 0; i < numToolExchanges; i++) {
+    out.push(...buildToolExchange(i, resultLength))
+  }
+  return out
+}
+
+function getResultMessages(messages: Msg[]): Msg[] {
+  return messages.filter(
+    m => Array.isArray(m.content) && m.content.some((b: any) => b.type === 'tool_result'),
+  )
+}
+
+function getResultBlock(msg: Msg): Block {
+  return (msg.content as Block[]).find((b: any) => b.type === 'tool_result') as Block
+}
+
+function getResultText(msg: Msg): string {
+  const block = getResultBlock(msg)
+  const c = block.content
+  if (typeof c === 'string') return c
+  if (Array.isArray(c)) {
+    return c
+      .filter((b: any) => b.type === 'text')
+      .map((b: any) => b.text)
+      .join('\n')
+  }
+  return ''
+}
+
+// ---------- getTiers ----------
+
+test('getTiers: < 16k window → recent=2, mid=3', () => {
+  expect(getTiers(8_000)).toEqual({ recent: 2, mid: 3 })
+})
+
+test('getTiers: 16k–32k → recent=3, mid=5', () => {
+  expect(getTiers(20_000)).toEqual({ recent: 3, mid: 5 })
+})
+
+test('getTiers: 32k–64k → recent=4, mid=8', () => {
+  expect(getTiers(48_000)).toEqual({ recent: 4, mid: 8 })
+})
+
+test('getTiers: 64k–128k (Copilot gpt-4o) → recent=5, mid=10', () => {
+  expect(getTiers(100_000)).toEqual({ recent: 5, mid: 10 })
+})
+
+test('getTiers: 128k–256k (Copilot Claude) → recent=8, mid=15', () => {
+  expect(getTiers(200_000)).toEqual({ recent: 8, mid: 15 })
+})
+
+test('getTiers: 256k–500k → recent=12, mid=25', () => {
+  expect(getTiers(400_000)).toEqual({ recent: 12, mid: 25 })
+})
+
+test('getTiers: ≥ 500k (gpt-4.1 1M) → recent=25, mid=50', () => {
+  expect(getTiers(1_000_000)).toEqual({ recent: 25, mid: 50 })
+})
+
+// ---------- master switch ----------
+
+test('pass-through when toolHistoryCompressionEnabled is false', () => {
+  mockState.enabled = false
+  const messages = buildConversation(20)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  expect(result).toBe(messages) // same reference (no transformation)
+})
+
+test('pass-through when total tool_results <= recent tier', () => {
+  // 100k effective → recent=5; only 4 exchanges → no compression
+  const messages = buildConversation(4)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  expect(result).toBe(messages)
+})
+
+// ---------- per-tier behavior ----------
+
+test('recent tier: tool_result content untouched', () => {
+  // 100k effective → recent=5, mid=10. With 6 exchanges, only the oldest is touched.
+  const messages = buildConversation(6, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // Last 5 should be untouched (full 5000 chars)
+  for (let i = resultMsgs.length - 5; i < resultMsgs.length; i++) {
+    expect(getResultText(resultMsgs[i]).length).toBe(5_000)
+  }
+})
+
+test('mid tier: long content truncated to MID_MAX_CHARS with marker', () => {
+  // 100k → recent=5, mid=10. 10 exchanges: 5 recent + 5 mid (none old).
+  const messages = buildConversation(10, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // First 5 are mid tier — should be truncated to ~2000 chars + marker
+  for (let i = 0; i < 5; i++) {
+    const text = getResultText(resultMsgs[i])
+    expect(text).toContain('[…truncated')
+    expect(text).toContain('chars from tool history]')
+    // Should be roughly 2000 chars + marker (under 2200)
+    expect(text.length).toBeLessThan(2_200)
+    expect(text.length).toBeGreaterThan(2_000)
+  }
+})
+
+test('mid tier: short content (< MID_MAX_CHARS) untouched', () => {
+  const messages = buildConversation(10, 500) // 500 < MID_MAX_CHARS
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  for (let i = 0; i < 5; i++) {
+    expect(getResultText(resultMsgs[i])).toBe(bigText(500))
+  }
+})
+
+test('old tier: content replaced with stub [name args={...} → N chars omitted]', () => {
+  // 100k → recent=5, mid=10, old=rest. 20 exchanges → 5 old + 10 mid + 5 recent.
+  const messages = buildConversation(20, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // First 5 are old tier — should be stubs
+  for (let i = 0; i < 5; i++) {
+    const text = getResultText(resultMsgs[i])
+    expect(text).toMatch(/^\[Read args=\{.*\} → 5000 chars omitted\]$/)
+  }
+})
+
+test('old tier: stub args truncated to 200 chars', () => {
+  const longArg = bigText(500)
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    {
+      role: 'assistant',
+      content: [
+        {
+          type: 'tool_use',
+          id: 'toolu_x',
+          name: 'Bash',
+          input: { command: longArg },
+        },
+      ],
+    },
+    {
+      role: 'user',
+      content: [
+        { type: 'tool_result', tool_use_id: 'toolu_x', content: 'output' },
+      ],
+    },
+    // Pad with enough recent exchanges to push the above into old tier
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+  const text = getResultText(resultMsgs[0])
+
+  // Stub format: [Bash args=<json≤200chars> → N chars omitted]
+  // The args portion (between args= and →) must be ≤ 200 chars.
+  const argsMatch = text.match(/args=(.*?) →/)
+  expect(argsMatch).not.toBeNull()
+  expect(argsMatch![1].length).toBeLessThanOrEqual(200)
+})
+
+test('old tier: orphan tool_result (no matching tool_use) falls back to "tool"', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    // Orphan: tool_result without matching tool_use in history
+    {
+      role: 'user',
+      content: [
+        { type: 'tool_result', tool_use_id: 'orphan_id', content: 'data' },
+      ],
+    },
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+  const text = getResultText(resultMsgs[0])
+
+  expect(text).toMatch(/^\[tool args=\{\} → 4 chars omitted\]$/)
+})
+
+// ---------- structural preservation ----------
+
+test('tool_use blocks always preserved', () => {
+  const messages = buildConversation(20, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4o')
+
+  const useCount = (msgs: Msg[]) =>
+    msgs.reduce((sum, m) => {
+      if (!Array.isArray(m.content)) return sum
+      return sum + m.content.filter((b: any) => b.type === 'tool_use').length
+    }, 0)
+
+  expect(useCount(result as Msg[])).toBe(useCount(messages))
+})
+
+test('text blocks always preserved', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'first' },
+    {
+      role: 'assistant',
+      content: [
+        { type: 'text', text: 'reasoning before tool' },
+        { type: 'tool_use', id: 'toolu_1', name: 'Read', input: {} },
+      ],
+    },
+    {
+      role: 'user',
+      content: [{ type: 'tool_result', tool_use_id: 'toolu_1', content: bigText(5000) }],
+    },
+    ...buildConversation(20, 5_000).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const assistantMsg = (result as Msg[])[1]
+  const textBlock = (assistantMsg.content as Block[]).find((b: any) => b.type === 'text')
+
+  expect(textBlock).toEqual({ type: 'text', text: 'reasoning before tool' })
+})
+
+test('thinking blocks always preserved', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'first' },
+    {
+      role: 'assistant',
+      content: [
+        { type: 'thinking', thinking: 'internal reasoning', signature: 'sig' },
+        { type: 'tool_use', id: 'toolu_1', name: 'Read', input: {} },
+      ],
+    },
+    {
+      role: 'user',
+      content: [{ type: 'tool_result', tool_use_id: 'toolu_1', content: bigText(5000) }],
+    },
+    ...buildConversation(20, 5_000).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const assistantMsg = (result as Msg[])[1]
+  const thinking = (assistantMsg.content as Block[]).find((b: any) => b.type === 'thinking')
+
+  expect(thinking).toEqual({
+    type: 'thinking',
+    thinking: 'internal reasoning',
+    signature: 'sig',
+  })
+})
+
+test('non-array content (string) handled gracefully', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'plain string content' },
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  expect((result as Msg[])[0].content).toBe('plain string content')
+})
+
+test('empty content array handled gracefully', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: [] },
+    ...buildConversation(20, 100).slice(1),
+  ]
+  expect(() => compressToolHistory(messages, 'gpt-4o')).not.toThrow()
+})
+
+// ---------- message shape compatibility ----------
+
+test('wrapped shape ({ message: { role, content } }) handled', () => {
+  type WrappedMsg = { message: { role: string; content: Block[] | string } }
+  const wrap = (m: Msg): WrappedMsg => ({ message: { role: m.role, content: m.content } })
+  const messages = buildConversation(20, 5_000).map(wrap)
+  const result = compressToolHistory(messages as any, 'gpt-4o')
+
+  // First wrapped tool-result message should have stub content (old tier)
+  const firstResultMsg = (result as WrappedMsg[]).find(
+    m =>
+      Array.isArray(m.message.content) &&
+      m.message.content.some((b: any) => b.type === 'tool_result'),
+  )
+  const block = (firstResultMsg!.message.content as Block[]).find(
+    (b: any) => b.type === 'tool_result',
+  ) as Block
+  const text = ((block.content as Block[])[0] as any).text
+  expect(text).toMatch(/^\[Read args=.*→ 5000 chars omitted\]$/)
+})
+
+test('flat shape ({ role, content }) handled', () => {
+  const messages = buildConversation(20, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  expect(getResultText(resultMsgs[0])).toMatch(/^\[Read args=.*→ 5000 chars omitted\]$/)
+})
+
+// ---------- tier boundary correctness ----------
+
+test('tier boundaries: 6 exchanges → 1 mid + 5 recent (recent=5)', () => {
+  const messages = buildConversation(6, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // Oldest: mid (truncated)
+  expect(getResultText(resultMsgs[0])).toContain('[…truncated')
+  // Last 5: untouched
+  for (let i = 1; i < 6; i++) {
+    expect(getResultText(resultMsgs[i]).length).toBe(5_000)
+  }
+})
+
+test('tier boundaries: 16 exchanges → 1 old + 10 mid + 5 recent', () => {
+  const messages = buildConversation(16, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // Oldest 1: stub (old tier)
+  expect(getResultText(resultMsgs[0])).toMatch(/^\[Read .*chars omitted\]$/)
+  // Next 10: mid (truncated)
+  for (let i = 1; i < 11; i++) {
+    expect(getResultText(resultMsgs[i])).toContain('[…truncated')
+  }
+  // Last 5: untouched
+  for (let i = 11; i < 16; i++) {
+    expect(getResultText(resultMsgs[i]).length).toBe(5_000)
+  }
+})
+
+test('large window (1M) with 30 exchanges: all untouched (recent=25 ≥ 30 - 5)', () => {
+  // ≥500k → recent=25, mid=50. 30 exchanges → 5 mid + 25 recent. None old.
+  mockState.effectiveWindow = 1_000_000
+  const messages = buildConversation(30, 5_000)
+  const result = compressToolHistory(messages, 'gpt-4.1')
+  const resultMsgs = getResultMessages(result)
+
+  // Last 25: untouched
+  for (let i = 5; i < 30; i++) {
+    expect(getResultText(resultMsgs[i]).length).toBe(5_000)
+  }
+})
+
+// ---------- attribute preservation ----------
+
+test('is_error flag preserved in mid tier', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    {
+      role: 'assistant',
+      content: [{ type: 'tool_use', id: 'toolu_err', name: 'Bash', input: {} }],
+    },
+    {
+      role: 'user',
+      content: [
+        {
+          type: 'tool_result',
+          tool_use_id: 'toolu_err',
+          is_error: true,
+          content: bigText(5_000),
+        },
+      ],
+    },
+    // Pad with enough recent exchanges to push the above into MID tier
+    ...buildConversation(10, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+  const block = getResultBlock(resultMsgs[0]) as { is_error?: boolean; content: unknown }
+
+  expect(block.is_error).toBe(true)
+  expect(getResultText(resultMsgs[0])).toContain('[…truncated')
+})
+
+test('is_error flag preserved in old tier (stub)', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    {
+      role: 'assistant',
+      content: [{ type: 'tool_use', id: 'toolu_err', name: 'Bash', input: {} }],
+    },
+    {
+      role: 'user',
+      content: [
+        {
+          type: 'tool_result',
+          tool_use_id: 'toolu_err',
+          is_error: true,
+          content: bigText(5_000),
+        },
+      ],
+    },
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+  const block = getResultBlock(resultMsgs[0]) as { is_error?: boolean; content: unknown }
+
+  expect(block.is_error).toBe(true)
+  expect(getResultText(resultMsgs[0])).toMatch(/^\[Bash .*chars omitted\]$/)
+})
+
+// ---------- COMPACTABLE_TOOLS filter ----------
+
+test('non-compactable tool (e.g. Task/Agent) is NEVER compressed', () => {
+  // Build conversation where the OLDEST exchange uses a non-compactable tool name
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    {
+      role: 'assistant',
+      content: [
+        { type: 'tool_use', id: 'task_1', name: 'Task', input: { goal: 'plan' } },
+      ],
+    },
+    {
+      role: 'user',
+      content: [
+        { type: 'tool_result', tool_use_id: 'task_1', content: bigText(5_000) },
+      ],
+    },
+    // Pad with 20 compactable exchanges to push Task into old tier
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // First tool_result is for Task (non-compactable) → must remain full
+  expect(getResultText(resultMsgs[0]).length).toBe(5_000)
+  expect(getResultText(resultMsgs[0])).not.toContain('chars omitted')
+  expect(getResultText(resultMsgs[0])).not.toContain('[…truncated')
+})
+
+test('mcp__ prefixed tools ARE compactable (matches microCompact behavior)', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    {
+      role: 'assistant',
+      content: [
+        { type: 'tool_use', id: 'mcp_1', name: 'mcp__github__get_issue', input: {} },
+      ],
+    },
+    {
+      role: 'user',
+      content: [
+        { type: 'tool_result', tool_use_id: 'mcp_1', content: bigText(5_000) },
+      ],
+    },
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // MCP tool result is compressed (gets stub since it's in old tier)
+  expect(getResultText(resultMsgs[0])).toMatch(/^\[mcp__github__get_issue .*chars omitted\]$/)
+})
+
+// ---------- skip already-cleared blocks ----------
+
+test('blocks already cleared by microCompact are NOT re-compressed', () => {
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    {
+      role: 'assistant',
+      content: [{ type: 'tool_use', id: 'cleared_1', name: 'Read', input: {} }],
+    },
+    {
+      role: 'user',
+      content: [
+        {
+          type: 'tool_result',
+          tool_use_id: 'cleared_1',
+          content: '[Old tool result content cleared]', // microCompact's marker
+        },
+      ],
+    },
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+
+  // Already-cleared marker survives untouched (no double processing)
+  expect(getResultText(resultMsgs[0])).toBe('[Old tool result content cleared]')
+})
+
+test('extra block attributes (e.g. cache_control) preserved across rewrites', () => {
+  const cacheControl = { type: 'ephemeral' }
+  const messages: Msg[] = [
+    { role: 'user', content: 'start' },
+    {
+      role: 'assistant',
+      content: [{ type: 'tool_use', id: 'toolu_cc', name: 'Read', input: {} }],
+    },
+    {
+      role: 'user',
+      content: [
+        {
+          type: 'tool_result',
+          tool_use_id: 'toolu_cc',
+          cache_control: cacheControl,
+          content: bigText(5_000),
+        },
+      ],
+    },
+    ...buildConversation(20, 100).slice(1),
+  ]
+  const result = compressToolHistory(messages, 'gpt-4o')
+  const resultMsgs = getResultMessages(result)
+  const block = getResultBlock(resultMsgs[0]) as { cache_control?: unknown }
+
+  // The custom attribute survived the stub rewrite via ...block spread
+  expect(block.cache_control).toEqual(cacheControl)
+})

src/services/api/compressToolHistory.ts

@@ -0,0 +1,255 @@
+/**
+ * Compresses old tool_result content for stateless OpenAI-compatible providers
+ * (Copilot, Mistral, Ollama). Preserves all conversation structure — tool_use,
+ * tool_result pairing, text, thinking, and is_error all survive intact. Only
+ * the BULK text of older tool_results is shrunk to delay context saturation.
+ *
+ * Tier sizes scale with the model's effective context window via
+ * getEffectiveContextWindowSize() — same calculation used by auto-compact, so
+ * the two systems stay aligned.
+ *
+ * Complements (does not replace) microCompact.ts:
+ * - microCompact: time/cache-based, runs from query.ts, binary clear/keep,
+ *   limited to Claude (cache editing) or idle gaps (time-based).
+ * - compressToolHistory: size-based, runs at the shim layer, tiered
+ *   compression, covers the gap for active sessions on non-Claude providers.
+ *
+ * Reuses isCompactableTool from microCompact to avoid touching tools the
+ * project already classifies as unsafe to compress (e.g. Task, Agent).
+ * Skips blocks already cleared by microCompact (TOOL_RESULT_CLEARED_MESSAGE).
+ *
+ * Anthropic native bypasses both shims, so it is unaffected by this module.
+ */
+import { getEffectiveContextWindowSize } from '../compact/autoCompact.js'
+import { isCompactableTool } from '../compact/microCompact.js'
+import { TOOL_RESULT_CLEARED_MESSAGE } from '../../utils/toolResultStorage.js'
+import { getGlobalConfig } from '../../utils/config.js'
+
+// Mid-tier truncation budget. 2k chars ≈ 500 tokens, enough to preserve the
+// shape of most tool outputs (file headers, command stderr, top grep hits)
+// without ballooning context. Bump too high and the tier loses its purpose.
+const MID_MAX_CHARS = 2_000
+
+// Stub args budget. JSON.stringify of a typical tool input fits in 200 chars
+// (file paths, short commands, small queries). Long inputs are rare and clamping
+// here keeps the stub size bounded even when callers pass oversized arguments.
+const STUB_ARGS_MAX_CHARS = 200
+
+type AnyMessage = {
+  role?: string
+  message?: { role?: string; content?: unknown }
+  content?: unknown
+}
+
+type ToolResultBlock = {
+  type: 'tool_result'
+  tool_use_id?: string
+  is_error?: boolean
+  content?: unknown
+}
+
+type ToolUseBlock = {
+  type: 'tool_use'
+  id?: string
+  name?: string
+  input?: unknown
+}
+
+type Tiers = { recent: number; mid: number }
+
+// Tier sizes scale with effective window. Targets roughly:
+// - recent tier stays under ~25% of available window (full fidelity kept)
+// - recent + mid tier stays under ~50% of available window (bounded bulk)
+// - everything older collapses to ~15-token stubs
+// Values assume ~5KB avg tool_result, which matches the Copilot default case
+// (parallel_tool_calls=true means multiple Read/Bash outputs per turn). For
+// ≥ 500k models the tiers are so generous that compression is effectively
+// inert for any realistic session — see compressToolHistory.test.ts.
+export function getTiers(effectiveWindow: number): Tiers {
+  if (effectiveWindow < 16_000) return { recent: 2, mid: 3 }
+  if (effectiveWindow < 32_000) return { recent: 3, mid: 5 }
+  if (effectiveWindow < 64_000) return { recent: 4, mid: 8 }
+  if (effectiveWindow < 128_000) return { recent: 5, mid: 10 }
+  if (effectiveWindow < 256_000) return { recent: 8, mid: 15 }
+  if (effectiveWindow < 500_000) return { recent: 12, mid: 25 }
+  return { recent: 25, mid: 50 }
+}
+
+function extractText(content: unknown): string {
+  if (typeof content === 'string') return content
+  if (Array.isArray(content)) {
+    return content
+      .filter(
+        (b: { type?: string; text?: string }) =>
+          b?.type === 'text' && typeof b.text === 'string',
+      )
+      .map((b: { text?: string }) => b.text ?? '')
+      .join('\n')
+  }
+  return ''
+}
+
+// Old-tier compression strategy. Replaces content entirely with a one-line
+// metadata marker ~10× more token-efficient than a 500-char truncation AND
+// unambiguous — partial truncations can look authoritative to the model. The
+// stub format encodes tool name + args so the model can re-invoke the same
+// tool if it needs the omitted output back.
+function buildStub(
+  block: ToolResultBlock,
+  toolUsesById: Map<string, ToolUseBlock>,
+): ToolResultBlock {
+  const original = extractText(block.content)
+  const toolUse = toolUsesById.get(block.tool_use_id ?? '')
+  const name = toolUse?.name ?? 'tool'
+  const args = toolUse?.input
+    ? JSON.stringify(toolUse.input).slice(0, STUB_ARGS_MAX_CHARS)
+    : '{}'
+  return {
+    ...block,
+    content: [
+      {
+        type: 'text',
+        text: `[${name} args=${args} → ${original.length} chars omitted]`,
+      },
+    ],
+  }
+}
+
+// Mid-tier compression. The trailing marker is load-bearing: without it, the
+// model can't distinguish "tool returned 2000 chars" from "tool returned 20k
+// chars that we cut to 2000". Distinguishing those matters for the model's
+// decision to re-invoke the tool.
+function truncateBlock(
+  block: ToolResultBlock,
+  maxChars: number,
+): ToolResultBlock {
+  const text = extractText(block.content)
+  if (text.length <= maxChars) return block
+  const omitted = text.length - maxChars
+  return {
+    ...block,
+    content: [
+      {
+        type: 'text',
+        text: `${text.slice(0, maxChars)}\n[…truncated ${omitted} chars from tool history]`,
+      },
+    ],
+  }
+}
+
+function getInner(msg: AnyMessage): { role?: string; content?: unknown } {
+  return (msg.message ?? msg) as { role?: string; content?: unknown }
+}
+
+function indexToolUses(messages: AnyMessage[]): Map<string, ToolUseBlock> {
+  const map = new Map<string, ToolUseBlock>()
+  for (const msg of messages) {
+    const content = getInner(msg).content
+    if (!Array.isArray(content)) continue
+    for (const b of content as Array<{ type?: string; id?: string }>) {
+      if (b?.type === 'tool_use' && b.id) {
+        map.set(b.id, b as ToolUseBlock)
+      }
+    }
+  }
+  return map
+}
+
+function indexToolResultMessages(messages: AnyMessage[]): number[] {
+  const indices: number[] = []
+  for (let i = 0; i < messages.length; i++) {
+    const inner = getInner(messages[i])
+    const role = inner.role ?? messages[i].role
+    const content = inner.content
+    if (
+      role === 'user' &&
+      Array.isArray(content) &&
+      content.some((b: { type?: string }) => b?.type === 'tool_result')
+    ) {
+      indices.push(i)
+    }
+  }
+  return indices
+}
+
+function rewriteMessage<T extends AnyMessage>(
+  msg: T,
+  newContent: unknown[],
+): T {
+  if (msg.message) {
+    return { ...msg, message: { ...msg.message, content: newContent } }
+  }
+  return { ...msg, content: newContent }
+}
+
+// microCompact.maybeTimeBasedMicrocompact may have already replaced old
+// tool_result content with TOOL_RESULT_CLEARED_MESSAGE before we see it.
+// Re-compressing produces a stub over a marker (e.g. `[Read args={} → 40
+// chars omitted]`), wasteful and less informative than the canonical marker.
+function isAlreadyCleared(block: ToolResultBlock): boolean {
+  const text = extractText(block.content)
+  return text === TOOL_RESULT_CLEARED_MESSAGE
+}
+
+function shouldCompressBlock(
+  block: ToolResultBlock,
+  toolUsesById: Map<string, ToolUseBlock>,
+): boolean {
+  if (isAlreadyCleared(block)) return false
+  const toolUse = toolUsesById.get(block.tool_use_id ?? '')
+  // Unknown tool name (orphan tool_result with no matching tool_use) falls
+  // through to compression with a generic "tool" stub. Safer default: the
+  // original tool_use vanished so there's no downstream use for the output.
+  if (!toolUse?.name) return true
+  // Respect microCompact's curated safe-to-compress set (Read/Bash/Grep/…/
+  // mcp__*) so user-facing flow tools (Task, Agent, custom) stay intact.
+  return isCompactableTool(toolUse.name)
+}
+
+export function compressToolHistory<T extends AnyMessage>(
+  messages: T[],
+  model: string,
+): T[] {
+  // Master kill-switch. Returns the original reference so callers skip a
+  // defensive copy when the feature is disabled.
+  if (!getGlobalConfig().toolHistoryCompressionEnabled) return messages
+
+  const tiers = getTiers(getEffectiveContextWindowSize(model))
+
+  const toolResultIndices = indexToolResultMessages(messages)
+  const total = toolResultIndices.length
+  // If every tool-result fits in the recent tier, no boundary crosses; return
+  // the same reference for the same copy-elision reason.
+  if (total <= tiers.recent) return messages
+
+  // O(1) lookup: messageIndex → tool-result position (0 = oldest). Replaces
+  // the naive Array.indexOf(i) that was O(n²) across the .map below.
+  const positionByIndex = new Map<number, number>()
+  for (let pos = 0; pos < toolResultIndices.length; pos++) {
+    positionByIndex.set(toolResultIndices[pos], pos)
+  }
+
+  const toolUsesById = indexToolUses(messages)
+
+  return messages.map((msg, i) => {
+    const pos = positionByIndex.get(i)
+    if (pos === undefined) return msg
+
+    const fromEnd = total - 1 - pos
+    if (fromEnd < tiers.recent) return msg
+
+    const inMidWindow = fromEnd < tiers.recent + tiers.mid
+    const content = getInner(msg).content as unknown[]
+    const newContent = content.map(block => {
+      const b = block as { type?: string }
+      if (b?.type !== 'tool_result') return block
+      const tr = block as ToolResultBlock
+      if (!shouldCompressBlock(tr, toolUsesById)) return block
+      return inMidWindow
+        ? truncateBlock(tr, MID_MAX_CHARS)
+        : buildStub(tr, toolUsesById)
+    })
+
+    return rewriteMessage(msg, newContent)
+  })
+}

src/services/api/openaiErrorClassification.ts

@@ -320,10 +320,7 @@ export function classifyOpenAIHttpFailure(options: {
     }
   }
 
-  if (
-    (options.status >= 200 && options.status < 300 && isMalformedProviderResponse(body)) ||
-    (options.status >= 400 && isMalformedProviderResponse(body))
-  ) {
+  if (options.status >= 400 && isMalformedProviderResponse(body)) {
     return {
       source: 'http',
       category: 'malformed_provider_response',

src/services/api/openaiShim.compression.test.ts

@@ -0,0 +1,317 @@
+import { afterEach, beforeEach, expect, mock, test } from 'bun:test'
+import { createOpenAIShimClient } from './openaiShim.js'
+
+type FetchType = typeof globalThis.fetch
+const originalFetch = globalThis.fetch
+
+const originalEnv = {
+  OPENAI_BASE_URL: process.env.OPENAI_BASE_URL,
+  OPENAI_API_KEY: process.env.OPENAI_API_KEY,
+  OPENAI_MODEL: process.env.OPENAI_MODEL,
+}
+
+// Mock config + autoCompact so the shim sees deterministic state.
+const mockState = {
+  enabled: true,
+  effectiveWindow: 100_000, // Copilot gpt-4o tier
+}
+
+mock.module('../../utils/config.js', () => ({
+  getGlobalConfig: () => ({
+    toolHistoryCompressionEnabled: mockState.enabled,
+    autoCompactEnabled: false,
+  }),
+}))
+
+mock.module('../compact/autoCompact.js', () => ({
+  getEffectiveContextWindowSize: () => mockState.effectiveWindow,
+}))
+
+type OpenAIShimClient = {
+  beta: {
+    messages: {
+      create: (
+        params: Record<string, unknown>,
+        options?: Record<string, unknown>,
+      ) => Promise<unknown>
+    }
+  }
+}
+
+function bigText(n: number): string {
+  return 'A'.repeat(n)
+}
+
+function buildToolExchange(id: number, resultLength: number) {
+  return [
+    {
+      role: 'assistant',
+      content: [
+        {
+          type: 'tool_use',
+          id: `toolu_${id}`,
+          name: 'Read',
+          input: { file_path: `/path/to/file${id}.ts` },
+        },
+      ],
+    },
+    {
+      role: 'user',
+      content: [
+        {
+          type: 'tool_result',
+          tool_use_id: `toolu_${id}`,
+          content: bigText(resultLength),
+        },
+      ],
+    },
+  ]
+}
+
+function buildLongConversation(numExchanges: number, resultLength = 5_000) {
+  const out: Array<{ role: string; content: unknown }> = [
+    { role: 'user', content: 'start the work' },
+  ]
+  for (let i = 0; i < numExchanges; i++) {
+    out.push(...buildToolExchange(i, resultLength))
+  }
+  return out
+}
+
+function makeFakeResponse(): Response {
+  return new Response(
+    JSON.stringify({
+      id: 'chatcmpl-1',
+      model: 'gpt-4o',
+      choices: [
+        {
+          message: { role: 'assistant', content: 'done' },
+          finish_reason: 'stop',
+        },
+      ],
+      usage: { prompt_tokens: 8, completion_tokens: 2, total_tokens: 10 },
+    }),
+    { headers: { 'Content-Type': 'application/json' } },
+  )
+}
+
+beforeEach(() => {
+  process.env.OPENAI_BASE_URL = 'http://example.test/v1'
+  process.env.OPENAI_API_KEY = 'test-key'
+  delete process.env.OPENAI_MODEL
+  mockState.enabled = true
+  mockState.effectiveWindow = 100_000
+})
+
+afterEach(() => {
+  if (originalEnv.OPENAI_BASE_URL === undefined) delete process.env.OPENAI_BASE_URL
+  else process.env.OPENAI_BASE_URL = originalEnv.OPENAI_BASE_URL
+  if (originalEnv.OPENAI_API_KEY === undefined) delete process.env.OPENAI_API_KEY
+  else process.env.OPENAI_API_KEY = originalEnv.OPENAI_API_KEY
+  if (originalEnv.OPENAI_MODEL === undefined) delete process.env.OPENAI_MODEL
+  else process.env.OPENAI_MODEL = originalEnv.OPENAI_MODEL
+  globalThis.fetch = originalFetch
+})
+
+async function captureRequestBody(
+  messages: Array<{ role: string; content: unknown }>,
+  model: string,
+): Promise<Record<string, unknown>> {
+  let captured: Record<string, unknown> | undefined
+
+  globalThis.fetch = (async (_input, init) => {
+    captured = JSON.parse(String(init?.body))
+    return makeFakeResponse()
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+  await client.beta.messages.create({
+    model,
+    system: 'system prompt',
+    messages,
+  })
+
+  if (!captured) throw new Error('request not captured')
+  return captured
+}
+
+function getToolMessages(body: Record<string, unknown>): Array<{ content: string }> {
+  const messages = body.messages as Array<{ role: string; content: string }>
+  return messages.filter(m => m.role === 'tool')
+}
+
+function getAssistantToolCalls(body: Record<string, unknown>): unknown[] {
+  const messages = body.messages as Array<{
+    role: string
+    tool_calls?: unknown[]
+  }>
+  return messages
+    .filter(m => m.role === 'assistant' && Array.isArray(m.tool_calls))
+    .flatMap(m => m.tool_calls ?? [])
+}
+
+// ============================================================================
+// BUG REPRO: without compression, full tool history is resent every turn
+// ============================================================================
+
+test('BUG REPRO: without compression, all 30 tool results are sent at full size', async () => {
+  mockState.enabled = false
+  const messages = buildLongConversation(30, 5_000)
+
+  const body = await captureRequestBody(messages, 'gpt-4o')
+  const toolMessages = getToolMessages(body)
+  const payloadSize = JSON.stringify(body).length
+
+  // All 30 tool results present, none truncated
+  expect(toolMessages.length).toBe(30)
+  for (const m of toolMessages) {
+    expect(m.content.length).toBeGreaterThanOrEqual(5_000)
+    expect(m.content).not.toContain('[…truncated')
+    expect(m.content).not.toContain('chars omitted')
+  }
+
+  // Total payload is large (~150KB raw) — this is the cost being paid every turn
+  expect(payloadSize).toBeGreaterThan(150_000)
+})
+
+// ============================================================================
+// FIX: with compression, recent kept full, mid truncated, old stubbed
+// ============================================================================
+
+test('FIX: with compression on Copilot gpt-4o (tier 5/10/rest), 30 turns shrinks dramatically', async () => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 100_000 // 64–128k → recent=5, mid=10
+  const messages = buildLongConversation(30, 5_000)
+
+  const body = await captureRequestBody(messages, 'gpt-4o')
+  const toolMessages = getToolMessages(body)
+  const payloadSize = JSON.stringify(body).length
+
+  // Structure preserved: still 30 tool messages, no orphan tool_calls
+  expect(toolMessages.length).toBe(30)
+  expect(getAssistantToolCalls(body).length).toBe(30)
+
+  // Tier breakdown (oldest → newest):
+  //   indices 0..14  → old tier (stubs)
+  //   indices 15..24 → mid tier (truncated)
+  //   indices 25..29 → recent (full)
+  for (let i = 0; i <= 14; i++) {
+    expect(toolMessages[i].content).toMatch(/^\[Read args=.*chars omitted\]$/)
+  }
+  for (let i = 15; i <= 24; i++) {
+    expect(toolMessages[i].content).toContain('[…truncated')
+  }
+  for (let i = 25; i <= 29; i++) {
+    expect(toolMessages[i].content.length).toBe(5_000)
+    expect(toolMessages[i].content).not.toContain('[…truncated')
+    expect(toolMessages[i].content).not.toContain('chars omitted')
+  }
+
+  // Significant reduction: from ~150KB to <60KB (10 mid×2KB + structure overhead)
+  expect(payloadSize).toBeLessThan(60_000)
+})
+
+// ============================================================================
+// FIX: large-context model gets generous tiers — compression effectively inert
+// ============================================================================
+
+test('FIX: gpt-4.1 (1M context) with 25 exchanges keeps all full (recent tier=25)', async () => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 1_000_000 // ≥500k → recent=25, mid=50
+  const messages = buildLongConversation(25, 5_000)
+
+  const body = await captureRequestBody(messages, 'gpt-4.1')
+  const toolMessages = getToolMessages(body)
+
+  expect(toolMessages.length).toBe(25)
+  for (const m of toolMessages) {
+    expect(m.content.length).toBe(5_000)
+    expect(m.content).not.toContain('[…truncated')
+    expect(m.content).not.toContain('chars omitted')
+  }
+})
+
+test('FIX: gpt-4.1 (1M context) with 30 exchanges → only first 5 mid-truncated', async () => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 1_000_000 // recent=25, mid=50
+  const messages = buildLongConversation(30, 5_000)
+
+  const body = await captureRequestBody(messages, 'gpt-4.1')
+  const toolMessages = getToolMessages(body)
+
+  // 30 total: indices 0..4 mid, indices 5..29 recent
+  for (let i = 0; i < 5; i++) {
+    expect(toolMessages[i].content).toContain('[…truncated')
+  }
+  for (let i = 5; i < 30; i++) {
+    expect(toolMessages[i].content.length).toBe(5_000)
+  }
+})
+
+// ============================================================================
+// FIX: stub preserves tool name and args — model can re-invoke if needed
+// ============================================================================
+
+test('FIX: stub format includes original tool name and arguments', async () => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 100_000
+  const messages = buildLongConversation(30, 5_000)
+
+  const body = await captureRequestBody(messages, 'gpt-4o')
+  const toolMessages = getToolMessages(body)
+  const oldestStub = toolMessages[0].content
+
+  // Format: [<tool_name> args=<json> → <N> chars omitted]
+  expect(oldestStub).toMatch(/^\[Read /)
+  expect(oldestStub).toMatch(/file_path/)
+  expect(oldestStub).toMatch(/→ 5000 chars omitted\]$/)
+})
+
+// ============================================================================
+// FIX: tool_use blocks (assistant tool_calls) are never modified
+// ============================================================================
+
+test('FIX: every tool_call retains its full id, name, and arguments', async () => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 100_000
+  const messages = buildLongConversation(30, 5_000)
+
+  const body = await captureRequestBody(messages, 'gpt-4o')
+  const toolCalls = getAssistantToolCalls(body) as Array<{
+    id: string
+    function: { name: string; arguments: string }
+  }>
+
+  expect(toolCalls.length).toBe(30)
+  for (let i = 0; i < toolCalls.length; i++) {
+    expect(toolCalls[i].id).toBe(`toolu_${i}`)
+    expect(toolCalls[i].function.name).toBe('Read')
+    expect(JSON.parse(toolCalls[i].function.arguments)).toEqual({
+      file_path: `/path/to/file${i}.ts`,
+    })
+  }
+})
+
+// ============================================================================
+// FIX: small-context provider (Mistral 32k) gets aggressive compression
+// ============================================================================
+
+test('FIX: 32k window (Mistral tier) → recent=3 keeps last 3 only', async () => {
+  mockState.enabled = true
+  mockState.effectiveWindow = 24_000 // 16–32k → recent=3, mid=5
+  const messages = buildLongConversation(15, 3_000)
+
+  const body = await captureRequestBody(messages, 'mistral-large-latest')
+  const toolMessages = getToolMessages(body)
+
+  // 15 total: indices 0..6 old, 7..11 mid, 12..14 recent
+  for (let i = 0; i <= 6; i++) {
+    expect(toolMessages[i].content).toContain('chars omitted')
+  }
+  for (let i = 7; i <= 11; i++) {
+    expect(toolMessages[i].content).toContain('[…truncated')
+  }
+  for (let i = 12; i <= 14; i++) {
+    expect(toolMessages[i].content.length).toBe(3_000)
+  }
+})

src/services/api/openaiShim.diagnostics.test.ts

@@ -117,3 +117,170 @@ test('redacts credentials in transport diagnostic URL logs', async () => {
   expect(logLine).not.toContain('user:supersecret')
   expect(logLine).not.toContain('supersecret@')
 })
+test('logs self-heal localhost fallback with redacted from/to URLs', async () => {
+  const debugSpy = mock(() => {})
+  mock.module('../../utils/debug.js', () => ({
+    logForDebugging: debugSpy,
+  }))
+
+  const nonce = `${Date.now()}-${Math.random()}`
+  const { createOpenAIShimClient } = await import(`./openaiShim.ts?ts=${nonce}`)
+
+  process.env.OPENAI_BASE_URL = 'http://user:supersecret@localhost:11434/v1'
+  process.env.OPENAI_API_KEY = 'supersecret'
+
+  globalThis.fetch = mock(async (input: string | Request) => {
+    const url = typeof input === 'string' ? input : input.url
+    if (url.includes('localhost')) {
+      throw Object.assign(new TypeError('fetch failed'), {
+        code: 'ENOTFOUND',
+      })
+    }
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'qwen2.5-coder:7b',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'ok',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 5,
+          completion_tokens: 2,
+          total_tokens: 7,
+        },
+      }),
+      {
+        status: 200,
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as typeof globalThis.fetch
+
+  const client = createOpenAIShimClient({}) as {
+    beta: {
+      messages: {
+        create: (params: Record<string, unknown>) => Promise<unknown>
+      }
+    }
+  }
+
+  await expect(
+    client.beta.messages.create({
+      model: 'qwen2.5-coder:7b',
+      messages: [{ role: 'user', content: 'hello' }],
+      max_tokens: 64,
+      stream: false,
+    }),
+  ).resolves.toBeDefined()
+
+  const fallbackLog = debugSpy.mock.calls.find(call =>
+    typeof call?.[0] === 'string' &&
+    call[0].includes('self-heal retry reason=localhost_resolution_failed'),
+  )
+
+  expect(fallbackLog).toBeDefined()
+  const logLine = String(fallbackLog?.[0])
+  expect(logLine).toContain('from=http://redacted:redacted@localhost:11434/v1/chat/completions')
+  expect(logLine).toContain('to=http://redacted:redacted@127.0.0.1:11434/v1/chat/completions')
+  expect(logLine).not.toContain('supersecret')
+})
+
+test('logs self-heal toolless retry for local tool-call incompatibility', async () => {
+  const debugSpy = mock(() => {})
+  mock.module('../../utils/debug.js', () => ({
+    logForDebugging: debugSpy,
+  }))
+
+  const nonce = `${Date.now()}-${Math.random()}`
+  const { createOpenAIShimClient } = await import(`./openaiShim.ts?ts=${nonce}`)
+
+  process.env.OPENAI_BASE_URL = 'http://localhost:11434/v1'
+  process.env.OPENAI_API_KEY = 'ollama'
+
+  let callCount = 0
+  globalThis.fetch = mock(async () => {
+    callCount += 1
+    if (callCount === 1) {
+      return new Response('tool_calls are not supported', {
+        status: 400,
+        headers: {
+          'Content-Type': 'text/plain',
+        },
+      })
+    }
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'qwen2.5-coder:7b',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'ok',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 7,
+          completion_tokens: 3,
+          total_tokens: 10,
+        },
+      }),
+      {
+        status: 200,
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as typeof globalThis.fetch
+
+  const client = createOpenAIShimClient({}) as {
+    beta: {
+      messages: {
+        create: (params: Record<string, unknown>) => Promise<unknown>
+      }
+    }
+  }
+
+  await expect(
+    client.beta.messages.create({
+      model: 'qwen2.5-coder:7b',
+      messages: [{ role: 'user', content: 'hello' }],
+      tools: [
+        {
+          name: 'Read',
+          description: 'Read file',
+          input_schema: {
+            type: 'object',
+            properties: {
+              filePath: { type: 'string' },
+            },
+            required: ['filePath'],
+          },
+        },
+      ],
+      max_tokens: 64,
+      stream: false,
+    }),
+  ).resolves.toBeDefined()
+
+  const fallbackLog = debugSpy.mock.calls.find(call =>
+    typeof call?.[0] === 'string' &&
+    call[0].includes('self-heal retry reason=tool_call_incompatible mode=toolless'),
+  )
+
+  expect(fallbackLog).toBeDefined()
+  expect(fallbackLog?.[1]).toEqual({ level: 'warn' })
+})

src/services/api/openaiShim.test.ts

@@ -2513,7 +2513,7 @@ test('non-streaming: real content takes precedence over reasoning_content', asyn
   ])
 })
 
-test('non-streaming: strips leaked reasoning preamble from assistant content', async () => {
+test('non-streaming: strips <think> tag block from assistant content', async () => {
   globalThis.fetch = (async () => {
     return new Response(
       JSON.stringify({
@@ -2524,7 +2524,7 @@ test('non-streaming: strips leaked reasoning preamble from assistant content', a
             message: {
               role: 'assistant',
               content:
-                'The user just said "hey" - a simple greeting. I should respond briefly and friendly.\n\nHey! How can I help you today?',
+                '<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?',
             },
             finish_reason: 'stop',
           },
@@ -2645,7 +2645,7 @@ test('streaming: thinking block closed before tool call', async () => {
   expect(thinkingStart?.content_block?.type).toBe('thinking')
 })
 
-test('streaming: strips leaked reasoning preamble from assistant content deltas', async () => {
+test('streaming: strips <think> tag block from assistant content deltas', async () => {
   globalThis.fetch = (async () => {
     const chunks = makeStreamChunks([
       {
@@ -2658,7 +2658,7 @@ test('streaming: strips leaked reasoning preamble from assistant content deltas'
             delta: {
               role: 'assistant',
               content:
-                'The user just said "hey" - a simple greeting. I should respond briefly and friendly.\n\nHey! How can I help you today?',
+                '<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?',
             },
             finish_reason: null,
           },
@@ -2700,10 +2700,10 @@ test('streaming: strips leaked reasoning preamble from assistant content deltas'
     }
   }
 
-  expect(textDeltas).toEqual(['Hey! How can I help you today?'])
+  expect(textDeltas.join('')).toBe('Hey! How can I help you today?')
 })
 
-test('streaming: strips leaked reasoning preamble when split across multiple content chunks', async () => {
+test('streaming: strips <think> tag split across multiple content chunks', async () => {
   globalThis.fetch = (async () => {
     const chunks = makeStreamChunks([
       {
@@ -2715,7 +2715,7 @@ test('streaming: strips leaked reasoning preamble when split across multiple con
             index: 0,
             delta: {
               role: 'assistant',
-              content: 'The user said "hey" - this is a simple greeting. ',
+              content: '<think>user wants a greeting,',
             },
             finish_reason: null,
           },
@@ -2729,8 +2729,21 @@ test('streaming: strips leaked reasoning preamble when split across multiple con
           {
             index: 0,
             delta: {
-              content:
-                'I should respond in a friendly, concise way.\n\nHey! How can I help you today?',
+              content: ' respond briefly</th',
+            },
+            finish_reason: null,
+          },
+        ],
+      },
+      {
+        id: 'chatcmpl-1',
+        object: 'chat.completion.chunk',
+        model: 'gpt-5-mini',
+        choices: [
+          {
+            index: 0,
+            delta: {
+              content: 'ink>Hey! How can I help you today?',
             },
             finish_reason: null,
           },
@@ -2773,7 +2786,69 @@ test('streaming: strips leaked reasoning preamble when split across multiple con
     }
   }
 
-  expect(textDeltas).toEqual(['Hey! How can I help you today?'])
+  expect(textDeltas.join('')).toBe('Hey! How can I help you today?')
+})
+
+test('streaming: preserves prose without tags (no phrase-based false positive)', async () => {
+  // Regression: older phrase-based sanitizer would strip "I should..." prose.
+  // The tag-based approach leaves legitimate assistant output alone.
+  globalThis.fetch = (async () => {
+    const chunks = makeStreamChunks([
+      {
+        id: 'chatcmpl-1',
+        object: 'chat.completion.chunk',
+        model: 'gpt-5-mini',
+        choices: [
+          {
+            index: 0,
+            delta: {
+              role: 'assistant',
+              content:
+                'I should note that the user role requires a briefly concise friendly response format.',
+            },
+            finish_reason: null,
+          },
+        ],
+      },
+      {
+        id: 'chatcmpl-1',
+        object: 'chat.completion.chunk',
+        model: 'gpt-5-mini',
+        choices: [
+          {
+            index: 0,
+            delta: {},
+            finish_reason: 'stop',
+          },
+        ],
+      },
+    ])
+
+    return makeSseResponse(chunks)
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+  const result = await client.beta.messages
+    .create({
+      model: 'gpt-5-mini',
+      system: 'test system',
+      messages: [{ role: 'user', content: 'hey' }],
+      max_tokens: 64,
+      stream: true,
+    })
+    .withResponse()
+
+  const textDeltas: string[] = []
+  for await (const event of result.data) {
+    const delta = (event as { delta?: { type?: string; text?: string } }).delta
+    if (delta?.type === 'text_delta' && typeof delta.text === 'string') {
+      textDeltas.push(delta.text)
+    }
+  }
+
+  expect(textDeltas.join('')).toBe(
+    'I should note that the user role requires a briefly concise friendly response format.',
+  )
 })
 
 test('classifies localhost transport failures with actionable category marker', async () => {
@@ -2856,6 +2931,204 @@ test('classifies chat-completions endpoint 404 failures with endpoint_not_found
     }),
   ).rejects.toThrow('openai_category=endpoint_not_found')
 })
+test('self-heals localhost resolution failures by retrying local loopback base URL', async () => {
+  process.env.OPENAI_BASE_URL = 'http://localhost:11434/v1'
+
+  const requestUrls: string[] = []
+  globalThis.fetch = (async (input, _init) => {
+    const url = typeof input === 'string' ? input : input.url
+    requestUrls.push(url)
+
+    if (url.includes('localhost')) {
+      const error = Object.assign(new TypeError('fetch failed'), {
+        code: 'ENOTFOUND',
+      })
+      throw error
+    }
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'qwen2.5-coder:7b',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'hello from loopback',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 4,
+          completion_tokens: 3,
+          total_tokens: 7,
+        },
+      }),
+      {
+        status: 200,
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await expect(
+    client.beta.messages.create({
+      model: 'qwen2.5-coder:7b',
+      messages: [{ role: 'user', content: 'hello' }],
+      max_tokens: 64,
+      stream: false,
+    }),
+  ).resolves.toBeDefined()
+
+  expect(requestUrls[0]).toBe('http://localhost:11434/v1/chat/completions')
+  expect(requestUrls).toContain('http://127.0.0.1:11434/v1/chat/completions')
+})
+
+test('self-heals local endpoint_not_found by retrying with /v1 base URL', async () => {
+  process.env.OPENAI_BASE_URL = 'http://localhost:11434'
+
+  const requestUrls: string[] = []
+  globalThis.fetch = (async (input, _init) => {
+    const url = typeof input === 'string' ? input : input.url
+    requestUrls.push(url)
+
+    if (url === 'http://localhost:11434/chat/completions') {
+      return new Response('Not Found', {
+        status: 404,
+        headers: {
+          'Content-Type': 'text/plain',
+        },
+      })
+    }
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'qwen2.5-coder:7b',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'hello from /v1',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 5,
+          completion_tokens: 2,
+          total_tokens: 7,
+        },
+      }),
+      {
+        status: 200,
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await expect(
+    client.beta.messages.create({
+      model: 'qwen2.5-coder:7b',
+      messages: [{ role: 'user', content: 'hello' }],
+      max_tokens: 64,
+      stream: false,
+    }),
+  ).resolves.toBeDefined()
+
+  expect(requestUrls).toEqual([
+    'http://localhost:11434/chat/completions',
+    'http://localhost:11434/v1/chat/completions',
+  ])
+})
+
+test('self-heals tool-call incompatibility by retrying local Ollama requests without tools', async () => {
+  process.env.OPENAI_BASE_URL = 'http://localhost:11434/v1'
+
+  const requestBodies: Array<Record<string, unknown>> = []
+  globalThis.fetch = (async (_input, init) => {
+    const requestBody = JSON.parse(String(init?.body)) as Record<string, unknown>
+    requestBodies.push(requestBody)
+
+    if (requestBodies.length === 1) {
+      return new Response('tool_calls are not supported', {
+        status: 400,
+        headers: {
+          'Content-Type': 'text/plain',
+        },
+      })
+    }
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'qwen2.5-coder:7b',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'fallback without tools',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 8,
+          completion_tokens: 4,
+          total_tokens: 12,
+        },
+      }),
+      {
+        status: 200,
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await expect(
+    client.beta.messages.create({
+      model: 'qwen2.5-coder:7b',
+      messages: [{ role: 'user', content: 'hello' }],
+      tools: [
+        {
+          name: 'Read',
+          description: 'Read a file',
+          input_schema: {
+            type: 'object',
+            properties: {
+              filePath: { type: 'string' },
+            },
+            required: ['filePath'],
+          },
+        },
+      ],
+      max_tokens: 64,
+      stream: false,
+    }),
+  ).resolves.toBeDefined()
+
+  expect(requestBodies).toHaveLength(2)
+  expect(Array.isArray(requestBodies[0]?.tools)).toBe(true)
+  expect(requestBodies[0]?.tool_choice).toBeUndefined()
+  expect(
+    requestBodies[1]?.tools === undefined ||
+      (Array.isArray(requestBodies[1]?.tools) && requestBodies[1]?.tools.length === 0),
+  ).toBe(true)
+  expect(requestBodies[1]?.tool_choice).toBeUndefined()
+})
 
 test('preserves valid tool_result and drops orphan tool_result', async () => {
   let requestBody: Record<string, unknown> | undefined
@@ -2924,7 +3197,7 @@ test('preserves valid tool_result and drops orphan tool_result', async () => {
           {
             role: 'user',
             content: 'What happened?',
-          }
+          },
         ],
       },
     ],
@@ -2933,14 +3206,393 @@ test('preserves valid tool_result and drops orphan tool_result', async () => {
   })
 
   const messages = requestBody?.messages as Array<Record<string, unknown>>
-  
+
   // Should have: system, user, assistant (tool_use), tool (valid_call_1), user
   // Should NOT have: tool (orphan_call_2)
-  
+
   const toolMessages = messages.filter(m => m.role === 'tool')
   expect(toolMessages.length).toBe(1)
   expect(toolMessages[0].tool_call_id).toBe('valid_call_1')
-  
+
   const orphanMessage = toolMessages.find(m => m.tool_call_id === 'orphan_call_2')
   expect(orphanMessage).toBeUndefined()
+  
+  // Actually, the semantic message IS injected here because the user block with orphan 
+  // tool result is converted to:
+  // 1. Tool result (valid_call_1) -> role 'tool'
+  // 2. User content ("What happened?") -> role 'user'
+  // This triggers the tool -> assistant injection.
+  const assistantMessages = messages.filter(m => m.role === 'assistant')
+  expect(assistantMessages.some(m => m.content === '[Tool execution interrupted by user]')).toBe(true)
+})
+
+test('drops empty assistant message when only thinking block was present and stripped', async () => {
+  let requestBody: Record<string, unknown> | undefined
+
+  globalThis.fetch = (async (_input, init) => {
+    requestBody = JSON.parse(String(init?.body))
+    return new Response(JSON.stringify({
+      id: 'chatcmpl-1',
+      object: 'chat.completion',
+      created: 123456789,
+      model: 'mistral-large-latest',
+      choices: [{ message: { role: 'assistant', content: 'hi' }, finish_reason: 'stop' }],
+      usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }
+    }), { headers: { 'Content-Type': 'application/json' } })
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await client.beta.messages.create({
+    model: 'mistral-large-latest',
+    messages: [
+      { role: 'user', content: 'Initial' },
+      { role: 'assistant', content: [{ type: 'thinking', thinking: 'I am thinking...', signature: 'sig' }] },
+      { role: 'user', content: 'Interrupting query' },
+    ],
+    max_tokens: 64,
+    stream: false,
+  })
+
+  const messages = requestBody?.messages as Array<Record<string, unknown>>
+  // The assistant msg is dropped because thinking is stripped.
+  // The two user messages are coalesced.
+  expect(messages.length).toBe(1)
+  expect(messages[0].role).toBe('user')
+  expect(String(messages[0].content)).toContain('Initial')
+  expect(String(messages[0].content)).toContain('Interrupting query')
+})
+
+test('injects semantic assistant message when tool result is followed by user message', async () => {
+  let requestBody: Record<string, unknown> | undefined
+
+  globalThis.fetch = (async (_input, init) => {
+    requestBody = JSON.parse(String(init?.body))
+    return new Response(JSON.stringify({
+      id: 'chatcmpl-2',
+      object: 'chat.completion',
+      created: 123456789,
+      model: 'mistral-large-latest',
+      choices: [{ message: { role: 'assistant', content: 'hi' }, finish_reason: 'stop' }],
+      usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }
+    }), { headers: { 'Content-Type': 'application/json' } })
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await client.beta.messages.create({
+    model: 'mistral-large-latest',
+    messages: [
+      { 
+        role: 'assistant', 
+        content: [{ type: 'tool_use', id: 'call_1', name: 'search', input: {} }] 
+      },
+      { 
+        role: 'user', 
+        content: [
+          { type: 'tool_result', tool_use_id: 'call_1', content: 'Result' }
+        ] 
+      },
+      { role: 'user', content: 'Next user query' },
+    ],
+    max_tokens: 64,
+    stream: false,
+  })
+
+  const messages = requestBody?.messages as Array<Record<string, unknown>>
+  // Roles should be: assistant (tool_calls) -> tool -> assistant (semantic) -> user
+  const roles = messages.map(m => m.role)
+  expect(roles).toEqual(['assistant', 'tool', 'assistant', 'user'])
+  
+  const semanticMsg = messages[2]
+  expect(semanticMsg.role).toBe('assistant')
+  expect(semanticMsg.content).toBe('[Tool execution interrupted by user]')
+})
+
+test('Moonshot: uses max_tokens (not max_completion_tokens) and strips store', async () => {
+  process.env.OPENAI_BASE_URL = 'https://api.moonshot.ai/v1'
+  process.env.OPENAI_API_KEY = 'sk-moonshot-test'
+
+  let requestBody: Record<string, unknown> | undefined
+  globalThis.fetch = (async (_input, init) => {
+    requestBody = JSON.parse(String(init?.body))
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'kimi-k2.6',
+        choices: [
+          { message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' },
+        ],
+        usage: { prompt_tokens: 3, completion_tokens: 1, total_tokens: 4 },
+      }),
+      { headers: { 'Content-Type': 'application/json' } },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+  await client.beta.messages.create({
+    model: 'kimi-k2.6',
+    system: 'you are kimi',
+    messages: [{ role: 'user', content: 'hi' }],
+    max_tokens: 256,
+    stream: false,
+  })
+
+  expect(requestBody?.max_tokens).toBe(256)
+  expect(requestBody?.max_completion_tokens).toBeUndefined()
+  expect(requestBody?.store).toBeUndefined()
+})
+
+test('Moonshot: cn host is also detected', async () => {
+  process.env.OPENAI_BASE_URL = 'https://api.moonshot.cn/v1'
+  process.env.OPENAI_API_KEY = 'sk-moonshot-test'
+
+  let requestBody: Record<string, unknown> | undefined
+  globalThis.fetch = (async (_input, init) => {
+    requestBody = JSON.parse(String(init?.body))
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'kimi-k2.6',
+        choices: [
+          { message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' },
+        ],
+        usage: { prompt_tokens: 3, completion_tokens: 1, total_tokens: 4 },
+      }),
+      { headers: { 'Content-Type': 'application/json' } },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+  await client.beta.messages.create({
+    model: 'kimi-k2.6',
+    system: 'you are kimi',
+    messages: [{ role: 'user', content: 'hi' }],
+    max_tokens: 256,
+    stream: false,
+  })
+
+  expect(requestBody?.store).toBeUndefined()
+})
+
+
+test('collapses multiple text blocks in tool_result to string for DeepSeek compatibility (issue #774)', async () => {
+  let requestBody: Record<string, unknown> | undefined
+
+  globalThis.fetch = (async (_input, init) => {
+    requestBody = JSON.parse(String(init?.body))
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'deepseek-reasoner',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'done',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 12,
+          completion_tokens: 4,
+          total_tokens: 16,
+        },
+      }),
+      {
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await client.beta.messages.create({
+    model: 'deepseek-reasoner',
+    system: 'test system',
+    messages: [
+      { role: 'user', content: 'Run ls' },
+      {
+        role: 'assistant',
+        content: [
+          {
+            type: 'tool_use',
+            id: 'call_1',
+            name: 'Bash',
+            input: { command: 'ls' },
+          },
+        ],
+      },
+      {
+        role: 'user',
+        content: [
+          {
+            type: 'tool_result',
+            tool_use_id: 'call_1',
+            content: [
+              { type: 'text', text: 'line one' },
+              { type: 'text', text: 'line two' },
+            ],
+          },
+        ],
+      },
+    ],
+    max_tokens: 64,
+    stream: false,
+  })
+
+  const messages = requestBody?.messages as Array<Record<string, unknown>>
+  const toolMessages = messages.filter(m => m.role === 'tool')
+  expect(toolMessages.length).toBe(1)
+  expect(toolMessages[0].tool_call_id).toBe('call_1')
+  expect(typeof toolMessages[0].content).toBe('string')
+  expect(toolMessages[0].content).toBe('line one\n\nline two')
+})
+
+test('collapses multiple text blocks into a single string for DeepSeek compatibility (issue #774)', async () => {
+  let requestBody: Record<string, unknown> | undefined
+
+  globalThis.fetch = (async (_input, init) => {
+    requestBody = JSON.parse(String(init?.body))
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'deepseek-reasoner',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'done',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 12,
+          completion_tokens: 4,
+          total_tokens: 16,
+        },
+      }),
+      {
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await client.beta.messages.create({
+    model: 'deepseek-reasoner',
+    system: 'test system',
+    messages: [
+      {
+        role: 'user',
+        content: [
+          { type: 'text', text: 'Hello!' },
+          { type: 'text', text: 'How are you?' },
+        ],
+      },
+    ],
+    max_tokens: 64,
+    stream: false,
+  })
+
+  const messages = requestBody?.messages as Array<Record<string, unknown>>
+  expect(messages.length).toBe(2) // system + user
+  expect(messages[1].role).toBe('user')
+  expect(typeof messages[1].content).toBe('string')
+  expect(messages[1].content).toBe('Hello!\n\nHow are you?')
+})
+
+test('preserves mixed text and image tool results as multipart content', async () => {
+  let requestBody: Record<string, unknown> | undefined
+
+  globalThis.fetch = (async (_input, init) => {
+    requestBody = JSON.parse(String(init?.body))
+
+    return new Response(
+      JSON.stringify({
+        id: 'chatcmpl-1',
+        model: 'gpt-4o',
+        choices: [
+          {
+            message: {
+              role: 'assistant',
+              content: 'done',
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 12,
+          completion_tokens: 4,
+          total_tokens: 16,
+        },
+      }),
+      {
+        headers: {
+          'Content-Type': 'application/json',
+        },
+      },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await client.beta.messages.create({
+    model: 'gpt-4o',
+    system: 'test system',
+    messages: [
+      { role: 'user', content: 'Show me' },
+      {
+        role: 'assistant',
+        content: [
+          {
+            type: 'tool_use',
+            id: 'call_1',
+            name: 'Bash',
+            input: { command: 'cat image.png' },
+          },
+        ],
+      },
+      {
+        role: 'user',
+        content: [
+          {
+            type: 'tool_result',
+            tool_use_id: 'call_1',
+            content: [
+              { type: 'text', text: 'Here is the image:' },
+              {
+                type: 'image',
+                source: {
+                  type: 'base64',
+                  media_type: 'image/png',
+                  data: 'iVBORw0KGgo=',
+                },
+              },
+            ],
+          },
+        ],
+      },
+    ],
+    max_tokens: 64,
+    stream: false,
+  })
+
+  const messages = requestBody?.messages as Array<Record<string, unknown>>
+  const toolMessages = messages.filter(m => m.role === 'tool')
+  expect(toolMessages.length).toBe(1)
+  expect(Array.isArray(toolMessages[0].content)).toBe(true)
+  const content = toolMessages[0].content as Array<Record<string, unknown>>
+  expect(content.length).toBe(2)
+  expect(content[0].type).toBe('text')
+  expect(content[1].type).toBe('image_url')
 })

src/services/api/openaiShim.ts

@@ -32,10 +32,9 @@ import { resolveGeminiCredential } from '../../utils/geminiAuth.js'
 import { hydrateGeminiAccessTokenFromSecureStorage } from '../../utils/geminiCredentials.js'
 import { hydrateGithubModelsTokenFromSecureStorage } from '../../utils/githubModelsCredentials.js'
 import {
-  looksLikeLeakedReasoningPrefix,
-  shouldBufferPotentialReasoningPrefix,
-  stripLeakedReasoningPreamble,
-} from './reasoningLeakSanitizer.js'
+  createThinkTagFilter,
+  stripThinkTags,
+} from './thinkTagSanitizer.js'
 import {
   codexStreamToAnthropic,
   collectCodexCompletedResponse,
@@ -47,12 +46,15 @@ import {
   type AnthropicUsage,
   type ShimCreateParams,
 } from './codexShim.js'
+import { compressToolHistory } from './compressToolHistory.js'
 import { fetchWithProxyRetry } from './fetchWithProxyRetry.js'
 import {
+  getLocalProviderRetryBaseUrls,
+  getGithubEndpointType,
   isLocalProviderUrl,
   resolveRuntimeCodexCredentials,
   resolveProviderRequest,
-  getGithubEndpointType,
+  shouldAttemptLocalToollessRetry,
 } from './providerConfig.js'
 import {
   buildOpenAICompatibilityErrorMessage,
@@ -80,6 +82,10 @@ const GITHUB_429_MAX_RETRIES = 3
 const GITHUB_429_BASE_DELAY_SEC = 1
 const GITHUB_429_MAX_DELAY_SEC = 32
 const GEMINI_API_HOST = 'generativelanguage.googleapis.com'
+const MOONSHOT_API_HOSTS = new Set([
+  'api.moonshot.ai',
+  'api.moonshot.cn',
+])
 
 const COPILOT_HEADERS: Record<string, string> = {
   'User-Agent': 'GitHubCopilotChat/0.26.7',
@@ -145,6 +151,15 @@ function hasGeminiApiHost(baseUrl: string | undefined): boolean {
   }
 }
 
+function isMoonshotBaseUrl(baseUrl: string | undefined): boolean {
+  if (!baseUrl) return false
+  try {
+    return MOONSHOT_API_HOSTS.has(new URL(baseUrl).hostname.toLowerCase())
+  } catch {
+    return false
+  }
+}
+
 function formatRetryAfterHint(response: Response): string {
   const ra = response.headers.get('retry-after')
   return ra ? ` (Retry-After: ${ra})` : ''
@@ -276,6 +291,15 @@ function convertToolResultContent(
     const text = parts[0].text ?? ''
     return isError ? `Error: ${text}` : text
   }
+
+  // Collapse arrays of only text blocks into a single string for DeepSeek
+  // compatibility (issue #774). DeepSeek rejects arrays in role: "tool" messages.
+  const allText = parts.every(p => p.type === 'text')
+  if (allText) {
+    const text = parts.map(p => p.text ?? '').join('\n\n')
+    return isError ? `Error: ${text}` : text
+  }
+
   if (isError && parts[0]?.type === 'text') {
     parts[0] = { ...parts[0], text: `Error: ${parts[0].text ?? ''}` }
   } else if (isError) {
@@ -334,6 +358,14 @@ function convertContentBlocks(
 
   if (parts.length === 0) return ''
   if (parts.length === 1 && parts[0].type === 'text') return parts[0].text ?? ''
+
+  // Collapse arrays of only text blocks into a single string for DeepSeek
+  // compatibility (issue #774).
+  const allText = parts.every(p => p.type === 'text')
+  if (allText) {
+    return parts.map(p => p.text ?? '').join('\n\n')
+  }
+
   return parts
 }
 
@@ -345,19 +377,43 @@ function isGeminiMode(): boolean {
 }
 
 function convertMessages(
-  messages: Array<{ role: string; message?: { role?: string; content?: unknown }; content?: unknown }>,
+  messages: Array<{
+    role: string
+    message?: { role?: string; content?: unknown }
+    content?: unknown
+  }>,
   system: unknown,
 ): OpenAIMessage[] {
   const result: OpenAIMessage[] = []
   const knownToolCallIds = new Set<string>()
 
+  // Pre-scan for all tool results in the history to identify valid tool calls
+  const toolResultIds = new Set<string>()
+  for (const msg of messages) {
+    const inner = msg.message ?? msg
+    const content = (inner as { content?: unknown }).content
+    if (Array.isArray(content)) {
+      for (const block of content) {
+        if (
+          (block as { type?: string }).type === 'tool_result' &&
+          (block as { tool_use_id?: string }).tool_use_id
+        ) {
+          toolResultIds.add((block as { tool_use_id: string }).tool_use_id)
+        }
+      }
+    }
+  }
+
   // System message first
   const sysText = convertSystemPrompt(system)
   if (sysText) {
     result.push({ role: 'system', content: sysText })
   }
 
-  for (const msg of messages) {
+  for (let i = 0; i < messages.length; i++) {
+    const msg = messages[i]
+    const isLastInHistory = i === messages.length - 1
+
     // Claude Code wraps messages in { role, message: { role, content } }
     const inner = msg.message ?? msg
     const role = (inner as { role?: string }).role ?? msg.role
@@ -366,8 +422,12 @@ function convertMessages(
     if (role === 'user') {
       // Check for tool_result blocks in user messages
       if (Array.isArray(content)) {
-        const toolResults = content.filter((b: { type?: string }) => b.type === 'tool_result')
-        const otherContent = content.filter((b: { type?: string }) => b.type !== 'tool_result')
+        const toolResults = content.filter(
+          (b: { type?: string }) => b.type === 'tool_result',
+        )
+        const otherContent = content.filter(
+          (b: { type?: string }) => b.type !== 'tool_result',
+        )
 
         // Emit tool results as tool messages, but ONLY if we have a matching tool_use ID.
         // Mistral/OpenAI strictly require tool messages to follow an assistant message with tool_calls.
@@ -382,7 +442,9 @@ function convertMessages(
               content: convertToolResultContent(tr.content, tr.is_error),
             })
           } else {
-            logForDebugging(`Dropping orphan tool_result for ID: ${id} to prevent API error`)
+            logForDebugging(
+              `Dropping orphan tool_result for ID: ${id} to prevent API error`,
+            )
           }
         }
 
@@ -402,8 +464,12 @@ function convertMessages(
     } else if (role === 'assistant') {
       // Check for tool_use blocks
       if (Array.isArray(content)) {
-        const toolUses = content.filter((b: { type?: string }) => b.type === 'tool_use')
-        const thinkingBlock = content.find((b: { type?: string }) => b.type === 'thinking')
+        const toolUses = content.filter(
+          (b: { type?: string }) => b.type === 'tool_use',
+        )
+        const thinkingBlock = content.find(
+          (b: { type?: string }) => b.type === 'thinking',
+        )
         const textContent = content.filter(
           (b: { type?: string }) => b.type !== 'tool_use' && b.type !== 'thinking',
         )
@@ -412,70 +478,108 @@ function convertMessages(
           role: 'assistant',
           content: (() => {
             const c = convertContentBlocks(textContent)
-            return typeof c === 'string' ? c : Array.isArray(c) ? c.map((p: { text?: string }) => p.text ?? '').join('') : ''
+            return typeof c === 'string'
+              ? c
+              : Array.isArray(c)
+                ? c.map((p: { text?: string }) => p.text ?? '').join('')
+                : ''
           })(),
         }
 
         if (toolUses.length > 0) {
-          assistantMsg.tool_calls = toolUses.map(
-            (tu: {
-              id?: string
-              name?: string
-              input?: unknown
-              extra_content?: Record<string, unknown>
-              signature?: string
-            }) => {
-              const id = tu.id ?? `call_${crypto.randomUUID().replace(/-/g, '')}`
-              knownToolCallIds.add(id)
-              const toolCall: NonNullable<OpenAIMessage['tool_calls']>[number] = {
-                id,
-                type: 'function' as const,
-                function: {
-                  name: tu.name ?? 'unknown',
-                  arguments:
-                    typeof tu.input === 'string'
-                      ? tu.input
-                      : JSON.stringify(tu.input ?? {}),
-                },
-              }
+          const mappedToolCalls = toolUses
+            .map(
+              (tu: {
+                id?: string
+                name?: string
+                input?: unknown
+                extra_content?: Record<string, unknown>
+                signature?: string
+              }) => {
+                const id = tu.id ?? `call_${crypto.randomUUID().replace(/-/g, '')}`
 
-              // Preserve existing extra_content if present
-              if (tu.extra_content) {
-                toolCall.extra_content = { ...tu.extra_content }
-              }
+                // Only keep tool calls that have a corresponding result in the history,
+                // or if it's the last message (prefill scenario).
+                // Orphaned tool calls (e.g. from user interruption) cause 400 errors.
+                if (!toolResultIds.has(id) && !isLastInHistory) {
+                  return null
+                }
 
-              // Handle Gemini thought_signature
-              if (isGeminiMode()) {
-                // If the model provided a signature in the tool_use block itself (e.g. from a previous Turn/Step)
-                // Use thinkingBlock.signature for ALL tool calls in the same assistant turn if available.
-                // The API requires the same signature on every replayed function call part in a parallel set.
-                const signature = tu.signature ?? (thinkingBlock as any)?.signature
+                knownToolCallIds.add(id)
+                const toolCall: NonNullable<
+                  OpenAIMessage['tool_calls']
+                >[number] = {
+                  id,
+                  type: 'function' as const,
+                  function: {
+                    name: tu.name ?? 'unknown',
+                    arguments:
+                      typeof tu.input === 'string'
+                        ? tu.input
+                        : JSON.stringify(tu.input ?? {}),
+                  },
+                }
 
-                // Merge into existing google-specific metadata if present
-                const existingGoogle = (toolCall.extra_content?.google as Record<string, unknown>) ?? {}
-                toolCall.extra_content = {
-                  ...toolCall.extra_content,
-                  google: {
-                    ...existingGoogle,
-                    thought_signature: signature ?? "skip_thought_signature_validator"
+                // Preserve existing extra_content if present
+                if (tu.extra_content) {
+                  toolCall.extra_content = { ...tu.extra_content }
+                }
+
+                // Handle Gemini thought_signature
+                if (isGeminiMode()) {
+                  // If the model provided a signature in the tool_use block itself (e.g. from a previous Turn/Step)
+                  // Use thinkingBlock.signature for ALL tool calls in the same assistant turn if available.
+                  // The API requires the same signature on every replayed function call part in a parallel set.
+                  const signature =
+                    tu.signature ?? (thinkingBlock as any)?.signature
+
+                  // Merge into existing google-specific metadata if present
+                  const existingGoogle =
+                    (toolCall.extra_content?.google as Record<
+                      string,
+                      unknown
+                    >) ?? {}
+                  toolCall.extra_content = {
+                    ...toolCall.extra_content,
+                    google: {
+                      ...existingGoogle,
+                      thought_signature:
+                        signature ?? 'skip_thought_signature_validator',
+                    },
                   }
                 }
-              }
 
-              return toolCall
-            },
-          )
+                return toolCall
+              },
+            )
+            .filter((tc): tc is NonNullable<typeof tc> => tc !== null)
+
+          if (mappedToolCalls.length > 0) {
+            assistantMsg.tool_calls = mappedToolCalls
+          }
         }
 
-        result.push(assistantMsg)
+        // Only push assistant message if it has content or tool calls.
+        // Stripped thinking-only blocks from user interruptions are empty and cause 400s.
+        if (assistantMsg.content || assistantMsg.tool_calls?.length) {
+          result.push(assistantMsg)
+        }
       } else {
-        result.push({
+        const assistantMsg: OpenAIMessage = {
           role: 'assistant',
           content: (() => {
             const c = convertContentBlocks(content)
-            return typeof c === 'string' ? c : Array.isArray(c) ? c.map((p: { text?: string }) => p.text ?? '').join('') : ''
+            return typeof c === 'string'
+              ? c
+              : Array.isArray(c)
+                ? c.map((p: { text?: string }) => p.text ?? '').join('')
+                : ''
           })(),
-        })
+        }
+
+        if (assistantMsg.content) {
+          result.push(assistantMsg)
+        }
       }
     }
   }
@@ -489,25 +593,56 @@ function convertMessages(
   for (const msg of result) {
     const prev = coalesced[coalesced.length - 1]
 
-    if (prev && prev.role === msg.role && msg.role !== 'tool' && msg.role !== 'system') {
-      const prevContent = prev.content
+    // Mistral/Devstral: 'tool' message must be followed by an 'assistant' message.
+    // If a 'tool' result is followed by a 'user' message, we must inject a semantic
+    // assistant response to satisfy the strict role sequence:
+    // ... -> assistant (calls) -> tool (results) -> assistant (semantic) -> user (next)
+    if (prev && prev.role === 'tool' && msg.role === 'user') {
+      coalesced.push({
+        role: 'assistant',
+        content: '[Tool execution interrupted by user]',
+      })
+    }
+
+    const lastAfterPossibleInjection = coalesced[coalesced.length - 1]
+    if (
+      lastAfterPossibleInjection &&
+      lastAfterPossibleInjection.role === msg.role &&
+      msg.role !== 'tool' &&
+      msg.role !== 'system'
+    ) {
+      const prevContent = lastAfterPossibleInjection.content
       const curContent = msg.content
 
       if (typeof prevContent === 'string' && typeof curContent === 'string') {
-        prev.content = prevContent + (prevContent && curContent ? '\n' : '') + curContent
+        lastAfterPossibleInjection.content =
+          prevContent + (prevContent && curContent ? '\n' : '') + curContent
       } else {
         const toArray = (
-          c: string | Array<{ type: string; text?: string; image_url?: { url: string } }> | undefined,
-        ): Array<{ type: string; text?: string; image_url?: { url: string } }> => {
+          c:
+            | string
+            | Array<{ type: string; text?: string; image_url?: { url: string } }>
+            | undefined,
+        ): Array<{
+          type: string
+          text?: string
+          image_url?: { url: string }
+        }> => {
           if (!c) return []
           if (typeof c === 'string') return c ? [{ type: 'text', text: c }] : []
           return c
         }
-        prev.content = [...toArray(prevContent), ...toArray(curContent)]
+        lastAfterPossibleInjection.content = [
+          ...toArray(prevContent),
+          ...toArray(curContent),
+        ]
       }
 
       if (msg.tool_calls?.length) {
-        prev.tool_calls = [...(prev.tool_calls ?? []), ...msg.tool_calls]
+        lastAfterPossibleInjection.tool_calls = [
+          ...(lastAfterPossibleInjection.tool_calls ?? []),
+          ...msg.tool_calls,
+        ]
       }
     } else {
       coalesced.push(msg)
@@ -718,8 +853,7 @@ async function* openaiStreamToAnthropic(
   let hasEmittedContentStart = false
   let hasEmittedThinkingStart = false
   let hasClosedThinking = false
-  let activeTextBuffer = ''
-  let textBufferMode: 'none' | 'pending' | 'strip' = 'none'
+  const thinkFilter = createThinkTagFilter()
   let lastStopReason: 'tool_use' | 'max_tokens' | 'end_turn' | null = null
   let hasEmittedFinalUsage = false
   let hasProcessedFinishReason = false
@@ -798,14 +932,12 @@ async function* openaiStreamToAnthropic(
   const closeActiveContentBlock = async function* () {
     if (!hasEmittedContentStart) return
 
-    if (textBufferMode !== 'none') {
-      const sanitized = stripLeakedReasoningPreamble(activeTextBuffer)
-      if (sanitized) {
-        yield {
-          type: 'content_block_delta',
-          index: contentBlockIndex,
-          delta: { type: 'text_delta', text: sanitized },
-        }
+    const tail = thinkFilter.flush()
+    if (tail) {
+      yield {
+        type: 'content_block_delta',
+        index: contentBlockIndex,
+        delta: { type: 'text_delta', text: tail },
       }
     }
 
@@ -815,8 +947,6 @@ async function* openaiStreamToAnthropic(
     }
     contentBlockIndex++
     hasEmittedContentStart = false
-    activeTextBuffer = ''
-    textBufferMode = 'none'
   }
 
   try {
@@ -873,7 +1003,6 @@ async function* openaiStreamToAnthropic(
             contentBlockIndex++
             hasClosedThinking = true
           }
-          activeTextBuffer += delta.content
           if (!hasEmittedContentStart) {
             yield {
               type: 'content_block_start',
@@ -883,38 +1012,13 @@ async function* openaiStreamToAnthropic(
             hasEmittedContentStart = true
           }
 
-          if (
-            textBufferMode === 'strip' ||
-            looksLikeLeakedReasoningPrefix(activeTextBuffer)
-          ) {
-            textBufferMode = 'strip'
-            continue
-          }
-
-          if (textBufferMode === 'pending') {
-            if (shouldBufferPotentialReasoningPrefix(activeTextBuffer)) {
-              continue
-            }
+          const visible = thinkFilter.feed(delta.content)
+          if (visible) {
             yield {
               type: 'content_block_delta',
               index: contentBlockIndex,
-              delta: {
-                type: 'text_delta',
-                text: activeTextBuffer,
-              },
+              delta: { type: 'text_delta', text: visible },
             }
-            textBufferMode = 'none'
-            continue
-          }
-
-          if (shouldBufferPotentialReasoningPrefix(activeTextBuffer)) {
-            textBufferMode = 'pending'
-            continue
-          }
-          yield {
-            type: 'content_block_delta',
-            index: contentBlockIndex,
-            delta: { type: 'text_delta', text: delta.content },
           }
         }
 
@@ -1329,14 +1433,15 @@ class OpenAIShimMessages {
     params: ShimCreateParams,
     options?: { signal?: AbortSignal; headers?: Record<string, string> },
   ): Promise<Response> {
-    const openaiMessages = convertMessages(
+    const compressedMessages = compressToolHistory(
       params.messages as Array<{
         role: string
         message?: { role?: string; content?: unknown }
         content?: unknown
       }>,
-      params.system,
+      request.resolvedModel,
     )
+    const openaiMessages = convertMessages(compressedMessages, params.system)
 
     const body: Record<string, unknown> = {
       model: request.resolvedModel,
@@ -1372,14 +1477,19 @@ class OpenAIShimMessages {
     const isGithubCopilot = isGithub && githubEndpointType === 'copilot'
     const isGithubModels = isGithub && (githubEndpointType === 'models' || githubEndpointType === 'custom')
 
-    if ((isGithub || isMistral || isLocal) && body.max_completion_tokens !== undefined) {
+    const isMoonshot = isMoonshotBaseUrl(request.baseUrl)
+
+    if ((isGithub || isMistral || isLocal || isMoonshot) && body.max_completion_tokens !== undefined) {
       body.max_tokens = body.max_completion_tokens
       delete body.max_completion_tokens
     }
 
     // mistral and gemini don't recognize body.store — Gemini returns 400
     // "Invalid JSON payload received. Unknown name 'store': Cannot find field."
-    if (isMistral || isGeminiMode()) {
+    // Moonshot (api.moonshot.ai/.cn) has not published support for the
+    // parameter either; strip it preemptively to avoid the same class of
+    // error on strict-parse providers.
+    if (isMistral || isGeminiMode() || isMoonshot) {
       delete body.store
     }
 
@@ -1459,48 +1569,95 @@ class OpenAIShimMessages {
       headers['X-GitHub-Api-Version'] = '2022-11-28'
     }
 
-    // Build the chat completions URL
-    // Azure Cognitive Services / Azure OpenAI require a deployment-specific path
-    // and an api-version query parameter.
-    // Standard format: {base}/openai/deployments/{model}/chat/completions?api-version={version}
-    // Non-Azure: {base}/chat/completions
-    let chatCompletionsUrl: string
-    if (isAzure) {
-      const apiVersion = process.env.AZURE_OPENAI_API_VERSION ?? '2024-12-01-preview'
-      const deployment = request.resolvedModel ?? process.env.OPENAI_MODEL ?? 'gpt-4o'
-      // If base URL already contains /deployments/, use it as-is with api-version
-      if (/\/deployments\//i.test(request.baseUrl)) {
-        const base = request.baseUrl.replace(/\/+$/, '')
-        chatCompletionsUrl = `${base}/chat/completions?api-version=${apiVersion}`
-      } else {
-        // Strip trailing /v1 or /openai/v1 if present, then build Azure path
-        const base = request.baseUrl.replace(/\/(openai\/)?v1\/?$/, '').replace(/\/+$/, '')
-        chatCompletionsUrl = `${base}/openai/deployments/${deployment}/chat/completions?api-version=${apiVersion}`
+    const buildChatCompletionsUrl = (baseUrl: string): string => {
+      // Azure Cognitive Services / Azure OpenAI require a deployment-specific
+      // path and an api-version query parameter.
+      if (isAzure) {
+        const apiVersion = process.env.AZURE_OPENAI_API_VERSION ?? '2024-12-01-preview'
+        const deployment = request.resolvedModel ?? process.env.OPENAI_MODEL ?? 'gpt-4o'
+
+        // If base URL already contains /deployments/, use it as-is with api-version.
+        if (/\/deployments\//i.test(baseUrl)) {
+          const normalizedBase = baseUrl.replace(/\/+$/, '')
+          return `${normalizedBase}/chat/completions?api-version=${apiVersion}`
+        }
+
+        // Strip trailing /v1 or /openai/v1 if present, then build Azure path.
+        const normalizedBase = baseUrl
+          .replace(/\/(openai\/)?v1\/?$/, '')
+          .replace(/\/+$/, '')
+
+        return `${normalizedBase}/openai/deployments/${deployment}/chat/completions?api-version=${apiVersion}`
       }
-    } else {
-      chatCompletionsUrl = `${request.baseUrl}/chat/completions`
+
+      return `${baseUrl}/chat/completions`
     }
 
-    const fetchInit = {
+    const localRetryBaseUrls = isLocal
+      ? getLocalProviderRetryBaseUrls(request.baseUrl)
+      : []
+
+    let activeBaseUrl = request.baseUrl
+    let chatCompletionsUrl = buildChatCompletionsUrl(activeBaseUrl)
+    const attemptedLocalBaseUrls = new Set<string>([activeBaseUrl])
+    let didRetryWithoutTools = false
+
+    const promoteNextLocalBaseUrl = (
+      reason: 'endpoint_not_found' | 'localhost_resolution_failed',
+    ): boolean => {
+      for (const candidateBaseUrl of localRetryBaseUrls) {
+        if (attemptedLocalBaseUrls.has(candidateBaseUrl)) {
+          continue
+        }
+
+        const previousUrl = chatCompletionsUrl
+        attemptedLocalBaseUrls.add(candidateBaseUrl)
+        activeBaseUrl = candidateBaseUrl
+        chatCompletionsUrl = buildChatCompletionsUrl(activeBaseUrl)
+
+        logForDebugging(
+          `[OpenAIShim] self-heal retry reason=${reason} method=POST from=${redactUrlForDiagnostics(previousUrl)} to=${redactUrlForDiagnostics(chatCompletionsUrl)} model=${request.resolvedModel}`,
+          { level: 'warn' },
+        )
+
+        return true
+      }
+
+      return false
+    }
+
+    let serializedBody = JSON.stringify(body)
+
+    const refreshSerializedBody = (): void => {
+      serializedBody = JSON.stringify(body)
+    }
+
+    const buildFetchInit = () => ({
       method: 'POST' as const,
       headers,
-      body: JSON.stringify(body),
+      body: serializedBody,
       signal: options?.signal,
-    }
+    })
 
-    const maxAttempts = isGithub ? GITHUB_429_MAX_RETRIES : 1
+    const maxSelfHealAttempts = isLocal
+      ? localRetryBaseUrls.length + 1
+      : 0
+    const maxAttempts = (isGithub ? GITHUB_429_MAX_RETRIES : 1) + maxSelfHealAttempts
 
     const throwClassifiedTransportError = (
       error: unknown,
       requestUrl: string,
+      preclassifiedFailure?: ReturnType<typeof classifyOpenAINetworkFailure>,
     ): never => {
       if (options?.signal?.aborted) {
         throw error
       }
 
-      const failure = classifyOpenAINetworkFailure(error, {
-        url: requestUrl,
-      })
+      const failure =
+        preclassifiedFailure ??
+        classifyOpenAINetworkFailure(error, {
+          url: requestUrl,
+        })
       const redactedUrl = redactUrlForDiagnostics(requestUrl)
       const safeMessage =
         redactSecretValueForDisplay(
@@ -1531,11 +1688,14 @@ class OpenAIShimMessages {
       responseHeaders: Headers,
       requestUrl: string,
       rateHint = '',
+      preclassifiedFailure?: ReturnType<typeof classifyOpenAIHttpFailure>,
     ): never => {
-      const failure = classifyOpenAIHttpFailure({
-        status,
-        body: errorBody,
-      })
+      const failure =
+        preclassifiedFailure ??
+        classifyOpenAIHttpFailure({
+          status,
+          body: errorBody,
+        })
       const redactedUrl = redactUrlForDiagnostics(requestUrl)
 
       logForDebugging(
@@ -1557,10 +1717,13 @@ class OpenAIShimMessages {
     let response: Response | undefined
     for (let attempt = 0; attempt < maxAttempts; attempt++) {
       try {
-        response = await fetchWithProxyRetry(chatCompletionsUrl, fetchInit)
+        response = await fetchWithProxyRetry(
+          chatCompletionsUrl,
+          buildFetchInit(),
+        )
       } catch (error) {
         const isAbortError =
-          fetchInit.signal?.aborted === true ||
+          options?.signal?.aborted === true ||
           (typeof DOMException !== 'undefined' &&
             error instanceof DOMException &&
             error.name === 'AbortError') ||
@@ -1573,7 +1736,19 @@ class OpenAIShimMessages {
           throw error
         }
 
-        throwClassifiedTransportError(error, chatCompletionsUrl)
+        const failure = classifyOpenAINetworkFailure(error, {
+          url: chatCompletionsUrl,
+        })
+
+        if (
+          isLocal &&
+          failure.category === 'localhost_resolution_failed' &&
+          promoteNextLocalBaseUrl('localhost_resolution_failed')
+        ) {
+          continue
+        }
+
+        throwClassifiedTransportError(error, chatCompletionsUrl, failure)
       }
 
       if (response.ok) {
@@ -1665,6 +1840,10 @@ class OpenAIShimMessages {
             return responsesResponse
           }
           const responsesErrorBody = await responsesResponse.text().catch(() => 'unknown error')
+          const responsesFailure = classifyOpenAIHttpFailure({
+            status: responsesResponse.status,
+            body: responsesErrorBody,
+          })
           let responsesErrorResponse: object | undefined
           try { responsesErrorResponse = JSON.parse(responsesErrorBody) } catch { /* raw text */ }
           throwClassifiedHttpError(
@@ -1673,10 +1852,49 @@ class OpenAIShimMessages {
             responsesErrorResponse,
             responsesResponse.headers,
             responsesUrl,
+            '',
+            responsesFailure,
           )
         }
       }
 
+      const failure = classifyOpenAIHttpFailure({
+        status: response.status,
+        body: errorBody,
+      })
+
+      if (
+        isLocal &&
+        failure.category === 'endpoint_not_found' &&
+        promoteNextLocalBaseUrl('endpoint_not_found')
+      ) {
+        continue
+      }
+
+      const hasToolsPayload =
+        Array.isArray(body.tools) &&
+        body.tools.length > 0
+
+      if (
+        !didRetryWithoutTools &&
+        failure.category === 'tool_call_incompatible' &&
+        shouldAttemptLocalToollessRetry({
+          baseUrl: activeBaseUrl,
+          hasTools: hasToolsPayload,
+        })
+      ) {
+        didRetryWithoutTools = true
+        delete body.tools
+        delete body.tool_choice
+        refreshSerializedBody()
+
+        logForDebugging(
+          `[OpenAIShim] self-heal retry reason=tool_call_incompatible mode=toolless method=POST url=${redactUrlForDiagnostics(chatCompletionsUrl)} model=${request.resolvedModel}`,
+          { level: 'warn' },
+        )
+        continue
+      }
+
       let errorResponse: object | undefined
       try { errorResponse = JSON.parse(errorBody) } catch { /* raw text */ }
       throwClassifiedHttpError(
@@ -1686,6 +1904,7 @@ class OpenAIShimMessages {
         response.headers as unknown as Headers,
         chatCompletionsUrl,
         rateHint,
+        failure,
       )
     }
 
@@ -1742,7 +1961,7 @@ class OpenAIShimMessages {
     if (typeof rawContent === 'string' && rawContent) {
       content.push({
         type: 'text',
-        text: stripLeakedReasoningPreamble(rawContent),
+        text: stripThinkTags(rawContent),
       })
     } else if (Array.isArray(rawContent) && rawContent.length > 0) {
       const parts: string[] = []
@@ -1760,7 +1979,7 @@ class OpenAIShimMessages {
       if (joined) {
         content.push({
           type: 'text',
-          text: stripLeakedReasoningPreamble(joined),
+          text: stripThinkTags(joined),
         })
       }
     }

src/services/api/providerConfig.local.test.ts

@@ -2,8 +2,10 @@ import { afterEach, expect, test } from 'bun:test'
 
 import {
   getAdditionalModelOptionsCacheScope,
+  getLocalProviderRetryBaseUrls,
   isLocalProviderUrl,
   resolveProviderRequest,
+  shouldAttemptLocalToollessRetry,
 } from './providerConfig.js'
 
 const originalEnv = {
@@ -83,3 +85,42 @@ test('skips local model cache scope for remote openai-compatible providers', ()
 
   expect(getAdditionalModelOptionsCacheScope()).toBeNull()
 })
+
+test('derives local retry base URLs with /v1 and loopback fallback candidates', () => {
+  expect(getLocalProviderRetryBaseUrls('http://localhost:11434')).toEqual([
+    'http://localhost:11434/v1',
+    'http://127.0.0.1:11434',
+    'http://127.0.0.1:11434/v1',
+  ])
+})
+
+test('does not derive local retry base URLs for remote providers', () => {
+  expect(getLocalProviderRetryBaseUrls('https://api.openai.com/v1')).toEqual([])
+})
+
+test('enables local toolless retry for likely Ollama endpoints with tools', () => {
+  expect(
+    shouldAttemptLocalToollessRetry({
+      baseUrl: 'http://localhost:11434/v1',
+      hasTools: true,
+    }),
+  ).toBe(true)
+})
+
+test('disables local toolless retry when no tools are present', () => {
+  expect(
+    shouldAttemptLocalToollessRetry({
+      baseUrl: 'http://localhost:11434/v1',
+      hasTools: false,
+    }),
+  ).toBe(false)
+})
+
+test('disables local toolless retry for non-Ollama local endpoints', () => {
+  expect(
+    shouldAttemptLocalToollessRetry({
+      baseUrl: 'http://localhost:1234/v1',
+      hasTools: true,
+    }),
+  ).toBe(false)
+})

src/services/api/providerConfig.ts

@@ -305,6 +305,101 @@ export function isLocalProviderUrl(baseUrl: string | undefined): boolean {
   }
 }
 
+function trimTrailingSlash(value: string): string {
+  return value.replace(/\/+$/, '')
+}
+
+function normalizePathWithV1(pathname: string): string {
+  const trimmed = trimTrailingSlash(pathname)
+  if (!trimmed || trimmed === '/') {
+    return '/v1'
+  }
+
+  if (trimmed.toLowerCase().endsWith('/v1')) {
+    return trimmed
+  }
+
+  return `${trimmed}/v1`
+}
+
+function isLikelyOllamaEndpoint(baseUrl: string): boolean {
+  try {
+    const parsed = new URL(baseUrl)
+    const hostname = parsed.hostname.toLowerCase()
+    const pathname = parsed.pathname.toLowerCase()
+
+    if (parsed.port === '11434') {
+      return true
+    }
+
+    return (
+      hostname.includes('ollama') ||
+      pathname.includes('ollama')
+    )
+  } catch {
+    return false
+  }
+}
+
+export function getLocalProviderRetryBaseUrls(baseUrl: string): string[] {
+  if (!isLocalProviderUrl(baseUrl)) {
+    return []
+  }
+
+  try {
+    const parsed = new URL(baseUrl)
+    const original = trimTrailingSlash(parsed.toString())
+    const seen = new Set<string>([original])
+    const candidates: string[] = []
+
+    const addCandidate = (hostname: string, pathname: string): void => {
+      const next = new URL(parsed.toString())
+      next.hostname = hostname
+      next.pathname = pathname
+      next.search = ''
+      next.hash = ''
+
+      const normalized = trimTrailingSlash(next.toString())
+      if (seen.has(normalized)) {
+        return
+      }
+
+      seen.add(normalized)
+      candidates.push(normalized)
+    }
+
+    const v1Pathname = normalizePathWithV1(parsed.pathname)
+    if (v1Pathname !== trimTrailingSlash(parsed.pathname)) {
+      addCandidate(parsed.hostname, v1Pathname)
+    }
+
+    const hostname = parsed.hostname.toLowerCase().replace(/^\[|\]$/g, '')
+    if (hostname === 'localhost' || hostname === '::1') {
+      addCandidate('127.0.0.1', parsed.pathname || '/')
+      addCandidate('127.0.0.1', v1Pathname)
+    }
+
+    return candidates
+  } catch {
+    return []
+  }
+}
+
+export function shouldAttemptLocalToollessRetry(options: {
+  baseUrl: string
+  hasTools: boolean
+}): boolean {
+  if (!options.hasTools) {
+    return false
+  }
+
+  if (!isLocalProviderUrl(options.baseUrl)) {
+    return false
+  }
+
+  return isLikelyOllamaEndpoint(options.baseUrl)
+}
+
 export function isCodexBaseUrl(baseUrl: string | undefined): boolean {
   if (!baseUrl) return false
   try {
@@ -412,6 +507,9 @@ export function resolveProviderRequest(options?: {
     ? normalizedGeminiEnvBaseUrl
     : asNamedEnvUrl(process.env.OPENAI_BASE_URL, 'OPENAI_BASE_URL')
 
+  // In Mistral mode, a literal "undefined" MISTRAL_BASE_URL is treated as
+  // misconfiguration and falls back to OPENAI_API_BASE, then
+  // DEFAULT_MISTRAL_BASE_URL for a safe default endpoint.
   const fallbackEnvBaseUrl = isMistralMode
     ? (primaryEnvBaseUrl === undefined
       ? asNamedEnvUrl(process.env.OPENAI_API_BASE, 'OPENAI_API_BASE') ?? DEFAULT_MISTRAL_BASE_URL

src/services/api/reasoningLeakSanitizer.test.ts

@@ -1,46 +0,0 @@
-import { describe, expect, test } from 'bun:test'
-
-import {
-  looksLikeLeakedReasoningPrefix,
-  shouldBufferPotentialReasoningPrefix,
-  stripLeakedReasoningPreamble,
-} from './reasoningLeakSanitizer.ts'
-
-describe('reasoning leak sanitizer', () => {
-  test('strips explicit internal reasoning preambles', () => {
-    const text =
-      'The user just said "hey" - a simple greeting. I should respond briefly and friendly.\n\nHey! How can I help you today?'
-
-    expect(looksLikeLeakedReasoningPrefix(text)).toBe(true)
-    expect(stripLeakedReasoningPreamble(text)).toBe(
-      'Hey! How can I help you today?',
-    )
-  })
-
-  test('does not strip normal user-facing advice that mentions "the user should"', () => {
-    const text =
-      'The user should reset their password immediately.\n\nHere are the steps...'
-
-    expect(looksLikeLeakedReasoningPrefix(text)).toBe(false)
-    expect(shouldBufferPotentialReasoningPrefix(text)).toBe(false)
-    expect(stripLeakedReasoningPreamble(text)).toBe(text)
-  })
-
-  test('does not strip legitimate first-person advice about responding to an incident', () => {
-    const text =
-      'I need to respond to this security incident immediately. The system is compromised.\n\nHere are the remediation steps...'
-
-    expect(looksLikeLeakedReasoningPrefix(text)).toBe(false)
-    expect(shouldBufferPotentialReasoningPrefix(text)).toBe(false)
-    expect(stripLeakedReasoningPreamble(text)).toBe(text)
-  })
-
-  test('does not strip legitimate first-person advice about answering a support ticket', () => {
-    const text =
-      'I need to answer the support ticket before end of day. The customer is waiting.\n\nHere is the response I drafted...'
-
-    expect(looksLikeLeakedReasoningPrefix(text)).toBe(false)
-    expect(shouldBufferPotentialReasoningPrefix(text)).toBe(false)
-    expect(stripLeakedReasoningPreamble(text)).toBe(text)
-  })
-})

src/services/api/reasoningLeakSanitizer.ts

@@ -1,54 +0,0 @@
-const EXPLICIT_REASONING_START_RE =
-  /^\s*(i should\b|i need to\b|let me think\b|the task\b|the request\b)/i
-
-const EXPLICIT_REASONING_META_RE =
-  /\b(user|request|question|prompt|message|task|greeting|small talk|briefly|friendly|concise)\b/i
-
-const USER_META_START_RE =
-  /^\s*the user\s+(just\s+)?(said|asked|is asking|wants|wanted|mentioned|seems|appears)\b/i
-
-const USER_REASONING_RE =
-  /^\s*the user\s+(just\s+)?(said|asked|is asking|wants|wanted|mentioned|seems|appears)\b[\s\S]*\b(i should|i need to|let me think|respond|reply|answer|greeting|small talk|briefly|friendly|concise)\b/i
-
-export function shouldBufferPotentialReasoningPrefix(text: string): boolean {
-  const normalized = text.trim()
-  if (!normalized) return false
-
-  if (looksLikeLeakedReasoningPrefix(normalized)) {
-    return true
-  }
-
-  const hasParagraphBoundary = /\n\s*\n/.test(normalized)
-  if (hasParagraphBoundary) {
-    return false
-  }
-
-  return (
-    EXPLICIT_REASONING_START_RE.test(normalized) ||
-    USER_META_START_RE.test(normalized)
-  )
-}
-
-export function looksLikeLeakedReasoningPrefix(text: string): boolean {
-  const normalized = text.trim()
-  if (!normalized) return false
-  return (
-    (EXPLICIT_REASONING_START_RE.test(normalized) &&
-      EXPLICIT_REASONING_META_RE.test(normalized)) ||
-    USER_REASONING_RE.test(normalized)
-  )
-}
-
-export function stripLeakedReasoningPreamble(text: string): string {
-  const normalized = text.replace(/\r\n/g, '\n')
-  const parts = normalized.split(/\n\s*\n/)
-  if (parts.length < 2) return text
-
-  const first = parts[0]?.trim() ?? ''
-  if (!looksLikeLeakedReasoningPrefix(first)) {
-    return text
-  }
-
-  const remainder = parts.slice(1).join('\n\n').trim()
-  return remainder || text
-}

src/services/api/smartModelRouting.test.ts

@@ -0,0 +1,191 @@
+import { describe, expect, test } from 'bun:test'
+
+import {
+  routeModel,
+  type SmartRoutingConfig,
+} from './smartModelRouting.ts'
+
+const ENABLED: SmartRoutingConfig = {
+  enabled: true,
+  simpleModel: 'claude-haiku-4-5',
+  strongModel: 'claude-opus-4-7',
+}
+
+describe('routeModel — disabled / misconfigured', () => {
+  test('disabled config routes to strong', () => {
+    const decision = routeModel(
+      { userText: 'hi' },
+      { ...ENABLED, enabled: false },
+    )
+    expect(decision.model).toBe('claude-opus-4-7')
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('disabled')
+  })
+
+  test('missing simpleModel falls back to strong', () => {
+    const decision = routeModel(
+      { userText: 'hi' },
+      { ...ENABLED, simpleModel: '' },
+    )
+    expect(decision.model).toBe('claude-opus-4-7')
+    expect(decision.complexity).toBe('strong')
+  })
+
+  test('simpleModel === strongModel routes to strong (no-op)', () => {
+    const decision = routeModel(
+      { userText: 'hi' },
+      { ...ENABLED, simpleModel: 'claude-opus-4-7' },
+    )
+    expect(decision.model).toBe('claude-opus-4-7')
+    expect(decision.complexity).toBe('strong')
+  })
+})
+
+describe('routeModel — simple path', () => {
+  test('short greeting routes to simple', () => {
+    const decision = routeModel({ userText: 'thanks!', turnNumber: 5 }, ENABLED)
+    expect(decision.model).toBe('claude-haiku-4-5')
+    expect(decision.complexity).toBe('simple')
+  })
+
+  test('empty input routes to simple', () => {
+    const decision = routeModel({ userText: '   ' }, ENABLED)
+    expect(decision.model).toBe('claude-haiku-4-5')
+    expect(decision.complexity).toBe('simple')
+  })
+
+  test('mid-length chatter routes to simple', () => {
+    const decision = routeModel(
+      { userText: 'yep looks good, go ahead', turnNumber: 10 },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('simple')
+  })
+})
+
+describe('routeModel — strong path', () => {
+  test('first turn always routes to strong, even when short', () => {
+    const decision = routeModel(
+      { userText: 'fix the bug', turnNumber: 1 },
+      ENABLED,
+    )
+    expect(decision.model).toBe('claude-opus-4-7')
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('first turn')
+  })
+
+  test('code fence routes to strong', () => {
+    const decision = routeModel(
+      {
+        userText: 'change this:\n```\nfoo()\n```',
+        turnNumber: 5,
+      },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('code')
+  })
+
+  test('inline code span routes to strong', () => {
+    const decision = routeModel(
+      { userText: 'rename `foo` to `bar`', turnNumber: 5 },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('strong')
+  })
+
+  test('reasoning keyword "plan" routes to strong even when short', () => {
+    const decision = routeModel(
+      { userText: 'plan the refactor', turnNumber: 5 },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('keyword')
+  })
+
+  test('reasoning keyword "debug" routes to strong', () => {
+    const decision = routeModel(
+      { userText: 'debug the test', turnNumber: 5 },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('strong')
+  })
+
+  test('"root cause" multi-word keyword routes to strong', () => {
+    const decision = routeModel(
+      { userText: 'find the root cause', turnNumber: 5 },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('strong')
+  })
+
+  test('multi-paragraph input routes to strong', () => {
+    const decision = routeModel(
+      {
+        userText: 'first thought.\n\nsecond thought.',
+        turnNumber: 5,
+      },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('multi-paragraph')
+  })
+
+  test('over-long input routes to strong', () => {
+    const long = 'ok '.repeat(100) // ~300 chars, 100 words
+    const decision = routeModel(
+      { userText: long, turnNumber: 5 },
+      ENABLED,
+    )
+    expect(decision.complexity).toBe('strong')
+  })
+
+  test('exactly at the boundary stays simple', () => {
+    const text = 'a'.repeat(160)
+    const decision = routeModel(
+      { userText: text, turnNumber: 5 },
+      { ...ENABLED, simpleMaxChars: 160, simpleMaxWords: 28 },
+    )
+    expect(decision.complexity).toBe('simple')
+  })
+
+  test('one char over the boundary routes to strong', () => {
+    const text = 'a'.repeat(161)
+    const decision = routeModel(
+      { userText: text, turnNumber: 5 },
+      { ...ENABLED, simpleMaxChars: 160, simpleMaxWords: 28 },
+    )
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('160 chars')
+  })
+})
+
+describe('routeModel — config overrides', () => {
+  test('custom simpleMaxChars is honored', () => {
+    const decision = routeModel(
+      { userText: 'abcdefghijklmnop', turnNumber: 5 },
+      { ...ENABLED, simpleMaxChars: 10 },
+    )
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('10 chars')
+  })
+
+  test('custom simpleMaxWords is honored', () => {
+    const decision = routeModel(
+      { userText: 'one two three four five', turnNumber: 5 },
+      { ...ENABLED, simpleMaxWords: 3 },
+    )
+    expect(decision.complexity).toBe('strong')
+    expect(decision.reason).toContain('3 words')
+  })
+})
+
+describe('routeModel — reason strings', () => {
+  test('simple decisions include char + word counts', () => {
+    const decision = routeModel(
+      { userText: 'sounds good', turnNumber: 5 },
+      ENABLED,
+    )
+    expect(decision.reason).toMatch(/\d+ chars, \d+ words/)
+  })
+})

src/services/api/smartModelRouting.ts

@@ -0,0 +1,215 @@
+/**
+ * Smart model routing — cheap-for-simple, strong-for-hard.
+ *
+ * For everyday short chatter ("ok", "thanks", "what does this do?") the
+ * incremental quality of Opus/GPT-5 over Haiku/Mini is negligible while the
+ * cost and latency are an order of magnitude worse. Smart routing opts a
+ * user into routing such "obviously simple" turns to a cheaper model while
+ * keeping the strong model for the anything-non-trivial path.
+ *
+ * This module is a pure primitive: it takes a turn description (the user's
+ * text + light context) and returns which model to use, based on config.
+ * It never reads env vars or state directly — caller supplies everything.
+ *
+ * Off by default. Users opt in via settings.smartRouting.enabled. Intent:
+ * make this a copy-paste-small config block rather than a hidden heuristic,
+ * so the tradeoff is visible and the user controls it.
+ */
+
+export type SmartRoutingConfig = {
+  enabled: boolean
+  /** Model to use for turns classified as "simple". */
+  simpleModel: string
+  /** Model to use for turns classified as "strong" (or when unsure). */
+  strongModel: string
+  /** Max characters in user input to qualify as "simple". Default 160. */
+  simpleMaxChars?: number
+  /** Max whitespace-separated words to qualify as "simple". Default 28. */
+  simpleMaxWords?: number
+}
+
+export type RoutingDecision = {
+  model: string
+  complexity: 'simple' | 'strong'
+  /** Human-readable reason — useful for the UI indicator and debug logs. */
+  reason: string
+}
+
+export type RoutingInput = {
+  /** The user's message text for this turn. */
+  userText: string
+  /**
+   * Optional: how many tool-use blocks the assistant has emitted in the
+   * recent conversation. High values correlate with "continue this work"
+   * follow-ups that can still be cheap, UNLESS the user also typed code
+   * or strong-keyword text.
+   */
+  recentToolUses?: number
+  /**
+   * Optional: turn number within the current session (1-indexed). The first
+   * turn is often task-setup and benefits from the strong model even if
+   * short — a bare "build X" opens the whole task.
+   */
+  turnNumber?: number
+}
+
+const DEFAULT_SIMPLE_MAX_CHARS = 160
+const DEFAULT_SIMPLE_MAX_WORDS = 28
+
+// Keywords that strongly suggest reasoning/planning/design work.
+// Matching is word-boundary / case-insensitive. Must include enough anchors
+// that short prompts like "plan the refactor" route to strong even under
+// the char/word cutoff.
+const STRONG_KEYWORDS = [
+  'plan',
+  'design',
+  'architect',
+  'architecture',
+  'refactor',
+  'debug',
+  'investigate',
+  'analyze',
+  'analyse',
+  'implement',
+  'optimize',
+  'optimise',
+  'review',
+  'audit',
+  'diagnose',
+  'root cause',
+  'root-cause',
+  'why does',
+  'why is',
+  'how should',
+  'why did',
+  'propose',
+  'trace',
+  'reproduce',
+]
+
+const STRONG_KEYWORD_RE = new RegExp(
+  `\\b(?:${STRONG_KEYWORDS.map(k => k.replace(/[-]/g, '[-\\s]')).join('|')})\\b`,
+  'i',
+)
+
+const CODE_FENCE_RE = /```[\s\S]*?```|`[^`\n]+`/
+
+function countWords(text: string): number {
+  const trimmed = text.trim()
+  if (!trimmed) return 0
+  return trimmed.split(/\s+/).length
+}
+
+function hasMultiParagraph(text: string): boolean {
+  return /\n\s*\n/.test(text)
+}
+
+function hasCode(text: string): boolean {
+  return CODE_FENCE_RE.test(text)
+}
+
+function hasStrongKeyword(text: string): boolean {
+  return STRONG_KEYWORD_RE.test(text)
+}
+
+/**
+ * Decide whether to route to the simple or strong model based on heuristics.
+ * Returns the chosen model + a reason. When routing is disabled or both
+ * models match, the strong model is used (safe default).
+ */
+export function routeModel(
+  input: RoutingInput,
+  config: SmartRoutingConfig,
+): RoutingDecision {
+  if (!config.enabled) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: 'smart-routing disabled',
+    }
+  }
+  if (!config.simpleModel || !config.strongModel) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: 'simpleModel or strongModel missing from config',
+    }
+  }
+  if (config.simpleModel === config.strongModel) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: 'simpleModel equals strongModel',
+    }
+  }
+
+  const text = input.userText ?? ''
+  const trimmed = text.trim()
+
+  if (!trimmed) {
+    // Empty input (e.g. resuming a tool-use chain) — cheap by default.
+    return {
+      model: config.simpleModel,
+      complexity: 'simple',
+      reason: 'empty user text',
+    }
+  }
+
+  // First turn of a session is task-setup — always use strong.
+  if (input.turnNumber === 1) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: 'first turn of session',
+    }
+  }
+
+  const maxChars = config.simpleMaxChars ?? DEFAULT_SIMPLE_MAX_CHARS
+  const maxWords = config.simpleMaxWords ?? DEFAULT_SIMPLE_MAX_WORDS
+
+  if (hasCode(trimmed)) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: 'contains code block or inline code',
+    }
+  }
+
+  if (hasStrongKeyword(trimmed)) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: 'contains reasoning/planning keyword',
+    }
+  }
+
+  if (hasMultiParagraph(trimmed)) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: 'multi-paragraph input',
+    }
+  }
+
+  if (trimmed.length > maxChars) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: `input > ${maxChars} chars`,
+    }
+  }
+
+  if (countWords(trimmed) > maxWords) {
+    return {
+      model: config.strongModel,
+      complexity: 'strong',
+      reason: `input > ${maxWords} words`,
+    }
+  }
+
+  return {
+    model: config.simpleModel,
+    complexity: 'simple',
+    reason: `short (${trimmed.length} chars, ${countWords(trimmed)} words)`,
+  }
+}

src/services/api/thinkTagSanitizer.test.ts

@@ -0,0 +1,183 @@
+import { describe, expect, test } from 'bun:test'
+
+import {
+  createThinkTagFilter,
+  stripThinkTags,
+} from './thinkTagSanitizer.ts'
+
+describe('stripThinkTags — whole-text cleanup', () => {
+  test('strips closed think pair', () => {
+    expect(stripThinkTags('<think>reasoning</think>Hello')).toBe('Hello')
+  })
+
+  test('strips closed thinking pair', () => {
+    expect(stripThinkTags('<thinking>x</thinking>Out')).toBe('Out')
+  })
+
+  test('strips closed reasoning pair', () => {
+    expect(stripThinkTags('<reasoning>x</reasoning>Out')).toBe('Out')
+  })
+
+  test('strips REASONING_SCRATCHPAD pair', () => {
+    expect(stripThinkTags('<REASONING_SCRATCHPAD>plan</REASONING_SCRATCHPAD>Answer'))
+      .toBe('Answer')
+  })
+
+  test('is case-insensitive', () => {
+    expect(stripThinkTags('<THINKING>x</THINKING>out')).toBe('out')
+    expect(stripThinkTags('<Think>x</Think>out')).toBe('out')
+  })
+
+  test('handles attributes on open tag', () => {
+    expect(stripThinkTags('<think id="plan-1">reason</think>ok')).toBe('ok')
+  })
+
+  test('strips unterminated open tag at block boundary', () => {
+    expect(stripThinkTags('<think>reasoning that never closes')).toBe('')
+  })
+
+  test('strips unterminated open tag after newline', () => {
+    // Block-boundary match consumes the leading newline, same as hermes.
+    expect(stripThinkTags('Answer: 42\n<think>second-guess myself'))
+      .toBe('Answer: 42')
+  })
+
+  test('strips orphan close tag', () => {
+    expect(stripThinkTags('trailing </think>done')).toBe('trailing done')
+  })
+
+  test('strips multiple blocks', () => {
+    expect(stripThinkTags('<think>a</think>B<think>c</think>D')).toBe('BD')
+  })
+
+  test('handles reasoning mid-response after content', () => {
+    expect(stripThinkTags('Answer: 42\n<think>double-check</think>\nDone'))
+      .toBe('Answer: 42\n\nDone')
+  })
+
+  test('handles nested-looking tags (lazy match + orphan cleanup)', () => {
+    expect(stripThinkTags('<think><think>x</think></think>y')).toBe('y')
+  })
+
+  test('preserves legitimate non-think tags', () => {
+    expect(stripThinkTags('use <div> and <span>')).toBe('use <div> and <span>')
+  })
+
+  test('preserves text without any tags', () => {
+    expect(stripThinkTags('Hello, world. I should respond briefly.')).toBe(
+      'Hello, world. I should respond briefly.',
+    )
+  })
+
+  test('handles empty input', () => {
+    expect(stripThinkTags('')).toBe('')
+  })
+})
+
+describe('createThinkTagFilter — streaming state machine', () => {
+  test('passes through plain text', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('Hello, ')).toBe('Hello, ')
+    expect(f.feed('world!')).toBe('world!')
+    expect(f.flush()).toBe('')
+  })
+
+  test('strips a complete think block in one chunk', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('pre<think>reason</think>post')).toBe('prepost')
+    expect(f.flush()).toBe('')
+  })
+
+  test('handles open tag split across deltas', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('before<th')).toBe('before')
+    expect(f.feed('ink>reason</think>after')).toBe('after')
+    expect(f.flush()).toBe('')
+  })
+
+  test('handles close tag split across deltas', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('<think>reason</th')).toBe('')
+    expect(f.feed('ink>keep')).toBe('keep')
+    expect(f.flush()).toBe('')
+  })
+
+  test('handles tag split on bare < boundary', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('leading <')).toBe('leading ')
+    expect(f.feed('think>inner</think>tail')).toBe('tail')
+    expect(f.flush()).toBe('')
+  })
+
+  test('preserves partial non-tag < at boundary when next char rules it out', () => {
+    const f = createThinkTagFilter()
+    // "<d" — 'd' cannot start any of our tag names, so emit immediately
+    expect(f.feed('pre<d')).toBe('pre<d')
+    expect(f.feed('iv>rest')).toBe('iv>rest')
+    expect(f.flush()).toBe('')
+  })
+
+  test('case-insensitive streaming', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('<THINKING>x</THINKING>out')).toBe('out')
+    expect(f.flush()).toBe('')
+  })
+
+  test('unterminated open tag — flush drops remainder', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('<think>reasoning with no close ')).toBe('')
+    expect(f.feed('and more reasoning')).toBe('')
+    expect(f.flush()).toBe('')
+    expect(f.isInsideBlock()).toBe(false)
+  })
+
+  test('multiple blocks in single feed', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('<think>a</think>B<think>c</think>D')).toBe('BD')
+    expect(f.flush()).toBe('')
+  })
+
+  test('flush after clean stream emits nothing extra', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('complete message')).toBe('complete message')
+    expect(f.flush()).toBe('')
+  })
+
+  test('flush of bare < at end emits it (not a tag prefix)', () => {
+    const f = createThinkTagFilter()
+    // bare '<' held back; flush emits it since it has no tag-name chars
+    expect(f.feed('x <')).toBe('x ')
+    expect(f.flush()).toBe('<')
+  })
+
+  test('flush of partial tag-name prefix at end drops it', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('x <thi')).toBe('x ')
+    expect(f.flush()).toBe('')
+  })
+
+  test('handles attributes on streaming open tag', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('<think type="plan">reason</think>ok')).toBe('ok')
+    expect(f.flush()).toBe('')
+  })
+
+  test('mid-delta transition: content, reasoning, content', () => {
+    const f = createThinkTagFilter()
+    expect(f.feed('Answer: 42\n<think>')).toBe('Answer: 42\n')
+    expect(f.feed('double-check')).toBe('')
+    expect(f.feed('</think>\nDone')).toBe('\nDone')
+    expect(f.flush()).toBe('')
+  })
+
+  test('orphan close tag mid-stream is stripped on flush via safety-net behavior', () => {
+    // Filter alone treats orphan close as "we're not inside", so it emits as-is.
+    // Safety net (stripThinkTags on final text) removes orphans.
+    const f = createThinkTagFilter()
+    const chunk1 = f.feed('trailing ')
+    const chunk2 = f.feed('</think>done')
+    const final = chunk1 + chunk2 + f.flush()
+    // Orphan close appears in stream output; safety net cleans it
+    expect(stripThinkTags(final)).toBe('trailing done')
+  })
+})

src/services/api/thinkTagSanitizer.ts

@@ -0,0 +1,162 @@
+/**
+ * Think-tag sanitizer for reasoning content leaks.
+ *
+ * Some OpenAI-compatible reasoning models (MiniMax M2.7, GLM-4.5/5, DeepSeek, Kimi K2,
+ * self-hosted vLLM builds) emit chain-of-thought inline inside the `content` field using
+ * XML-like tags instead of the separate `reasoning_content` channel. Example:
+ *
+ *   <think>the user wants foo, let me check bar</think>Here is the answer: ...
+ *
+ * This module strips those blocks structurally (tag-based), independent of English
+ * phrasings. Three layers:
+ *
+ *   1. `createThinkTagFilter()` — streaming state machine. Feeds deltas, emits only
+ *      the visible (non-reasoning) portion, and buffers partial tags across chunk
+ *      boundaries so `</th` + `ink>` still parses correctly.
+ *
+ *   2. `stripThinkTags()` — whole-text cleanup. Removes closed pairs, unterminated
+ *      opens at block boundaries, and orphan open/close tags. Used for non-streaming
+ *      responses and as a safety net after stream close.
+ *
+ *   3. Flush discards buffered partial tags at stream end (false-negative bias —
+ *      prefer losing a partial reasoning fragment over leaking it).
+ */
+
+const TAG_NAMES = [
+  'think',
+  'thinking',
+  'reasoning',
+  'thought',
+  'reasoning_scratchpad',
+] as const
+
+const TAG_ALT = TAG_NAMES.join('|')
+
+const OPEN_TAG_RE = new RegExp(`<\\s*(?:${TAG_ALT})\\b[^>]*>`, 'i')
+const CLOSE_TAG_RE = new RegExp(`<\\s*/\\s*(?:${TAG_ALT})\\s*>`, 'i')
+
+const CLOSED_PAIR_RE_G = new RegExp(
+  `<\\s*(${TAG_ALT})\\b[^>]*>[\\s\\S]*?<\\s*/\\s*\\1\\s*>`,
+  'gi',
+)
+const UNTERMINATED_OPEN_RE = new RegExp(
+  `(?:^|\\n)[ \\t]*<\\s*(?:${TAG_ALT})\\b[^>]*>[\\s\\S]*$`,
+  'i',
+)
+const ORPHAN_TAG_RE_G = new RegExp(
+  `<\\s*/?\\s*(?:${TAG_ALT})\\b[^>]*>\\s*`,
+  'gi',
+)
+
+const MAX_PARTIAL_TAG = 64
+
+/**
+ * Remove reasoning/thinking blocks from a complete text body.
+ *
+ * Handles:
+ *   - Closed pairs: <think>...</think> (lazy match, anywhere in text)
+ *   - Unterminated open tags at a block boundary: strips from the tag to end of string
+ *   - Orphan open or close tags (no matching partner)
+ *
+ * False-negative bias: prefers leaving a few tag characters in rare edge cases over
+ * stripping legitimate content.
+ */
+export function stripThinkTags(text: string): string {
+  if (!text) return text
+  let out = text
+  out = out.replace(CLOSED_PAIR_RE_G, '')
+  out = out.replace(UNTERMINATED_OPEN_RE, '')
+  out = out.replace(ORPHAN_TAG_RE_G, '')
+  return out
+}
+
+export interface ThinkTagFilter {
+  feed(chunk: string): string
+  flush(): string
+  isInsideBlock(): boolean
+}
+
+/**
+ * Streaming state machine. Feed deltas, emits visible (non-reasoning) text.
+ * Handles tags split across chunk boundaries by holding back a short tail buffer
+ * whenever the current buffer ends with what looks like a partial tag.
+ */
+export function createThinkTagFilter(): ThinkTagFilter {
+  let inside = false
+  let buffer = ''
+
+  function findPartialTagStart(s: string): number {
+    const lastLt = s.lastIndexOf('<')
+    if (lastLt === -1) return -1
+    if (s.indexOf('>', lastLt) !== -1) return -1
+    const tail = s.slice(lastLt)
+    if (tail.length > MAX_PARTIAL_TAG) return -1
+
+    const m = /^<\s*\/?\s*([a-zA-Z_]\w*)?\s*$/.exec(tail)
+    if (!m) return -1
+    const partialName = (m[1] ?? '').toLowerCase()
+    if (!partialName) return lastLt
+    if (TAG_NAMES.some(name => name.startsWith(partialName))) return lastLt
+    return -1
+  }
+
+  function feed(chunk: string): string {
+    if (!chunk) return ''
+    buffer += chunk
+    let out = ''
+
+    while (buffer.length > 0) {
+      if (!inside) {
+        const open = OPEN_TAG_RE.exec(buffer)
+        if (open) {
+          out += buffer.slice(0, open.index)
+          buffer = buffer.slice(open.index + open[0].length)
+          inside = true
+          continue
+        }
+
+        const partialStart = findPartialTagStart(buffer)
+        if (partialStart === -1) {
+          out += buffer
+          buffer = ''
+        } else {
+          out += buffer.slice(0, partialStart)
+          buffer = buffer.slice(partialStart)
+        }
+        return out
+      }
+
+      const close = CLOSE_TAG_RE.exec(buffer)
+      if (close) {
+        buffer = buffer.slice(close.index + close[0].length)
+        inside = false
+        continue
+      }
+
+      const partialStart = findPartialTagStart(buffer)
+      if (partialStart === -1) {
+        buffer = ''
+      } else {
+        buffer = buffer.slice(partialStart)
+      }
+      return out
+    }
+
+    return out
+  }
+
+  function flush(): string {
+    const held = buffer
+    const wasInside = inside
+    buffer = ''
+    inside = false
+
+    if (wasInside) return ''
+    if (!held) return ''
+
+    if (/^<\s*\/?\s*[a-zA-Z_]/.test(held)) return ''
+    return held
+  }
+
+  return { feed, flush, isInsideBlock: () => inside }
+}

src/services/autoFix/autoFixRunner.test.ts

@@ -70,7 +70,7 @@ describe('runAutoFixCheck', () => {
 
   test('handles timeout gracefully', async () => {
     const result = await runAutoFixCheck({
-      lint: 'sleep 10',
+      lint: 'node -e "setTimeout(() => {}, 10000)"',
       timeout: 100,
 
       cwd: '/tmp',

src/services/autoFix/autoFixRunner.ts

@@ -46,14 +46,31 @@ async function runCommand(
 
     const killTree = () => {
       try {
-        if (!isWindows && proc.pid) {
+        if (isWindows && proc.pid) {
+          // shell=true on Windows can leave child commands running unless we
+          // terminate the full process tree.
+          const killer = spawn('taskkill', ['/pid', String(proc.pid), '/T', '/F'], {
+            windowsHide: true,
+            stdio: 'ignore',
+          })
+          killer.unref()
+          return
+        }
+
+        if (proc.pid) {
           // Kill the entire process group
           process.kill(-proc.pid, 'SIGTERM')
-        } else {
-          proc.kill('SIGTERM')
+          return
         }
+
+        proc.kill('SIGTERM')
       } catch {
-        // Process may have already exited
+        // Process may have already exited; fallback to direct child kill.
+        try {
+          proc.kill('SIGTERM')
+        } catch {
+          // Ignore final fallback errors.
+        }
       }
     }
 

src/services/compact/autoCompact.test.ts

@@ -16,12 +16,21 @@ describe('getEffectiveContextWindowSize', () => {
     // 8k minus 20k summary reservation = -12k, causing infinite auto-compact.
     // Now the fallback is 128k and there's a floor, so effective is always
     // at least reservedTokensForSummary + buffer.
+    //
+    // The exact floor depends on the max-output-tokens slot-reservation cap
+    // (tengu_otk_slot_v1 GrowthBook flag). With cap enabled, the model's
+    // default output cap drops to CAPPED_DEFAULT_MAX_TOKENS (8k), so the
+    // summary reservation is 8k and the floor is 8k + 13k = 21k. With cap
+    // disabled it's 20k + 13k = 33k. Assert the worst case so the test is
+    // stable regardless of flag state in CI vs local.
     process.env.CLAUDE_CODE_USE_OPENAI = '1'
     try {
       const effective = getEffectiveContextWindowSize('some-unknown-3p-model')
       expect(effective).toBeGreaterThan(0)
-      // Must be at least summary reservation (20k) + buffer (13k) = 33k
-      expect(effective).toBeGreaterThanOrEqual(33_000)
+      // 21k = CAPPED_DEFAULT_MAX_TOKENS (8k) + AUTOCOMPACT_BUFFER_TOKENS (13k).
+      // Covers the anti-regression intent of issue #635 without assuming
+      // the GrowthBook flag state.
+      expect(effective).toBeGreaterThanOrEqual(21_000)
     } finally {
       delete process.env.CLAUDE_CODE_USE_OPENAI
     }

src/services/compact/microCompact.ts

@@ -38,7 +38,7 @@ export const TIME_BASED_MC_CLEARED_MESSAGE = '[Old tool result content cleared]'
 const IMAGE_MAX_TOKEN_SIZE = 2000
 
 // Only compact these built-in tools (MCP tools are also compactable via prefix match)
-const COMPACTABLE_TOOLS = new Set<string>([
+export const COMPACTABLE_TOOLS = new Set<string>([
   FILE_READ_TOOL_NAME,
   ...SHELL_TOOL_NAMES,
   GREP_TOOL_NAME,
@@ -51,7 +51,7 @@ const COMPACTABLE_TOOLS = new Set<string>([
 
 const MCP_TOOL_PREFIX = 'mcp__'
 
-function isCompactableTool(name: string): boolean {
+export function isCompactableTool(name: string): boolean {
   return COMPACTABLE_TOOLS.has(name) || name.startsWith(MCP_TOOL_PREFIX)
 }
 

src/services/mcp/client.ts

@@ -2524,7 +2524,7 @@ export async function transformResultContent(
       return [
         {
           type: 'text',
-          text: resultContent.text,
+          text: recursivelySanitizeUnicode(resultContent.text) as string,
         },
       ]
     case 'audio': {
@@ -2569,7 +2569,9 @@ export async function transformResultContent(
         return [
           {
             type: 'text',
-            text: `${prefix}${resource.text}`,
+            text: recursivelySanitizeUnicode(
+              `${prefix}${resource.text}`,
+            ) as string,
           },
         ]
       } else if ('blob' in resource) {

src/services/wiki/init.test.ts

@@ -26,10 +26,10 @@ test('initializeWiki creates the expected wiki scaffold', async () => {
 
   expect(result.alreadyExisted).toBe(false)
   expect(result.createdFiles).toEqual([
-    '.openclaude/wiki/schema.md',
-    '.openclaude/wiki/index.md',
-    '.openclaude/wiki/log.md',
-    '.openclaude/wiki/pages/architecture.md',
+    join('.openclaude', 'wiki', 'schema.md'),
+    join('.openclaude', 'wiki', 'index.md'),
+    join('.openclaude', 'wiki', 'log.md'),
+    join('.openclaude', 'wiki', 'pages', 'architecture.md'),
   ])
   expect(await readFile(paths.schemaFile, 'utf8')).toContain(
     '# OpenClaude Wiki Schema',

src/tools/ConfigTool/prompt.ts

@@ -59,7 +59,7 @@ export function generatePrompt(): string {
 ## Configurable settings list
 The following settings are available for you to change:
 
-### Global Settings (stored in ~/.claude.json)
+### Global Settings (stored in ~/.openclaude.json)
 ${globalSettings.join('\n')}
 
 ### Project Settings (stored in settings.json)

src/tools/WebFetchTool/utils.ts

@@ -15,6 +15,7 @@ import {
 } from '../../utils/mcpOutputStorage.js'
 import { getSettings_DEPRECATED } from '../../utils/settings/settings.js'
 import { asSystemPrompt } from '../../utils/systemPromptType.js'
+import { ssrfGuardedLookup } from '../../utils/hooks/ssrfGuard.js'
 import { isPreapprovedHost } from './preapproved.js'
 import { makeSecondaryModelPrompt } from './prompt.js'
 
@@ -281,6 +282,7 @@ export async function getWithPermittedRedirects(
       maxRedirects: 0,
       responseType: 'arraybuffer',
       maxContentLength: MAX_HTTP_CONTENT_LENGTH,
+      lookup: ssrfGuardedLookup,
       headers: {
         Accept: 'text/markdown, text/html, */*',
         'User-Agent': getWebFetchUserAgent(),

src/utils/auth.ts

@@ -693,7 +693,7 @@ export function refreshAwsAuth(awsAuthRefresh: string): Promise<boolean> {
               'AWS auth refresh timed out after 3 minutes. Run your auth command manually in a separate terminal.',
             )
           : chalk.red(
-              'Error running awsAuthRefresh (in settings or ~/.claude.json):',
+              'Error running awsAuthRefresh (in settings or ~/.openclaude.json):',
             )
         // biome-ignore lint/suspicious/noConsole:: intentional console output
         console.error(message)
@@ -771,7 +771,7 @@ async function getAwsCredsFromCredentialExport(): Promise<{
       }
     } catch (e) {
       const message = chalk.red(
-        'Error getting AWS credentials from awsCredentialExport (in settings or ~/.claude.json):',
+        'Error getting AWS credentials from awsCredentialExport (in settings or ~/.openclaude.json):',
       )
       if (e instanceof Error) {
         // biome-ignore lint/suspicious/noConsole:: intentional console output
@@ -961,7 +961,7 @@ export function refreshGcpAuth(gcpAuthRefresh: string): Promise<boolean> {
               'GCP auth refresh timed out after 3 minutes. Run your auth command manually in a separate terminal.',
             )
           : chalk.red(
-              'Error running gcpAuthRefresh (in settings or ~/.claude.json):',
+              'Error running gcpAuthRefresh (in settings or ~/.openclaude.json):',
             )
         // biome-ignore lint/suspicious/noConsole:: intentional console output
         console.error(message)
@@ -1959,7 +1959,7 @@ export async function validateForceLoginOrg(): Promise<OrgValidationResult> {
 
   // Always fetch the authoritative org UUID from the profile endpoint.
   // Even keychain-sourced tokens verify server-side: the cached org UUID
-  // in ~/.claude.json is user-writable and cannot be trusted.
+  // in ~/.openclaude.json is user-writable and cannot be trusted.
   const { source } = getAuthTokenSource()
   const isEnvVarToken =
     source === 'CLAUDE_CODE_OAUTH_TOKEN' ||

src/utils/caCertsConfig.ts

@@ -28,7 +28,7 @@ import { getSettingsForSource } from './settings/settings.js'
  * is lazy-initialized) and ensure Node.js compatibility.
  *
  * This is safe to call before the trust dialog because we only read from
- * user-controlled files (~/.claude/settings.json and ~/.claude.json),
+ * user-controlled files (~/.claude/settings.json and ~/.openclaude.json),
  * not from project-level settings.
  */
 export function applyExtraCACertsFromConfig(): void {
@@ -52,7 +52,7 @@ export function applyExtraCACertsFromConfig(): void {
  * after the trust dialog. But we need the CA cert early to establish the TLS
  * connection to an HTTPS proxy during init().
  *
- * We read from global config (~/.claude.json) and user settings
+ * We read from global config (~/.openclaude.json) and user settings
  * (~/.claude/settings.json). These are user-controlled files that don't
  * require trust approval.
  */

src/utils/claudeInChrome/setup.ts

@@ -355,7 +355,7 @@ exec ${command}
  *
  * Only positive detections are persisted. A negative result from the
  * filesystem scan is not cached, because it may come from a machine that
- * shares ~/.claude.json but has no local Chrome (e.g. a remote dev
+ * shares ~/.openclaude.json but has no local Chrome (e.g. a remote dev
  * environment using the bridge), and caching it would permanently poison
  * auto-enable for every session on every machine that reads that config.
  */

src/utils/config.ts

@@ -244,6 +244,7 @@ export type GlobalConfig = {
   bypassPermissionsModeAccepted?: boolean
   hasUsedBackslashReturn?: boolean
   autoCompactEnabled: boolean // Controls whether auto-compact is enabled
+  toolHistoryCompressionEnabled: boolean // Compress old tool_result content for small-context providers
   showTurnDuration: boolean // Controls whether to show turn duration message (e.g., "Cooked for 1m 6s")
   /**
    * @deprecated Use settings.env instead.
@@ -622,6 +623,7 @@ function createDefaultGlobalConfig(): GlobalConfig {
     verbose: false,
     editorMode: 'normal',
     autoCompactEnabled: true,
+    toolHistoryCompressionEnabled: true,
     showTurnDuration: true,
     hasSeenTasksHint: false,
     hasUsedStash: false,
@@ -668,6 +670,7 @@ export const GLOBAL_CONFIG_KEYS = [
   'editorMode',
   'hasUsedBackslashReturn',
   'autoCompactEnabled',
+  'toolHistoryCompressionEnabled',
   'showTurnDuration',
   'diffTool',
   'env',
@@ -918,7 +921,7 @@ let configCacheHits = 0
 let configCacheMisses = 0
 // Session-total count of actual disk writes to the global config file.
 // Exposed for internal-only dev diagnostics (see inc-4552) so anomalous write
-// rates surface in the UI before they corrupt ~/.claude.json.
+// rates surface in the UI before they corrupt ~/.openclaude.json.
 let globalConfigWriteCount = 0
 
 export function getGlobalConfigWriteCount(): number {
@@ -1257,7 +1260,7 @@ function saveConfigWithLock<A extends object>(
     const currentConfig = getConfig(file, createDefault)
     if (file === getGlobalClaudeFile() && wouldLoseAuthState(currentConfig)) {
       logForDebugging(
-        'saveConfigWithLock: re-read config is missing auth that cache has; refusing to write to avoid wiping ~/.claude.json. See GH #3117.',
+        'saveConfigWithLock: re-read config is missing auth that cache has; refusing to write to avoid wiping ~/.openclaude.json. See GH #3117.',
         { level: 'error' },
       )
       logEvent('tengu_config_auth_loss_prevented', {})

src/utils/deepLink/registerProtocol.ts

@@ -253,7 +253,7 @@ async function resolveClaudePath(): Promise<string> {
  * Check whether the OS-level protocol handler is already registered AND
  * points at the expected `claude` binary. Reads the registration artifact
  * directly (symlink target, .desktop Exec line, registry value) rather than
- * a cached flag in ~/.claude.json, so:
+ * a cached flag in ~/.openclaude.json, so:
  *   - the check is per-machine (config can sync across machines; OS state can't)
  *   - stale paths self-heal (install-method change → re-register next session)
  *   - deleted artifacts self-heal
@@ -311,7 +311,7 @@ export async function ensureDeepLinkProtocolRegistered(): Promise<void> {
   // EACCES/ENOSPC are deterministic — retrying next session won't help.
   // Throttle to once per 24h so a read-only ~/.local/share/applications
   // doesn't generate a failure event on every startup. Marker lives in
-  // ~/.claude (per-machine, not synced) rather than ~/.claude.json (can sync).
+  // ~/.claude (per-machine, not synced) rather than ~/.openclaude.json (can sync).
   const failureMarkerPath = path.join(
     getClaudeConfigHomeDir(),
     '.deep-link-register-failed',

src/utils/env.test.ts

@@ -0,0 +1,62 @@
+import { afterEach, beforeEach, expect, test } from 'bun:test'
+import { mkdtempSync, rmSync, writeFileSync } from 'fs'
+import { tmpdir } from 'os'
+import { join } from 'path'
+
+const originalEnv = {
+  CLAUDE_CONFIG_DIR: process.env.CLAUDE_CONFIG_DIR,
+  CLAUDE_CODE_CUSTOM_OAUTH_URL: process.env.CLAUDE_CODE_CUSTOM_OAUTH_URL,
+  USER_TYPE: process.env.USER_TYPE,
+}
+
+let tempDir: string
+
+beforeEach(() => {
+  tempDir = mkdtempSync(join(tmpdir(), 'openclaude-env-test-'))
+  process.env.CLAUDE_CONFIG_DIR = tempDir
+  delete process.env.CLAUDE_CODE_CUSTOM_OAUTH_URL
+  delete process.env.USER_TYPE
+})
+
+afterEach(() => {
+  rmSync(tempDir, { recursive: true, force: true })
+  if (originalEnv.CLAUDE_CONFIG_DIR === undefined) {
+    delete process.env.CLAUDE_CONFIG_DIR
+  } else {
+    process.env.CLAUDE_CONFIG_DIR = originalEnv.CLAUDE_CONFIG_DIR
+  }
+  if (originalEnv.CLAUDE_CODE_CUSTOM_OAUTH_URL === undefined) {
+    delete process.env.CLAUDE_CODE_CUSTOM_OAUTH_URL
+  } else {
+    process.env.CLAUDE_CODE_CUSTOM_OAUTH_URL = originalEnv.CLAUDE_CODE_CUSTOM_OAUTH_URL
+  }
+  if (originalEnv.USER_TYPE === undefined) {
+    delete process.env.USER_TYPE
+  } else {
+    process.env.USER_TYPE = originalEnv.USER_TYPE
+  }
+})
+
+async function importFreshEnvModule() {
+  return import(`./env.js?ts=${Date.now()}-${Math.random()}`)
+}
+
+// getGlobalClaudeFile — three migration branches
+
+test('getGlobalClaudeFile: new install returns .openclaude.json when neither file exists', async () => {
+  const { getGlobalClaudeFile } = await importFreshEnvModule()
+  expect(getGlobalClaudeFile()).toBe(join(tempDir, '.openclaude.json'))
+})
+
+test('getGlobalClaudeFile: existing user keeps .claude.json when only legacy file exists', async () => {
+  writeFileSync(join(tempDir, '.claude.json'), '{}')
+  const { getGlobalClaudeFile } = await importFreshEnvModule()
+  expect(getGlobalClaudeFile()).toBe(join(tempDir, '.claude.json'))
+})
+
+test('getGlobalClaudeFile: migrated user uses .openclaude.json when both files exist', async () => {
+  writeFileSync(join(tempDir, '.claude.json'), '{}')
+  writeFileSync(join(tempDir, '.openclaude.json'), '{}')
+  const { getGlobalClaudeFile } = await importFreshEnvModule()
+  expect(getGlobalClaudeFile()).toBe(join(tempDir, '.openclaude.json'))
+})

src/utils/env.ts

@@ -21,8 +21,21 @@ export const getGlobalClaudeFile = memoize((): string => {
     return join(getClaudeConfigHomeDir(), '.config.json')
   }
 
-  const filename = `.claude${fileSuffixForOauthConfig()}.json`
-  return join(process.env.CLAUDE_CONFIG_DIR || homedir(), filename)
+  const oauthSuffix = fileSuffixForOauthConfig()
+  const configDir = process.env.CLAUDE_CONFIG_DIR || homedir()
+
+  // Default to .openclaude.json. Fall back to .claude.json only if the new
+  // file doesn't exist yet and the legacy one does (same migration pattern
+  // as resolveClaudeConfigHomeDir for the config directory).
+  const newFilename = `.openclaude${oauthSuffix}.json`
+  const legacyFilename = `.claude${oauthSuffix}.json`
+  if (
+    !getFsImplementation().existsSync(join(configDir, newFilename)) &&
+    getFsImplementation().existsSync(join(configDir, legacyFilename))
+  ) {
+    return join(configDir, legacyFilename)
+  }
+  return join(configDir, newFilename)
 })
 
 const hasInternetAccess = memoize(async (): Promise<boolean> => {

src/utils/json.ts

@@ -24,7 +24,7 @@ type CachedParse = { ok: true; value: unknown } | { ok: false }
 // lodash memoize default resolver = first arg only).
 // Skip caching above this size — the LRU stores the full string as the key,
 // so a 200KB config file would pin ~10MB in #keyList across 50 slots. Large
-// inputs like ~/.claude.json also change between reads (numStartups bumps on
+// inputs like ~/.openclaude.json also change between reads (numStartups bumps on
 // every CC startup), so the cache never hits anyway.
 const PARSE_CACHE_MAX_KEY_BYTES = 8 * 1024
 

src/utils/localInstaller.ts

@@ -44,9 +44,10 @@ function getCandidateLocalBinaryPaths(localInstallDir: string): string[] {
 }
 
 export function isManagedLocalInstallationPath(execPath: string): boolean {
+  const normalizedExecPath = execPath.replace(/\\+/g, '/')
   return (
-    execPath.includes('/.openclaude/local/node_modules/') ||
-    execPath.includes('/.claude/local/node_modules/')
+    normalizedExecPath.includes('/.openclaude/local/node_modules/') ||
+    normalizedExecPath.includes('/.claude/local/node_modules/')
   )
 }
 

src/utils/managedEnv.ts

@@ -131,7 +131,7 @@ export function applySafeConfigEnvironmentVariables(): void {
         : null
   }
 
-  // Global config (~/.claude.json) is user-controlled. In CCD mode,
+  // Global config (~/.openclaude.json) is user-controlled. In CCD mode,
   // filterSettingsEnv strips keys that were in the spawn env snapshot so
   // the desktop host's operational vars (OTEL, etc.) are not overridden.
   Object.assign(process.env, filterSettingsEnv(getGlobalConfig().env))

src/utils/managedEnvConstants.ts

@@ -123,7 +123,6 @@ export const SAFE_ENV_VARS = new Set([
   'ANTHROPIC_DEFAULT_SONNET_MODEL_DESCRIPTION',
   'ANTHROPIC_DEFAULT_SONNET_MODEL_NAME',
   'ANTHROPIC_DEFAULT_SONNET_MODEL_SUPPORTED_CAPABILITIES',
-  'ANTHROPIC_FOUNDRY_API_KEY',
   'ANTHROPIC_MODEL',
   'ANTHROPIC_SMALL_FAST_MODEL_AWS_REGION',
   'ANTHROPIC_SMALL_FAST_MODEL',

src/utils/model/benchmark.ts

@@ -0,0 +1,205 @@
+/**
+ * Model Benchmarking for OpenClaude
+ * 
+ * Tests and compares model speed/quality for informed model selection.
+ * Supports OpenAI-compatible, Ollama, Anthropic, Bedrock, Vertex.
+ */
+
+import { getAPIProvider } from './providers.js'
+
+export interface BenchmarkResult {
+  model: string
+  provider: string
+  firstTokenMs: number
+  totalTokens: number
+  tokensPerSecond: number
+  success: boolean
+  error?: string
+}
+
+const TEST_PROMPT = 'Write a short hello world in Python.'
+const MAX_TOKENS = 50
+const TIMEOUT_MS = 30000
+
+function getBenchmarkEndpoint(): string | null {
+  const provider = getAPIProvider()
+  const baseUrl = process.env.OPENAI_BASE_URL
+  
+  // Check for Ollama (local)
+  if (baseUrl?.includes('localhost:11434') || baseUrl?.includes('localhost:11435')) {
+    return `${baseUrl}/chat/completions`
+  }
+  // OpenAI-compatible endpoints
+  if (provider === 'openai' || provider === 'firstParty') {
+    return `${baseUrl || 'https://api.openai.com/v1'}/chat/completions`
+  }
+  // NVIDIA NIM or MiniMax via OPENAI_BASE_URL
+  if (baseUrl?.includes('nvidia') || baseUrl?.includes('minimax')) {
+    return `${baseUrl}/chat/completions`
+  }
+  return null
+}
+
+function getBenchmarkAuthHeader(): string | null {
+  const apiKey = process.env.OPENAI_API_KEY
+  if (!apiKey) return null
+  return `Bearer ${apiKey}`
+}
+
+export async function benchmarkModel(
+  model: string,
+  onChunk?: (text: string) => void,
+): Promise<BenchmarkResult> {
+  const endpoint = getBenchmarkEndpoint()
+  const authHeader = getBenchmarkAuthHeader()
+  
+  if (!endpoint || !authHeader) {
+    return {
+      model,
+      provider: getAPIProvider(),
+      firstTokenMs: 0,
+      totalTokens: 0,
+      tokensPerSecond: 0,
+      success: false,
+      error: 'Benchmark not supported for this provider',
+    }
+  }
+  
+  const startTime = performance.now()
+  let totalTokens = 0
+  let firstTokenMs: number | null = null
+
+  try {
+    const response = await fetch(endpoint, {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'application/json',
+        'Authorization': authHeader,
+      },
+      body: JSON.stringify({
+        model,
+        messages: [{ role: 'user', content: TEST_PROMPT }],
+        max_tokens: MAX_TOKENS,
+        stream: true,
+      }),
+      signal: AbortSignal.timeout(TIMEOUT_MS),
+    })
+
+    if (!response.ok) {
+      let errorMsg = `HTTP ${response.status}`
+      try {
+        const error = await response.json()
+        errorMsg = error.error?.message || errorMsg
+      } catch {
+        // ignore
+      }
+      return {
+        model,
+        provider: getAPIProvider(),
+        firstTokenMs: 0,
+        totalTokens: 0,
+        tokensPerSecond: 0,
+        success: false,
+        error: errorMsg,
+      }
+    }
+
+    const reader = response.body?.getReader()
+    if (!reader) {
+      throw new Error('No response body')
+    }
+
+    const decoder = new TextDecoder()
+    let buffer = ''
+
+    while (true) {
+      const { done, value } = await reader.read()
+      if (done) break
+
+      buffer += decoder.decode(value, { stream: true })
+      const lines = buffer.split('\n')
+      buffer = lines.pop() || ''
+
+      for (const line of lines) {
+        if (line.startsWith('data: ')) {
+          const data = line.slice(6)
+          if (data === '[DONE]') continue
+
+          try {
+            const json = JSON.parse(data)
+            const content = json.choices?.[0]?.delta?.content
+            if (content) {
+              if (firstTokenMs === null) {
+                firstTokenMs = performance.now() - startTime
+              }
+              totalTokens += content.length / 4
+              onChunk?.(content)
+            }
+          } catch {
+            // skip invalid JSON
+          }
+        }
+      }
+    }
+
+    const totalMs = performance.now() - startTime
+    const tokensPerSecond = totalMs > 0 ? (totalTokens / totalMs) * 1000 : 0
+
+    return {
+      model,
+      provider: getAPIProvider(),
+      firstTokenMs: firstTokenMs ?? 0,
+      totalTokens,
+      tokensPerSecond,
+      success: true,
+    }
+  } catch (error) {
+    return {
+      model,
+      provider: getAPIProvider(),
+      firstTokenMs: 0,
+      totalTokens: 0,
+      tokensPerSecond: 0,
+      success: false,
+      error: error instanceof Error ? error.message : 'Unknown error',
+    }
+  }
+}
+
+export async function benchmarkMultipleModels(
+  models: string[],
+  onProgress?: (completed: number, total: number, result: BenchmarkResult) => void,
+): Promise<BenchmarkResult[]> {
+  const results: BenchmarkResult[] = []
+
+  for (let i = 0; i < models.length; i++) {
+    const result = await benchmarkModel(models[i])
+    results.push(result)
+    onProgress?.(i + 1, models.length, result)
+  }
+
+  return results
+}
+
+export function formatBenchmarkResults(results: BenchmarkResult[]): string {
+  const header = 'Model'.padEnd(40) + 'TPS' + '  First Token' + '  Status'
+  const divider = '-'.repeat(70)
+  
+  const rows = results
+    .sort((a, b) => b.tokensPerSecond - a.tokensPerSecond)
+    .map(r => {
+      const name = r.model.length > 38 ? r.model.slice(0, 37) + '…' : r.model
+      const tps = r.tokensPerSecond.toFixed(1).padStart(6)
+      const first = r.firstTokenMs > 0 ? `${r.firstTokenMs.toFixed(0)}ms`.padStart(12) : 'N/A'.padStart(12)
+      const status = r.success ? '✓' : '✗'
+      return name.padEnd(40) + tps + '  ' + first + '  ' + status
+    })
+
+  return [header, divider, ...rows].join('\n')
+}
+
+export function isBenchmarkSupported(): boolean {
+  const endpoint = getBenchmarkEndpoint()
+  const authHeader = getBenchmarkAuthHeader()
+  return endpoint !== null && authHeader !== null
+}

src/utils/model/configs.ts

@@ -20,7 +20,7 @@ export const OPENAI_MODEL_DEFAULTS = {
 // Override with GEMINI_MODEL env var.
 // ---------------------------------------------------------------------------
 export const GEMINI_MODEL_DEFAULTS = {
-  opus: 'gemini-2.5-pro-preview-03-25',   // most capable
+  opus: 'gemini-2.5-pro',   // most capable
   sonnet: 'gemini-2.0-flash',              // balanced
   haiku: 'gemini-2.0-flash-lite',          // fast & cheap
 } as const
@@ -112,7 +112,7 @@ export const CLAUDE_OPUS_4_CONFIG = {
   vertex: 'claude-opus-4@20250514',
   foundry: 'claude-opus-4',
   openai: 'gpt-4o',
-  gemini: 'gemini-2.5-pro-preview-03-25',
+  gemini: 'gemini-2.5-pro',
   github: 'github:copilot',
   codex: 'gpt-5.4',
   'nvidia-nim': 'nvidia/llama-3.1-nemotron-70b-instruct',
@@ -125,7 +125,7 @@ export const CLAUDE_OPUS_4_1_CONFIG = {
   vertex: 'claude-opus-4-1@20250805',
   foundry: 'claude-opus-4-1',
   openai: 'gpt-4o',
-  gemini: 'gemini-2.5-pro-preview-03-25',
+  gemini: 'gemini-2.5-pro',
   github: 'github:copilot',
   codex: 'gpt-5.4',
   'nvidia-nim': 'nvidia/llama-3.1-nemotron-70b-instruct',
@@ -138,7 +138,7 @@ export const CLAUDE_OPUS_4_5_CONFIG = {
   vertex: 'claude-opus-4-5@20251101',
   foundry: 'claude-opus-4-5',
   openai: 'gpt-4o',
-  gemini: 'gemini-2.5-pro-preview-03-25',
+  gemini: 'gemini-2.5-pro',
   github: 'github:copilot',
   codex: 'gpt-5.4',
   'nvidia-nim': 'nvidia/llama-3.1-nemotron-70b-instruct',
@@ -151,7 +151,7 @@ export const CLAUDE_OPUS_4_6_CONFIG = {
   vertex: 'claude-opus-4-6',
   foundry: 'claude-opus-4-6',
   openai: 'gpt-4o',
-  gemini: 'gemini-2.5-pro-preview-03-25',
+  gemini: 'gemini-2.5-pro',
   github: 'github:copilot',
   codex: 'gpt-5.4',
   'nvidia-nim': 'nvidia/llama-3.1-nemotron-70b-instruct',

src/utils/model/model.openai-shim-providers.test.ts

@@ -0,0 +1,115 @@
+import { afterEach, beforeEach, expect, test } from 'bun:test'
+
+import { saveGlobalConfig } from '../config.js'
+import { getUserSpecifiedModelSetting } from './model.js'
+
+const SAVED_ENV = {
+  CLAUDE_CODE_USE_OPENAI: process.env.CLAUDE_CODE_USE_OPENAI,
+  CLAUDE_CODE_USE_GEMINI: process.env.CLAUDE_CODE_USE_GEMINI,
+  CLAUDE_CODE_USE_GITHUB: process.env.CLAUDE_CODE_USE_GITHUB,
+  CLAUDE_CODE_USE_MISTRAL: process.env.CLAUDE_CODE_USE_MISTRAL,
+  CLAUDE_CODE_USE_BEDROCK: process.env.CLAUDE_CODE_USE_BEDROCK,
+  CLAUDE_CODE_USE_VERTEX: process.env.CLAUDE_CODE_USE_VERTEX,
+  CLAUDE_CODE_USE_FOUNDRY: process.env.CLAUDE_CODE_USE_FOUNDRY,
+  NVIDIA_NIM: process.env.NVIDIA_NIM,
+  MINIMAX_API_KEY: process.env.MINIMAX_API_KEY,
+  OPENAI_MODEL: process.env.OPENAI_MODEL,
+  OPENAI_BASE_URL: process.env.OPENAI_BASE_URL,
+  CODEX_API_KEY: process.env.CODEX_API_KEY,
+  CHATGPT_ACCOUNT_ID: process.env.CHATGPT_ACCOUNT_ID,
+}
+
+function restoreEnv(key: keyof typeof SAVED_ENV): void {
+  if (SAVED_ENV[key] === undefined) {
+    delete process.env[key]
+  } else {
+    process.env[key] = SAVED_ENV[key]
+  }
+}
+
+beforeEach(() => {
+  delete process.env.CLAUDE_CODE_USE_OPENAI
+  delete process.env.CLAUDE_CODE_USE_GEMINI
+  delete process.env.CLAUDE_CODE_USE_GITHUB
+  delete process.env.CLAUDE_CODE_USE_MISTRAL
+  delete process.env.CLAUDE_CODE_USE_BEDROCK
+  delete process.env.CLAUDE_CODE_USE_VERTEX
+  delete process.env.CLAUDE_CODE_USE_FOUNDRY
+  delete process.env.NVIDIA_NIM
+  delete process.env.MINIMAX_API_KEY
+  delete process.env.OPENAI_MODEL
+  delete process.env.OPENAI_BASE_URL
+  delete process.env.CODEX_API_KEY
+  delete process.env.CHATGPT_ACCOUNT_ID
+  saveGlobalConfig(current => ({
+    ...current,
+    model: undefined,
+  }))
+})
+
+afterEach(() => {
+  for (const key of Object.keys(SAVED_ENV) as Array<keyof typeof SAVED_ENV>) {
+    restoreEnv(key)
+  }
+  saveGlobalConfig(current => ({
+    ...current,
+    model: undefined,
+  }))
+})
+
+test('codex provider reads OPENAI_MODEL, not stale settings.model', () => {
+  // Regression: switching from Moonshot (settings.model='kimi-k2.6' persisted
+  // from that session) to the Codex profile. Codex profile correctly sets
+  // OPENAI_MODEL=codexplan + base URL to chatgpt.com/backend-api/codex.
+  // getUserSpecifiedModelSetting previously ignored env for 'codex' provider
+  // and returned settings.model='kimi-k2.6', causing Codex's API to reject
+  // the request: "The 'kimi-k2.6' model is not supported when using Codex".
+  saveGlobalConfig(current => ({ ...current, model: 'kimi-k2.6' }))
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  process.env.OPENAI_BASE_URL = 'https://chatgpt.com/backend-api/codex'
+  process.env.OPENAI_MODEL = 'codexplan'
+  process.env.CODEX_API_KEY = 'codex-test'
+  process.env.CHATGPT_ACCOUNT_ID = 'acct_test'
+
+  const model = getUserSpecifiedModelSetting()
+  expect(model).toBe('codexplan')
+})
+
+test('nvidia-nim provider reads OPENAI_MODEL, not stale settings.model', () => {
+  saveGlobalConfig(current => ({ ...current, model: 'kimi-k2.6' }))
+  process.env.NVIDIA_NIM = '1'
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  process.env.OPENAI_MODEL = 'nvidia/llama-3.1-nemotron-70b-instruct'
+
+  const model = getUserSpecifiedModelSetting()
+  expect(model).toBe('nvidia/llama-3.1-nemotron-70b-instruct')
+})
+
+test('minimax provider reads OPENAI_MODEL, not stale settings.model', () => {
+  saveGlobalConfig(current => ({ ...current, model: 'kimi-k2.6' }))
+  process.env.MINIMAX_API_KEY = 'minimax-test'
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  process.env.OPENAI_MODEL = 'MiniMax-M2.5'
+
+  const model = getUserSpecifiedModelSetting()
+  expect(model).toBe('MiniMax-M2.5')
+})
+
+test('openai provider still reads OPENAI_MODEL (regression guard)', () => {
+  saveGlobalConfig(current => ({ ...current, model: 'stale-default' }))
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  process.env.OPENAI_MODEL = 'gpt-4o'
+
+  const model = getUserSpecifiedModelSetting()
+  expect(model).toBe('gpt-4o')
+})
+
+test('github provider still reads OPENAI_MODEL (regression guard)', () => {
+  saveGlobalConfig(current => ({ ...current, model: 'stale-default' }))
+  process.env.CLAUDE_CODE_USE_GITHUB = '1'
+  process.env.OPENAI_MODEL = 'github:copilot'
+
+  const model = getUserSpecifiedModelSetting()
+  expect(model).toBe('github:copilot')
+})
+

src/utils/model/model.ts

@@ -91,11 +91,24 @@ export function getUserSpecifiedModelSetting(): ModelSetting | undefined {
     const setting = normalizeModelSetting(settings.model)
     // Read the model env var that matches the active provider to prevent
     // cross-provider leaks (e.g. ANTHROPIC_MODEL sent to the OpenAI API).
+    //
+    // All OpenAI-shim providers (openai, codex, github, nvidia-nim, minimax)
+    // set CLAUDE_CODE_USE_OPENAI=1 + OPENAI_MODEL via
+    // applyProviderProfileToProcessEnv. Earlier this check only included
+    // openai/github — codex/nvidia-nim/minimax fell through to the stale
+    // settings.model, so switching from (say) Moonshot to Codex kept firing
+    // `kimi-k2.6` at the Codex endpoint and getting 400s.
     const provider = getAPIProvider()
+    const isOpenAIShimProvider =
+      provider === 'openai' ||
+      provider === 'codex' ||
+      provider === 'github' ||
+      provider === 'nvidia-nim' ||
+      provider === 'minimax'
     specifiedModel =
       (provider === 'gemini' ? process.env.GEMINI_MODEL : undefined) ||
       (provider === 'mistral' ? process.env.MISTRAL_MODEL : undefined) ||
-      (provider === 'openai' || provider === 'gemini' || provider === 'mistral' || provider === 'github' ? process.env.OPENAI_MODEL : undefined) ||
+      (isOpenAIShimProvider ? process.env.OPENAI_MODEL : undefined) ||
       (provider === 'firstParty' ? process.env.ANTHROPIC_MODEL : undefined) ||
       setting ||
       undefined
@@ -140,7 +153,7 @@ export function getDefaultOpusModel(): ModelName {
   }
   // Gemini provider
   if (getAPIProvider() === 'gemini') {
-    return process.env.GEMINI_MODEL || 'gemini-2.5-pro-preview-03-25'
+    return process.env.GEMINI_MODEL || 'gemini-2.5-pro'
   }
   // Mistral provider
   if (getAPIProvider() === 'mistral') {

src/utils/model/modelCache.test.ts

@@ -0,0 +1,30 @@
+import { describe, expect, it, beforeEach, afterEach, vi } from 'bun:test'
+import { isModelCacheValid, getCachedModelsFromDisk, saveModelsToCache } from '../model/modelCache.js'
+
+vi.mock('../model/ollamaModels.js', () => ({
+  isOllamaProvider: vi.fn(() => true),
+}))
+
+describe('modelCache', () => {
+  const mockModel = { value: 'llama3', label: 'Llama 3', description: 'Test model' }
+
+  describe('isModelCacheValid', () => {
+    it('returns false for non-existent cache', async () => {
+      const result = await isModelCacheValid('ollama')
+      expect(result).toBe(false)
+    })
+  })
+
+  describe('getCachedModelsFromDisk', () => {
+    it('returns null when not cache available', async () => {
+      const result = await getCachedModelsFromDisk()
+      expect(result).toBeNull()
+    })
+  })
+
+  describe('saveModelsToCache', () => {
+    it('has saveModelsToCache function', () => {
+      expect(typeof saveModelsToCache).toBe('function')
+    })
+  })
+})

src/utils/model/modelCache.ts

@@ -0,0 +1,165 @@
+/**
+ * Model Caching for OpenClaude
+ * 
+ * Caches model lists to disk for faster startup and offline access.
+ * Uses async fs operations to avoid blocking the event loop.
+ */
+
+import { access, readFile, writeFile, mkdir, unlink } from 'node:fs/promises'
+import { existsSync } from 'node:fs'
+import { join } from 'node:path'
+import { homedir } from 'node:os'
+import { getAPIProvider } from './providers.js'
+
+const CACHE_VERSION = '1'
+const CACHE_TTL_HOURS = 24
+const CACHE_DIR_NAME = '.openclaude-model-cache'
+
+interface ModelCache {
+  version: string
+  timestamp: number
+  provider: string
+  models: Array<{ value: string; label: string; description: string }>
+}
+
+function getCacheDir(): string {
+  const home = homedir()
+  const cacheDir = join(home, CACHE_DIR_NAME)
+  if (!existsSync(cacheDir)) {
+    mkdir(cacheDir, { recursive: true })
+  }
+  return cacheDir
+}
+
+function getCacheFilePath(provider: string): string {
+  return join(getCacheDir(), `${provider}.json`)
+}
+
+function isOpenAICompatibleProvider(): boolean {
+  const baseUrl = process.env.OPENAI_BASE_URL || ''
+  return baseUrl.includes('localhost') || baseUrl.includes('nvidia') || baseUrl.includes('minimax') || getAPIProvider() === 'openai'
+}
+
+export async function isModelCacheValid(provider: string): Promise<boolean> {
+  const cachePath = getCacheFilePath(provider)
+  
+  try {
+    await access(cachePath)
+  } catch {
+    return false
+  }
+
+  try {
+    const data = JSON.parse(await readFile(cachePath, 'utf-8')) as ModelCache
+    if (data.version !== CACHE_VERSION) {
+      return false
+    }
+    if (data.provider !== provider) {
+      return false
+    }
+
+    const ageHours = (Date.now() - data.timestamp) / (1000 * 60 * 60)
+    return ageHours < CACHE_TTL_HOURS
+  } catch {
+    return false
+  }
+}
+
+export async function getCachedModelsFromDisk<T>(): Promise<T[] | null> {
+  const provider = getAPIProvider()
+  const baseUrl = process.env.OPENAI_BASE_URL || ''
+  const isLocalOllama = baseUrl.includes('localhost:11434') || baseUrl.includes('localhost:11435')
+  const isNvidia = baseUrl.includes('nvidia') || baseUrl.includes('integrate.api.nvidia')
+  const isMiniMax = baseUrl.includes('minimax')
+  
+  if (!isLocalOllama && !isNvidia && !isMiniMax && provider !== 'openai') {
+    return null
+  }
+
+  const cachePath = getCacheFilePath(provider)
+  
+  if (!(await isModelCacheValid(provider))) {
+    return null
+  }
+
+  try {
+    const data = JSON.parse(await readFile(cachePath, 'utf-8')) as ModelCache
+    return data.models as T[]
+  } catch {
+    return null
+  }
+}
+
+export async function saveModelsToCache(
+  models: Array<{ value: string; label: string; description: string }>,
+): Promise<void> {
+  const provider = getAPIProvider()
+  if (!provider) return
+
+  const cachePath = getCacheFilePath(provider)
+  const cacheData: ModelCache = {
+    version: CACHE_VERSION,
+    timestamp: Date.now(),
+    provider,
+    models,
+  }
+  
+  try {
+    await writeFile(cachePath, JSON.stringify(cacheData, null, 2), 'utf-8')
+  } catch (error) {
+    console.warn('[ModelCache] Failed to save cache:', error)
+  }
+}
+
+export async function clearModelCache(provider?: string): Promise<void> {
+  if (provider) {
+    const cachePath = getCacheFilePath(provider)
+    try {
+      await unlink(cachePath)
+    } catch {
+      // ignore if doesn't exist
+    }
+  } else {
+    const cacheDir = getCacheDir()
+    try {
+      await unlink(join(cacheDir, 'ollama.json'))
+      await unlink(join(cacheDir, 'nvidia-nim.json'))
+      await unlink(join(cacheDir, 'minimax.json'))
+    } catch {
+      // ignore
+    }
+  }
+}
+
+export async function getModelCacheInfo(): Promise<{ provider: string; age: string } | null> {
+  const provider = getAPIProvider()
+  const cachePath = getCacheFilePath(provider)
+  
+  try {
+    await access(cachePath)
+  } catch {
+    return null
+  }
+
+  try {
+    const data = JSON.parse(await readFile(cachePath, 'utf-8')) as ModelCache
+    const ageMs = Date.now() - data.timestamp
+    const ageHours = Math.floor(ageMs / (1000 * 60 * 60))
+    const ageMins = Math.floor((ageMs % (1000 * 60 * 60)) / (1000 * 60))
+    
+    return {
+      provider: data.provider,
+      age: ageHours > 0 ? `${ageHours}h ${ageMins}m` : `${ageMins}m`,
+    }
+  } catch {
+    return null
+  }
+}
+
+export function isCacheAvailable(): boolean {
+  const baseUrl = process.env.OPENAI_BASE_URL || ''
+  const isLocalOllama = baseUrl.includes('localhost:11434') || baseUrl.includes('localhost:11435')
+  const isNvidia = baseUrl.includes('nvidia') || baseUrl.includes('integrate.api.nvidia')
+  const isMiniMax = baseUrl.includes('minimax')
+  return isLocalOllama || isNvidia || isMiniMax || getAPIProvider() === 'openai'
+}

src/utils/model/openaiContextWindows.ts

@@ -219,6 +219,17 @@ const OPENAI_CONTEXT_WINDOWS: Record<string, number> = {
   'kimi-k2.5':                262_144,
   'glm-5':                    202_752,
   'glm-4.7':                  202_752,
+
+  // Moonshot AI direct API (api.moonshot.ai/v1). Values from Moonshot's
+  // published model card — all K2 tier share 256K context. Prefix matching
+  // in lookupByKey catches variants like "kimi-k2.6-preview".
+  'kimi-k2.6':                262_144,
+  'kimi-k2':                  131_072,
+  'kimi-k2-instruct':         131_072,
+  'kimi-k2-thinking':         262_144,
+  'moonshot-v1-8k':             8_192,
+  'moonshot-v1-32k':           32_768,
+  'moonshot-v1-128k':         131_072,
 }
 
 /**
@@ -391,6 +402,15 @@ const OPENAI_MAX_OUTPUT_TOKENS: Record<string, number> = {
   'kimi-k2.5':                 32_768,
   'glm-5':                     16_384,
   'glm-4.7':                   16_384,
+
+  // Moonshot AI direct API
+  'kimi-k2.6':                 32_768,
+  'kimi-k2':                   32_768,
+  'kimi-k2-instruct':          32_768,
+  'kimi-k2-thinking':          32_768,
+  'moonshot-v1-8k':             4_096,
+  'moonshot-v1-32k':           16_384,
+  'moonshot-v1-128k':          32_768,
 }
 
 function lookupByModel<T>(table: Record<string, T>, model: string): T | undefined {

src/utils/model/providers.test.ts

@@ -107,3 +107,60 @@ test('official OpenAI base URLs now keep provider detection on openai for aliase
   const { getAPIProvider } = await importFreshProvidersModule()
   expect(getAPIProvider()).toBe('openai')
 })
+
+// isGithubNativeAnthropicMode
+
+test('isGithubNativeAnthropicMode: false when CLAUDE_CODE_USE_GITHUB is not set', async () => {
+  clearProviderEnv()
+  process.env.OPENAI_MODEL = 'claude-sonnet-4-5'
+  const { isGithubNativeAnthropicMode } = await importFreshProvidersModule()
+  expect(isGithubNativeAnthropicMode()).toBe(false)
+})
+
+test('isGithubNativeAnthropicMode: true for bare claude- model via OPENAI_MODEL', async () => {
+  clearProviderEnv()
+  process.env.CLAUDE_CODE_USE_GITHUB = '1'
+  process.env.OPENAI_MODEL = 'claude-sonnet-4-5'
+  const { isGithubNativeAnthropicMode } = await importFreshProvidersModule()
+  expect(isGithubNativeAnthropicMode()).toBe(true)
+})
+
+test('isGithubNativeAnthropicMode: true for github:copilot:claude- compound format', async () => {
+  clearProviderEnv()
+  process.env.CLAUDE_CODE_USE_GITHUB = '1'
+  process.env.OPENAI_MODEL = 'github:copilot:claude-sonnet-4'
+  const { isGithubNativeAnthropicMode } = await importFreshProvidersModule()
+  expect(isGithubNativeAnthropicMode()).toBe(true)
+})
+
+test('isGithubNativeAnthropicMode: true when resolvedModel is a claude- model', async () => {
+  clearProviderEnv()
+  process.env.CLAUDE_CODE_USE_GITHUB = '1'
+  process.env.OPENAI_MODEL = 'github:copilot'
+  const { isGithubNativeAnthropicMode } = await importFreshProvidersModule()
+  expect(isGithubNativeAnthropicMode('claude-haiku-4-5')).toBe(true)
+})
+
+test('isGithubNativeAnthropicMode: false for generic github:copilot alias', async () => {
+  clearProviderEnv()
+  process.env.CLAUDE_CODE_USE_GITHUB = '1'
+  process.env.OPENAI_MODEL = 'github:copilot'
+  const { isGithubNativeAnthropicMode } = await importFreshProvidersModule()
+  expect(isGithubNativeAnthropicMode()).toBe(false)
+})
+
+test('isGithubNativeAnthropicMode: false for non-Claude model', async () => {
+  clearProviderEnv()
+  process.env.CLAUDE_CODE_USE_GITHUB = '1'
+  process.env.OPENAI_MODEL = 'gpt-4o'
+  const { isGithubNativeAnthropicMode } = await importFreshProvidersModule()
+  expect(isGithubNativeAnthropicMode()).toBe(false)
+})
+
+test('isGithubNativeAnthropicMode: false for github:copilot:gpt- model', async () => {
+  clearProviderEnv()
+  process.env.CLAUDE_CODE_USE_GITHUB = '1'
+  process.env.OPENAI_MODEL = 'github:copilot:gpt-4o'
+  const { isGithubNativeAnthropicMode } = await importFreshProvidersModule()
+  expect(isGithubNativeAnthropicMode()).toBe(false)
+})

src/utils/model/providers.ts

@@ -45,6 +45,24 @@ export function getAPIProvider(): APIProvider {
 export function usesAnthropicAccountFlow(): boolean {
   return getAPIProvider() === 'firstParty'
 }
+
+/**
+ * Returns true when the GitHub provider should use Anthropic's native API
+ * format instead of the OpenAI-compatible shim.
+ *
+ * Enabled when CLAUDE_CODE_USE_GITHUB=1 and the model string contains "claude-"
+ * anywhere (handles bare names like "claude-sonnet-4" and compound formats like
+ * "github:copilot:claude-sonnet-4" or any future provider-prefixed variants).
+ *
+ * api.githubcopilot.com supports Anthropic native format for Claude models,
+ * enabling prompt caching via cache_control blocks which significantly reduces
+ * per-turn token costs by caching the system prompt and tool definitions.
+ */
+export function isGithubNativeAnthropicMode(resolvedModel?: string): boolean {
+  if (!isEnvTruthy(process.env.CLAUDE_CODE_USE_GITHUB)) return false
+  const model = resolvedModel?.trim() || process.env.OPENAI_MODEL?.trim() || ''
+  return model.toLowerCase().includes('claude-')
+}
 function isCodexModel(): boolean {
   return shouldUseCodexTransport(
     process.env.OPENAI_MODEL || '',

src/utils/permissions/filesystem.ts

@@ -64,6 +64,7 @@ export const DANGEROUS_FILES = [
   '.profile',
   '.ripgreprc',
   '.mcp.json',
+  '.openclaude.json',
   '.claude.json',
 ] as const
 

src/utils/plugins/marketplaceManager.ts

@@ -532,6 +532,7 @@ export async function gitPull(
 ): Promise<{ code: number; stderr: string }> {
   logForDebugging(`git pull: cwd=${cwd} ref=${ref ?? 'default'}`)
   const env = { ...process.env, ...GIT_NO_PROMPT_ENV }
+  const baseArgs = ['-c', 'core.hooksPath=/dev/null']
   const credentialArgs = options?.disableCredentialHelper
     ? ['-c', 'credential.helper=']
     : []
@@ -539,7 +540,7 @@ export async function gitPull(
   if (ref) {
     const fetchResult = await execFileNoThrowWithCwd(
       gitExe(),
-      [...credentialArgs, 'fetch', 'origin', ref],
+      [...baseArgs, ...credentialArgs, 'fetch', 'origin', ref],
       { cwd, timeout: getPluginGitTimeoutMs(), stdin: 'ignore', env },
     )
 
@@ -549,7 +550,7 @@ export async function gitPull(
 
     const checkoutResult = await execFileNoThrowWithCwd(
       gitExe(),
-      [...credentialArgs, 'checkout', ref],
+      [...baseArgs, ...credentialArgs, 'checkout', ref],
       { cwd, timeout: getPluginGitTimeoutMs(), stdin: 'ignore', env },
     )
 
@@ -559,7 +560,7 @@ export async function gitPull(
 
     const pullResult = await execFileNoThrowWithCwd(
       gitExe(),
-      [...credentialArgs, 'pull', 'origin', ref],
+      [...baseArgs, ...credentialArgs, 'pull', 'origin', ref],
       { cwd, timeout: getPluginGitTimeoutMs(), stdin: 'ignore', env },
     )
     if (pullResult.code !== 0) {
@@ -571,7 +572,7 @@ export async function gitPull(
 
   const result = await execFileNoThrowWithCwd(
     gitExe(),
-    [...credentialArgs, 'pull', 'origin', 'HEAD'],
+    [...baseArgs, ...credentialArgs, 'pull', 'origin', 'HEAD'],
     { cwd, timeout: getPluginGitTimeoutMs(), stdin: 'ignore', env },
   )
   if (result.code !== 0) {
@@ -625,6 +626,8 @@ async function gitSubmoduleUpdate(
     [
       '-c',
       'core.sshCommand=ssh -o BatchMode=yes -o StrictHostKeyChecking=yes',
+      '-c',
+      'core.hooksPath=/dev/null',
       ...credentialArgs,
       'submodule',
       'update',
@@ -810,6 +813,8 @@ export async function gitClone(
   const args = [
     '-c',
     'core.sshCommand=ssh -o BatchMode=yes -o StrictHostKeyChecking=yes',
+    '-c',
+    'core.hooksPath=/dev/null',
     'clone',
     '--depth',
     '1',

src/utils/providerAutoDetect.test.ts

@@ -0,0 +1,299 @@
+import { describe, expect, test } from 'bun:test'
+
+import {
+  detectBestProvider,
+  detectLocalService,
+  detectProviderFromEnv,
+} from './providerAutoDetect.ts'
+
+// Hermetic env scan: always report "no Codex auth on disk" so tests don't
+// depend on the dev machine's ~/.codex/auth.json state.
+function scan(env: Record<string, string | undefined>) {
+  return detectProviderFromEnv({ env, hasCodexAuth: () => false })
+}
+
+describe('detectProviderFromEnv — priority order', () => {
+  test('ANTHROPIC_API_KEY wins over all others', () => {
+    expect(
+      scan({
+        ANTHROPIC_API_KEY: 'sk-ant-x',
+        OPENAI_API_KEY: 'sk-x',
+        GEMINI_API_KEY: 'gem-x',
+      }),
+    ).toEqual({ kind: 'anthropic', source: 'ANTHROPIC_API_KEY set' })
+  })
+
+  test('CODEX_API_KEY beats OpenAI/Gemini/etc', () => {
+    expect(
+      scan({
+        CODEX_API_KEY: 'codex-x',
+        OPENAI_API_KEY: 'sk-x',
+      }),
+    ).toEqual({ kind: 'codex', source: 'CODEX_API_KEY set' })
+  })
+
+  test('CHATGPT_ACCOUNT_ID alone is enough for Codex', () => {
+    expect(
+      scan({
+        CHATGPT_ACCOUNT_ID: 'acct-123',
+      }),
+    ).toEqual({ kind: 'codex', source: 'CHATGPT_ACCOUNT_ID set' })
+  })
+
+  test('Codex auth file on disk is detected without any env', () => {
+    expect(
+      detectProviderFromEnv({ env: {}, hasCodexAuth: () => true }),
+    ).toEqual({ kind: 'codex', source: '~/.codex/auth.json present' })
+  })
+
+  test('GITHUB_TOKEN wins over OpenAI', () => {
+    expect(
+      scan({
+        GITHUB_TOKEN: 'ghp-x',
+        OPENAI_API_KEY: 'sk-x',
+      }),
+    ).toEqual({ kind: 'github', source: 'GITHUB_TOKEN set (GitHub Copilot)' })
+  })
+
+  test('GH_TOKEN is equivalent to GITHUB_TOKEN', () => {
+    expect(
+      scan({
+        GH_TOKEN: 'ghp-x',
+      }),
+    ).toEqual({ kind: 'github', source: 'GH_TOKEN set (GitHub Copilot)' })
+  })
+
+  test('OPENAI_API_KEYS (plural) detected', () => {
+    expect(
+      scan({
+        OPENAI_API_KEYS: 'sk-a,sk-b',
+      }),
+    ).toEqual({ kind: 'openai', source: 'OPENAI_API_KEYS set' })
+  })
+
+  test('OPENAI_API_KEY reports baseUrl when set', () => {
+    expect(
+      scan({
+        OPENAI_API_KEY: 'sk-x',
+        OPENAI_BASE_URL: 'https://openrouter.ai/api/v1',
+      }),
+    ).toEqual({
+      kind: 'openai',
+      source: 'OPENAI_API_KEY set',
+      baseUrl: 'https://openrouter.ai/api/v1',
+    })
+  })
+
+  test('GEMINI_API_KEY detected', () => {
+    expect(scan({ GEMINI_API_KEY: 'gem-x' })).toEqual({
+      kind: 'gemini',
+      source: 'GEMINI_API_KEY set',
+    })
+  })
+
+  test('GOOGLE_API_KEY also detects Gemini', () => {
+    expect(scan({ GOOGLE_API_KEY: 'gk-x' })).toEqual({
+      kind: 'gemini',
+      source: 'GOOGLE_API_KEY set',
+    })
+  })
+
+  test('MISTRAL_API_KEY detected', () => {
+    expect(scan({ MISTRAL_API_KEY: 'mis-x' })).toEqual({
+      kind: 'mistral',
+      source: 'MISTRAL_API_KEY set',
+    })
+  })
+
+  test('MINIMAX_API_KEY detected', () => {
+    expect(scan({ MINIMAX_API_KEY: 'mm-x' })).toEqual({
+      kind: 'minimax',
+      source: 'MINIMAX_API_KEY set',
+    })
+  })
+
+  test('empty-string values are ignored', () => {
+    expect(
+      scan({
+        ANTHROPIC_API_KEY: '',
+        OPENAI_API_KEY: '   ',
+        GEMINI_API_KEY: 'gem-x',
+      }),
+    ).toEqual({ kind: 'gemini', source: 'GEMINI_API_KEY set' })
+  })
+
+  test('no credentials → null', () => {
+    expect(scan({})).toBeNull()
+  })
+})
+
+describe('detectLocalService', () => {
+  test('returns Ollama when its /api/tags responds ok', async () => {
+    const fetchImpl = (async (input: URL | RequestInfo) => {
+      const url = typeof input === 'string' ? input : (input as URL).toString()
+      if (url.includes(':11434')) {
+        return new Response('{"models":[]}', { status: 200 })
+      }
+      return new Response('', { status: 404 })
+    }) as typeof fetch
+
+    const result = await detectLocalService({
+      env: {},
+      fetchImpl,
+      timeoutMs: 200,
+    })
+    expect(result?.kind).toBe('ollama')
+    expect(result?.baseUrl).toBe('http://localhost:11434')
+  })
+
+  test('Ollama wins over LM Studio even when both are reachable', async () => {
+    const fetchImpl = (async () => new Response('{}', { status: 200 })) as typeof fetch
+    const result = await detectLocalService({
+      env: {},
+      fetchImpl,
+      timeoutMs: 200,
+    })
+    expect(result?.kind).toBe('ollama')
+  })
+
+  test('falls back to LM Studio when Ollama is unreachable', async () => {
+    const fetchImpl = (async (input: URL | RequestInfo) => {
+      const url = typeof input === 'string' ? input : (input as URL).toString()
+      if (url.includes(':1234')) {
+        return new Response('{"data":[]}', { status: 200 })
+      }
+      return new Response('', { status: 404 })
+    }) as typeof fetch
+
+    const result = await detectLocalService({
+      env: {},
+      fetchImpl,
+      timeoutMs: 200,
+    })
+    expect(result?.kind).toBe('lm-studio')
+    expect(result?.baseUrl).toBe('http://localhost:1234')
+  })
+
+  test('returns null when no local services respond', async () => {
+    const fetchImpl = (async () =>
+      new Response('', { status: 500 })) as typeof fetch
+    const result = await detectLocalService({
+      env: {},
+      fetchImpl,
+      timeoutMs: 200,
+    })
+    expect(result).toBeNull()
+  })
+
+  test('honors OLLAMA_BASE_URL override', async () => {
+    const probedUrls: string[] = []
+    const fetchImpl = (async (input: URL | RequestInfo) => {
+      const url = typeof input === 'string' ? input : (input as URL).toString()
+      probedUrls.push(url)
+      return new Response('{"models":[]}', { status: 200 })
+    }) as typeof fetch
+
+    const result = await detectLocalService({
+      env: { OLLAMA_BASE_URL: 'http://10.0.0.5:11434' },
+      fetchImpl,
+      timeoutMs: 200,
+    })
+    expect(result?.baseUrl).toBe('http://10.0.0.5:11434')
+    expect(probedUrls).toContain('http://10.0.0.5:11434/api/tags')
+  })
+
+  test('probe timeout does not throw — returns null', async () => {
+    const fetchImpl = (async (_input: URL | RequestInfo, init?: RequestInit) => {
+      // Respect the caller's abort signal so the race with timeoutMs is fair.
+      return new Promise<Response>((_resolve, reject) => {
+        const onAbort = () => reject(new Error('aborted'))
+        init?.signal?.addEventListener('abort', onAbort)
+        setTimeout(() => {
+          init?.signal?.removeEventListener('abort', onAbort)
+          _resolve(new Response('ok'))
+        }, 500)
+      })
+    }) as typeof fetch
+
+    const result = await detectLocalService({
+      env: {},
+      fetchImpl,
+      timeoutMs: 50,
+    })
+    expect(result).toBeNull()
+  })
+
+  test('network errors do not throw', async () => {
+    const fetchImpl = (async () => {
+      throw new Error('ECONNREFUSED')
+    }) as typeof fetch
+
+    const result = await detectLocalService({
+      env: {},
+      fetchImpl,
+      timeoutMs: 200,
+    })
+    expect(result).toBeNull()
+  })
+})
+
+describe('detectBestProvider — orchestrator', () => {
+  test('env match short-circuits the local probe', async () => {
+    let probeCalled = false
+    const fetchImpl = (async () => {
+      probeCalled = true
+      return new Response('{}', { status: 200 })
+    }) as typeof fetch
+
+    const result = await detectBestProvider({
+      env: { ANTHROPIC_API_KEY: 'sk-ant' },
+      fetchImpl,
+      timeoutMs: 200,
+      hasCodexAuth: () => false,
+    })
+    expect(result?.kind).toBe('anthropic')
+    expect(probeCalled).toBe(false)
+  })
+
+  test('env miss falls through to local-service probe', async () => {
+    const fetchImpl = (async () => new Response('{}', { status: 200 })) as typeof fetch
+    const result = await detectBestProvider({
+      env: {},
+      fetchImpl,
+      timeoutMs: 200,
+      hasCodexAuth: () => false,
+    })
+    expect(result?.kind).toBe('ollama')
+  })
+
+  test('skipLocal prevents network probes', async () => {
+    let probeCalled = false
+    const fetchImpl = (async () => {
+      probeCalled = true
+      return new Response('{}', { status: 200 })
+    }) as typeof fetch
+
+    const result = await detectBestProvider({
+      env: {},
+      fetchImpl,
+      skipLocal: true,
+      hasCodexAuth: () => false,
+    })
+    expect(result).toBeNull()
+    expect(probeCalled).toBe(false)
+  })
+
+  test('completely empty environment returns null', async () => {
+    const fetchImpl = (async () => {
+      throw new Error('nothing reachable')
+    }) as typeof fetch
+
+    const result = await detectBestProvider({
+      env: {},
+      fetchImpl,
+      timeoutMs: 100,
+      hasCodexAuth: () => false,
+    })
+    expect(result).toBeNull()
+  })
+})

src/utils/providerAutoDetect.ts

@@ -0,0 +1,283 @@
+/**
+ * Zero-config provider autodetection.
+ *
+ * Scans the environment (API keys, OAuth tokens, stored credentials) and local
+ * network (Ollama, LM Studio) to pick the best provider for first-run users
+ * who have not explicitly configured one. Returns a structured detection
+ * result that callers can consume to build a launch-ready profile env, or
+ * null when nothing is detected — in which case the existing onboarding /
+ * picker flow should take over.
+ *
+ * Detection priority (first match wins):
+ *   1. ANTHROPIC_API_KEY → first-party Claude (most capable default)
+ *   2. Codex: CODEX_API_KEY, CHATGPT_ACCOUNT_ID, or valid ~/.codex/auth.json
+ *   3. GitHub Copilot: GITHUB_TOKEN or GH_TOKEN
+ *   4. OPENAI_API_KEY / OPENAI_API_KEYS
+ *   5. GEMINI_API_KEY or GOOGLE_API_KEY
+ *   6. MISTRAL_API_KEY
+ *   7. MINIMAX_API_KEY
+ *   8. Local Ollama reachable (default localhost:11434)
+ *   9. Local LM Studio reachable (default localhost:1234)
+ *
+ * Local-service probes are parallelized and cheap (short timeout, no
+ * request body). Env scans are synchronous and run first so we don't make
+ * network calls when a credential is already present.
+ *
+ * This module intentionally does NOT decide whether to apply the detection;
+ * callers should gate on hasExplicitProviderSelection() (providerProfile.ts)
+ * and the presence of a persisted profile file.
+ */
+
+import { existsSync } from 'fs'
+import { homedir } from 'os'
+import { join } from 'path'
+
+export type DetectedProviderKind =
+  | 'anthropic'
+  | 'codex'
+  | 'github'
+  | 'openai'
+  | 'gemini'
+  | 'mistral'
+  | 'minimax'
+  | 'ollama'
+  | 'lm-studio'
+
+export type DetectedProvider = {
+  kind: DetectedProviderKind
+  /** One-line human-readable reason, e.g. "ANTHROPIC_API_KEY set". */
+  source: string
+  /** Present when the detection already resolved a usable base URL. */
+  baseUrl?: string
+  /** Present when detection also narrowed down a specific model. */
+  model?: string
+}
+
+type EnvLike = NodeJS.ProcessEnv | Record<string, string | undefined>
+
+function envHasNonEmpty(env: EnvLike, key: string): boolean {
+  const value = env[key]
+  return typeof value === 'string' && value.trim().length > 0
+}
+
+function firstSet(env: EnvLike, keys: readonly string[]): string | undefined {
+  for (const key of keys) {
+    if (envHasNonEmpty(env, key)) return key
+  }
+  return undefined
+}
+
+function defaultHasCodexAuthFile(): boolean {
+  const paths = [
+    process.env.CODEX_AUTH_PATH,
+    join(homedir(), '.codex', 'auth.json'),
+  ]
+  return paths.some(p => p && existsSync(p))
+}
+
+export type DetectProviderFromEnvOptions = {
+  env?: EnvLike
+  /**
+   * Override Codex auth-file detection. Primarily for tests — the default
+   * implementation checks ~/.codex/auth.json and CODEX_AUTH_PATH on disk.
+   */
+  hasCodexAuth?: () => boolean
+}
+
+/**
+ * Synchronous env-only scan. Returns the highest-priority env-provided
+ * provider, or null if nothing is present. Intentionally does not touch
+ * the network — fast path for the common case where a user has exported
+ * one of the standard API-key env vars.
+ */
+function isOptionsObject(
+  value: EnvLike | DetectProviderFromEnvOptions | undefined,
+): value is DetectProviderFromEnvOptions {
+  if (!value || typeof value !== 'object') return false
+  if ('hasCodexAuth' in value && typeof value.hasCodexAuth === 'function') {
+    return true
+  }
+  if ('env' in value && typeof (value as { env?: unknown }).env === 'object') {
+    return true
+  }
+  return false
+}
+
+export function detectProviderFromEnv(
+  envOrOptions: EnvLike | DetectProviderFromEnvOptions = process.env,
+): DetectedProvider | null {
+  const options: DetectProviderFromEnvOptions = isOptionsObject(envOrOptions)
+    ? envOrOptions
+    : { env: envOrOptions as EnvLike }
+  const env = options.env ?? process.env
+  const hasCodexAuth = options.hasCodexAuth ?? defaultHasCodexAuthFile
+  if (envHasNonEmpty(env, 'ANTHROPIC_API_KEY')) {
+    return { kind: 'anthropic', source: 'ANTHROPIC_API_KEY set' }
+  }
+
+  if (
+    envHasNonEmpty(env, 'CODEX_API_KEY') ||
+    envHasNonEmpty(env, 'CHATGPT_ACCOUNT_ID') ||
+    envHasNonEmpty(env, 'CODEX_ACCOUNT_ID') ||
+    hasCodexAuth()
+  ) {
+    const sourceEnv =
+      firstSet(env, ['CODEX_API_KEY', 'CHATGPT_ACCOUNT_ID', 'CODEX_ACCOUNT_ID'])
+    return {
+      kind: 'codex',
+      source: sourceEnv ? `${sourceEnv} set` : '~/.codex/auth.json present',
+    }
+  }
+
+  const githubKey = firstSet(env, ['GITHUB_TOKEN', 'GH_TOKEN'])
+  if (githubKey) {
+    return {
+      kind: 'github',
+      source: `${githubKey} set (GitHub Copilot)`,
+    }
+  }
+
+  const openaiKey = firstSet(env, ['OPENAI_API_KEYS', 'OPENAI_API_KEY'])
+  if (openaiKey) {
+    return {
+      kind: 'openai',
+      source: `${openaiKey} set`,
+      baseUrl: env.OPENAI_BASE_URL ?? env.OPENAI_API_BASE,
+    }
+  }
+
+  const geminiKey = firstSet(env, ['GEMINI_API_KEY', 'GOOGLE_API_KEY'])
+  if (geminiKey) {
+    return { kind: 'gemini', source: `${geminiKey} set` }
+  }
+
+  if (envHasNonEmpty(env, 'MISTRAL_API_KEY')) {
+    return { kind: 'mistral', source: 'MISTRAL_API_KEY set' }
+  }
+
+  if (envHasNonEmpty(env, 'MINIMAX_API_KEY')) {
+    return { kind: 'minimax', source: 'MINIMAX_API_KEY set' }
+  }
+
+  return null
+}
+
+type LocalProbe = {
+  kind: DetectedProviderKind
+  url: string
+  timeoutMs: number
+  source: string
+  baseUrl: string
+}
+
+const DEFAULT_LOCAL_PROBE_TIMEOUT_MS = 1200
+
+async function probeReachable(
+  url: string,
+  timeoutMs: number,
+  fetchImpl: typeof fetch,
+): Promise<boolean> {
+  const controller = new AbortController()
+  const timer = setTimeout(() => controller.abort(), timeoutMs)
+  try {
+    const response = await fetchImpl(url, {
+      method: 'GET',
+      signal: controller.signal,
+    })
+    return response.ok
+  } catch {
+    return false
+  } finally {
+    clearTimeout(timer)
+  }
+}
+
+/**
+ * Returns the highest-priority local service reachable from the host.
+ * Runs probes in parallel and picks by priority rather than first-response,
+ * so slow-but-preferred services still win over fast-but-lower-priority ones.
+ */
+export async function detectLocalService(options?: {
+  env?: EnvLike
+  fetchImpl?: typeof fetch
+  timeoutMs?: number
+}): Promise<DetectedProvider | null> {
+  const env = options?.env ?? process.env
+  const fetchImpl = options?.fetchImpl ?? globalThis.fetch
+  const timeoutMs = options?.timeoutMs ?? DEFAULT_LOCAL_PROBE_TIMEOUT_MS
+
+  const ollamaBase = (env.OLLAMA_BASE_URL ?? 'http://localhost:11434').replace(
+    /\/+$/,
+    '',
+  )
+  const lmStudioBase = (env.LM_STUDIO_BASE_URL ?? 'http://localhost:1234').replace(
+    /\/+$/,
+    '',
+  )
+
+  const probes: LocalProbe[] = [
+    {
+      kind: 'ollama',
+      url: `${ollamaBase}/api/tags`,
+      timeoutMs,
+      source: `Ollama reachable at ${ollamaBase}`,
+      baseUrl: ollamaBase,
+    },
+    {
+      kind: 'lm-studio',
+      url: `${lmStudioBase}/v1/models`,
+      timeoutMs,
+      source: `LM Studio reachable at ${lmStudioBase}`,
+      baseUrl: lmStudioBase,
+    },
+  ]
+
+  const results = await Promise.all(
+    probes.map(async probe => ({
+      probe,
+      reachable: await probeReachable(probe.url, probe.timeoutMs, fetchImpl),
+    })),
+  )
+
+  for (const { probe, reachable } of results) {
+    if (reachable) {
+      return {
+        kind: probe.kind,
+        source: probe.source,
+        baseUrl: probe.baseUrl,
+      }
+    }
+  }
+
+  return null
+}
+
+/**
+ * Orchestrator: env scan first (sync, free), then local-service probes
+ * (async, ~1-2s worst case) only if nothing was found in env.
+ */
+export async function detectBestProvider(options?: {
+  env?: EnvLike
+  fetchImpl?: typeof fetch
+  timeoutMs?: number
+  /** Skip local-service probes — useful for tests or offline smoke checks. */
+  skipLocal?: boolean
+  /** Override for Codex auth-file detection. See detectProviderFromEnv. */
+  hasCodexAuth?: () => boolean
+}): Promise<DetectedProvider | null> {
+  const env = options?.env ?? process.env
+
+  const fromEnv = detectProviderFromEnv({
+    env,
+    hasCodexAuth: options?.hasCodexAuth,
+  })
+  if (fromEnv) return fromEnv
+
+  if (options?.skipLocal) return null
+
+  return detectLocalService({
+    env,
+    fetchImpl: options?.fetchImpl,
+    timeoutMs: options?.timeoutMs,
+  })
+}

src/utils/providerDiscovery.test.ts

@@ -1,9 +1,9 @@
 import { afterEach, expect, mock, test } from 'bun:test'
 
-import {
-  getLocalOpenAICompatibleProviderLabel,
-  listOpenAICompatibleModels,
-} from './providerDiscovery.js'
+async function loadProviderDiscoveryModule() {
+  // @ts-expect-error cache-busting query string for Bun module mocks
+  return import(`./providerDiscovery.js?ts=${Date.now()}-${Math.random()}`)
+}
 
 const originalFetch = globalThis.fetch
 const originalEnv = {
@@ -16,6 +16,8 @@ afterEach(() => {
 })
 
 test('lists models from a local openai-compatible /models endpoint', async () => {
+  const { listOpenAICompatibleModels } = await loadProviderDiscoveryModule()
+
   globalThis.fetch = mock((input, init) => {
     const url = typeof input === 'string' ? input : input.url
     expect(url).toBe('http://localhost:1234/v1/models')
@@ -47,6 +49,8 @@ test('lists models from a local openai-compatible /models endpoint', async () =>
 })
 
 test('returns null when a local openai-compatible /models request fails', async () => {
+  const { listOpenAICompatibleModels } = await loadProviderDiscoveryModule()
+
   globalThis.fetch = mock(() =>
     Promise.resolve(new Response('not available', { status: 503 })),
   ) as typeof globalThis.fetch
@@ -56,13 +60,19 @@ test('returns null when a local openai-compatible /models request fails', async
   ).resolves.toBeNull()
 })
 
-test('detects LM Studio from the default localhost port', () => {
+test('detects LM Studio from the default localhost port', async () => {
+  const { getLocalOpenAICompatibleProviderLabel } =
+    await loadProviderDiscoveryModule()
+
   expect(getLocalOpenAICompatibleProviderLabel('http://localhost:1234/v1')).toBe(
     'LM Studio',
   )
 })
 
-test('detects common local openai-compatible providers by hostname', () => {
+test('detects common local openai-compatible providers by hostname', async () => {
+  const { getLocalOpenAICompatibleProviderLabel } =
+    await loadProviderDiscoveryModule()
+
   expect(
     getLocalOpenAICompatibleProviderLabel('http://localai.local:8080/v1'),
   ).toBe('LocalAI')
@@ -71,8 +81,283 @@ test('detects common local openai-compatible providers by hostname', () => {
   ).toBe('vLLM')
 })
 
-test('falls back to a generic local openai-compatible label', () => {
+test('detects Moonshot (Kimi) from api.moonshot.ai hostname', async () => {
+  const { getLocalOpenAICompatibleProviderLabel } =
+    await loadProviderDiscoveryModule()
+
+  expect(
+    getLocalOpenAICompatibleProviderLabel('https://api.moonshot.ai/v1'),
+  ).toBe('Moonshot (Kimi)')
+})
+
+test('falls back to a generic local openai-compatible label', async () => {
+  const { getLocalOpenAICompatibleProviderLabel } =
+    await loadProviderDiscoveryModule()
+
   expect(
     getLocalOpenAICompatibleProviderLabel('http://127.0.0.1:8080/v1'),
   ).toBe('Local OpenAI-compatible')
+})
+
+test('ollama generation readiness reports unreachable when tags endpoint is down', async () => {
+  const { probeOllamaGenerationReadiness } = await loadProviderDiscoveryModule()
+
+  const calledUrls: string[] = []
+  globalThis.fetch = mock(input => {
+    const url = typeof input === 'string' ? input : input.url
+    calledUrls.push(url)
+    return Promise.resolve(new Response('not available', { status: 503 }))
+  }) as typeof globalThis.fetch
+
+  await expect(
+    probeOllamaGenerationReadiness({
+      baseUrl: 'http://localhost:11434',
+    }),
+  ).resolves.toMatchObject({
+    state: 'unreachable',
+    models: [],
+  })
+
+  expect(calledUrls).toEqual([
+    'http://localhost:11434/api/tags',
+  ])
+})
+
+test('ollama generation readiness reports no models when server is reachable', async () => {
+  const { probeOllamaGenerationReadiness } = await loadProviderDiscoveryModule()
+
+  const calledUrls: string[] = []
+  globalThis.fetch = mock(input => {
+    const url = typeof input === 'string' ? input : input.url
+    calledUrls.push(url)
+    return Promise.resolve(
+      new Response(JSON.stringify({ models: [] }), {
+        status: 200,
+        headers: { 'Content-Type': 'application/json' },
+      }),
+    )
+  }) as typeof globalThis.fetch
+
+  await expect(
+    probeOllamaGenerationReadiness({
+      baseUrl: 'http://localhost:11434',
+    }),
+  ).resolves.toMatchObject({
+    state: 'no_models',
+    models: [],
+  })
+
+  expect(calledUrls).toEqual([
+    'http://localhost:11434/api/tags',
+  ])
+})
+
+test('ollama generation readiness reports generation_failed when requested model is missing', async () => {
+  const { probeOllamaGenerationReadiness } = await loadProviderDiscoveryModule()
+
+  const calledUrls: string[] = []
+  globalThis.fetch = mock(input => {
+    const url = typeof input === 'string' ? input : input.url
+    calledUrls.push(url)
+    return Promise.resolve(
+      new Response(
+        JSON.stringify({
+          models: [{ name: 'llama3.1:8b', size: 1024 }],
+        }),
+        {
+          status: 200,
+          headers: { 'Content-Type': 'application/json' },
+        },
+      ),
+    )
+  }) as typeof globalThis.fetch
+
+  await expect(
+    probeOllamaGenerationReadiness({
+      baseUrl: 'http://localhost:11434',
+      model: 'qwen2.5-coder:7b',
+    }),
+  ).resolves.toMatchObject({
+    state: 'generation_failed',
+    probeModel: 'qwen2.5-coder:7b',
+    detail: 'requested model not installed: qwen2.5-coder:7b',
+  })
+
+  expect(calledUrls).toEqual(['http://localhost:11434/api/tags'])
+})
+
+test('ollama generation readiness reports generation failures when chat probe fails', async () => {
+  const { probeOllamaGenerationReadiness } = await loadProviderDiscoveryModule()
+
+  globalThis.fetch = mock(input => {
+    const url = typeof input === 'string' ? input : input.url
+    if (url.endsWith('/api/tags')) {
+      return Promise.resolve(
+        new Response(
+          JSON.stringify({
+            models: [{ name: 'qwen2.5-coder:7b', size: 42 }],
+          }),
+          {
+            status: 200,
+            headers: { 'Content-Type': 'application/json' },
+          },
+        ),
+      )
+    }
+
+    return Promise.resolve(new Response('model not found', { status: 404 }))
+  }) as typeof globalThis.fetch
+
+  await expect(
+    probeOllamaGenerationReadiness({
+      baseUrl: 'http://localhost:11434',
+      model: 'qwen2.5-coder:7b',
+    }),
+  ).resolves.toMatchObject({
+    state: 'generation_failed',
+    probeModel: 'qwen2.5-coder:7b',
+  })
+})
+
+test('ollama generation readiness reports generation_failed when chat probe returns invalid JSON', async () => {
+  const { probeOllamaGenerationReadiness } = await loadProviderDiscoveryModule()
+
+  globalThis.fetch = mock(input => {
+    const url = typeof input === 'string' ? input : input.url
+    if (url.endsWith('/api/tags')) {
+      return Promise.resolve(
+        new Response(
+          JSON.stringify({
+            models: [{ name: 'llama3.1:8b', size: 1024 }],
+          }),
+          {
+            status: 200,
+            headers: { 'Content-Type': 'application/json' },
+          },
+        ),
+      )
+    }
+
+    return Promise.resolve(
+      new Response('<html>proxy error</html>', {
+        status: 200,
+        headers: { 'Content-Type': 'text/html' },
+      }),
+    )
+  }) as typeof globalThis.fetch
+
+  await expect(
+    probeOllamaGenerationReadiness({
+      baseUrl: 'http://localhost:11434',
+    }),
+  ).resolves.toMatchObject({
+    state: 'generation_failed',
+    probeModel: 'llama3.1:8b',
+    detail: 'invalid JSON response',
+  })
+})
+
+test('ollama generation readiness reports ready when chat probe succeeds', async () => {
+  const { probeOllamaGenerationReadiness } = await loadProviderDiscoveryModule()
+
+  globalThis.fetch = mock(input => {
+    const url = typeof input === 'string' ? input : input.url
+    if (url.endsWith('/api/tags')) {
+      return Promise.resolve(
+        new Response(
+          JSON.stringify({
+            models: [{ name: 'llama3.1:8b', size: 1024 }],
+          }),
+          {
+            status: 200,
+            headers: { 'Content-Type': 'application/json' },
+          },
+        ),
+      )
+    }
+
+    return Promise.resolve(
+      new Response(
+        JSON.stringify({
+          message: { role: 'assistant', content: 'OK' },
+          done: true,
+        }),
+        {
+          status: 200,
+          headers: { 'Content-Type': 'application/json' },
+        },
+      ),
+    )
+  }) as typeof globalThis.fetch
+
+  await expect(
+    probeOllamaGenerationReadiness({
+      baseUrl: 'http://localhost:11434',
+    }),
+  ).resolves.toMatchObject({
+    state: 'ready',
+    probeModel: 'llama3.1:8b',
+  })
+})
+
+test('atomic chat readiness reports unreachable when /v1/models is down', async () => {
+  const { probeAtomicChatReadiness } = await loadProviderDiscoveryModule()
+
+  const calledUrls: string[] = []
+  globalThis.fetch = mock(input => {
+    const url = typeof input === 'string' ? input : input.url
+    calledUrls.push(url)
+    return Promise.resolve(new Response('unavailable', { status: 503 }))
+  }) as typeof globalThis.fetch
+
+  await expect(
+    probeAtomicChatReadiness({ baseUrl: 'http://127.0.0.1:1337' }),
+  ).resolves.toEqual({ state: 'unreachable' })
+
+  expect(calledUrls[0]).toBe('http://127.0.0.1:1337/v1/models')
+})
+
+test('atomic chat readiness reports no_models when server is reachable but empty', async () => {
+  const { probeAtomicChatReadiness } = await loadProviderDiscoveryModule()
+
+  globalThis.fetch = mock(() =>
+    Promise.resolve(
+      new Response(JSON.stringify({ data: [] }), {
+        status: 200,
+        headers: { 'Content-Type': 'application/json' },
+      }),
+    ),
+  ) as typeof globalThis.fetch
+
+  await expect(
+    probeAtomicChatReadiness({ baseUrl: 'http://127.0.0.1:1337' }),
+  ).resolves.toEqual({ state: 'no_models' })
+})
+
+test('atomic chat readiness returns loaded model ids when ready', async () => {
+  const { probeAtomicChatReadiness } = await loadProviderDiscoveryModule()
+
+  globalThis.fetch = mock(() =>
+    Promise.resolve(
+      new Response(
+        JSON.stringify({
+          data: [
+            { id: 'Qwen3_5-4B_Q4_K_M' },
+            { id: 'llama-3.1-8b-instruct' },
+          ],
+        }),
+        {
+          status: 200,
+          headers: { 'Content-Type': 'application/json' },
+        },
+      ),
+    ),
+  ) as typeof globalThis.fetch
+
+  await expect(
+    probeAtomicChatReadiness({ baseUrl: 'http://127.0.0.1:1337' }),
+  ).resolves.toEqual({
+    state: 'ready',
+    models: ['Qwen3_5-4B_Q4_K_M', 'llama-3.1-8b-instruct'],
+  })
 })

src/utils/providerDiscovery.ts

@@ -4,6 +4,13 @@ import { DEFAULT_OPENAI_BASE_URL } from '../services/api/providerConfig.js'
 export const DEFAULT_OLLAMA_BASE_URL = 'http://localhost:11434'
 export const DEFAULT_ATOMIC_CHAT_BASE_URL = 'http://127.0.0.1:1337'
 
+export type OllamaGenerationReadiness = {
+  state: 'ready' | 'unreachable' | 'no_models' | 'generation_failed'
+  models: OllamaModelDescriptor[]
+  probeModel?: string
+  detail?: string
+}
+
 function withTimeoutSignal(timeoutMs: number): {
   signal: AbortSignal
   clear: () => void
@@ -20,6 +27,83 @@ function trimTrailingSlash(value: string): string {
   return value.replace(/\/+$/, '')
 }
 
+function compactDetail(value: string, maxLength = 180): string {
+  const compact = value.trim().replace(/\s+/g, ' ')
+  if (!compact) {
+    return ''
+  }
+
+  if (compact.length <= maxLength) {
+    return compact
+  }
+
+  return `${compact.slice(0, maxLength)}...`
+}
+
+type OllamaTagsPayload = {
+  models?: Array<{
+    name?: string
+    size?: number
+    details?: {
+      family?: string
+      families?: string[]
+      parameter_size?: string
+      quantization_level?: string
+    }
+  }>
+}
+
+function normalizeOllamaModels(
+  payload: OllamaTagsPayload,
+): OllamaModelDescriptor[] {
+  return (payload.models ?? [])
+    .filter(model => Boolean(model.name))
+    .map(model => ({
+      name: model.name!,
+      sizeBytes: typeof model.size === 'number' ? model.size : null,
+      family: model.details?.family ?? null,
+      families: model.details?.families ?? [],
+      parameterSize: model.details?.parameter_size ?? null,
+      quantizationLevel: model.details?.quantization_level ?? null,
+    }))
+}
+
+async function fetchOllamaModelsProbe(
+  baseUrl?: string,
+  timeoutMs = 5000,
+): Promise<{
+  reachable: boolean
+  models: OllamaModelDescriptor[]
+}> {
+  const { signal, clear } = withTimeoutSignal(timeoutMs)
+  try {
+    const response = await fetch(`${getOllamaApiBaseUrl(baseUrl)}/api/tags`, {
+      method: 'GET',
+      signal,
+    })
+
+    if (!response.ok) {
+      return {
+        reachable: false,
+        models: [],
+      }
+    }
+
+    const payload = (await response.json().catch(() => ({}))) as OllamaTagsPayload
+    return {
+      reachable: true,
+      models: normalizeOllamaModels(payload),
+    }
+  } catch {
+    return {
+      reachable: false,
+      models: [],
+    }
+  } finally {
+    clear()
+  }
+}
+
 export function getOllamaApiBaseUrl(baseUrl?: string): string {
   const parsed = new URL(
     baseUrl || process.env.OLLAMA_BASE_URL || DEFAULT_OLLAMA_BASE_URL,
@@ -113,6 +197,10 @@ export function getLocalOpenAICompatibleProviderLabel(baseUrl?: string): string
     if (host.includes('minimax') || haystack.includes('minimax')) {
       return 'MiniMax'
     }
+    // Moonshot AI (Kimi) direct API
+    if (host.includes('moonshot') || haystack.includes('moonshot') || haystack.includes('kimi')) {
+      return 'Moonshot (Kimi)'
+    }
   } catch {
     // Fall back to the generic label when the base URL is malformed.
   }
@@ -121,61 +209,15 @@ export function getLocalOpenAICompatibleProviderLabel(baseUrl?: string): string
 }
 
 export async function hasLocalOllama(baseUrl?: string): Promise<boolean> {
-  const { signal, clear } = withTimeoutSignal(1200)
-  try {
-    const response = await fetch(`${getOllamaApiBaseUrl(baseUrl)}/api/tags`, {
-      method: 'GET',
-      signal,
-    })
-    return response.ok
-  } catch {
-    return false
-  } finally {
-    clear()
-  }
+  const { reachable } = await fetchOllamaModelsProbe(baseUrl, 1200)
+  return reachable
 }
 
 export async function listOllamaModels(
   baseUrl?: string,
 ): Promise<OllamaModelDescriptor[]> {
-  const { signal, clear } = withTimeoutSignal(5000)
-  try {
-    const response = await fetch(`${getOllamaApiBaseUrl(baseUrl)}/api/tags`, {
-      method: 'GET',
-      signal,
-    })
-    if (!response.ok) {
-      return []
-    }
-
-    const data = (await response.json()) as {
-      models?: Array<{
-        name?: string
-        size?: number
-        details?: {
-          family?: string
-          families?: string[]
-          parameter_size?: string
-          quantization_level?: string
-        }
-      }>
-    }
-
-    return (data.models ?? [])
-      .filter(model => Boolean(model.name))
-      .map(model => ({
-        name: model.name!,
-        sizeBytes: typeof model.size === 'number' ? model.size : null,
-        family: model.details?.family ?? null,
-        families: model.details?.families ?? [],
-        parameterSize: model.details?.parameter_size ?? null,
-        quantizationLevel: model.details?.quantization_level ?? null,
-      }))
-  } catch {
-    return []
-  } finally {
-    clear()
-  }
+  const { models } = await fetchOllamaModelsProbe(baseUrl, 5000)
+  return models
 }
 
 export async function listOpenAICompatibleModels(options?: {
@@ -260,6 +302,24 @@ export async function listAtomicChatModels(
   }
 }
 
+export type AtomicChatReadiness =
+  | { state: 'unreachable' }
+  | { state: 'no_models' }
+  | { state: 'ready'; models: string[] }
+
+export async function probeAtomicChatReadiness(options?: {
+  baseUrl?: string
+}): Promise<AtomicChatReadiness> {
+  if (!(await hasLocalAtomicChat(options?.baseUrl))) {
+    return { state: 'unreachable' }
+  }
+  const models = await listAtomicChatModels(options?.baseUrl)
+  if (models.length === 0) {
+    return { state: 'no_models' }
+  }
+  return { state: 'ready', models }
+}
+
 export async function benchmarkOllamaModel(
   modelName: string,
   baseUrl?: string,
@@ -294,3 +354,106 @@ export async function benchmarkOllamaModel(
     clear()
   }
 }
+
+export async function probeOllamaGenerationReadiness(options?: {
+  baseUrl?: string
+  model?: string
+  timeoutMs?: number
+}): Promise<OllamaGenerationReadiness> {
+  const timeoutMs = options?.timeoutMs ?? 8000
+  const { reachable, models } = await fetchOllamaModelsProbe(
+    options?.baseUrl,
+    timeoutMs,
+  )
+  if (!reachable) {
+    return {
+      state: 'unreachable',
+      models: [],
+    }
+  }
+
+  if (models.length === 0) {
+    return {
+      state: 'no_models',
+      models: [],
+    }
+  }
+
+  const requestedModel = options?.model?.trim() || undefined
+  if (requestedModel && !models.some(model => model.name === requestedModel)) {
+    return {
+      state: 'generation_failed',
+      models,
+      probeModel: requestedModel,
+      detail: `requested model not installed: ${requestedModel}`,
+    }
+  }
+
+  const probeModel = requestedModel ?? models[0]!.name
+  const { signal, clear } = withTimeoutSignal(timeoutMs)
+
+  try {
+    const response = await fetch(`${getOllamaApiBaseUrl(options?.baseUrl)}/api/chat`, {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'application/json',
+      },
+      signal,
+      body: JSON.stringify({
+        model: probeModel,
+        stream: false,
+        messages: [{ role: 'user', content: 'Reply with OK.' }],
+        options: {
+          temperature: 0,
+          num_predict: 8,
+        },
+      }),
+    })
+
+    if (!response.ok) {
+      const responseBody = await response.text().catch(() => '')
+      const detailSuffix = compactDetail(responseBody)
+      return {
+        state: 'generation_failed',
+        models,
+        probeModel,
+        detail: detailSuffix
+          ? `status ${response.status}: ${detailSuffix}`
+          : `status ${response.status}`,
+      }
+    }
+
+    try {
+      await response.json()
+    } catch {
+      return {
+        state: 'generation_failed',
+        models,
+        probeModel,
+        detail: 'invalid JSON response',
+      }
+    }
+
+    return {
+      state: 'ready',
+      models,
+      probeModel,
+    }
+  } catch (error) {
+    const detail =
+      error instanceof Error
+        ? error.name === 'AbortError'
+          ? 'request timed out'
+          : error.message
+        : String(error)
+
+    return {
+      state: 'generation_failed',
+      models,
+      probeModel,
+      detail,
+    }
+  } finally {
+    clear()
+  }
+}

src/utils/providerProfile.test.ts

@@ -572,31 +572,64 @@ test('buildStartupEnvFromProfile leaves explicit provider selections untouched',
   assert.equal(env.OPENAI_API_KEY, undefined)
 })
 
-test('buildStartupEnvFromProfile lets saved startup profile override profile-managed env', async () => {
+test('buildStartupEnvFromProfile preserves plural-profile env when the legacy file is stale', async () => {
+  // Regression: a user saves a provider via /provider (plural system).
+  // addProviderProfile does NOT sync the legacy .openclaude-profile.json,
+  // so the legacy file retains whatever it had from an earlier setup (e.g.
+  // OpenAI defaults). At startup, applyActiveProviderProfileFromConfig()
+  // correctly applies the active plural profile (Moonshot) first, marking
+  // env with CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED=1. The legacy-file
+  // load must NOT overwrite that env — it previously did, surfacing as
+  // "banner shows the wrong provider / model".
   const processEnv = {
     CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED: '1',
-    CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED_ID: 'saved_ollama',
+    CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED_ID: 'saved_moonshot',
     CLAUDE_CODE_USE_OPENAI: '1',
-    OPENAI_BASE_URL: 'http://localhost:11434/v1',
-    OPENAI_MODEL: 'llama3.1:8b',
+    OPENAI_BASE_URL: 'https://api.moonshot.ai/v1',
+    OPENAI_MODEL: 'kimi-k2.6',
   }
 
   const env = await buildStartupEnvFromProfile({
+    // Stale legacy file — points at SambaNova, but user's active plural
+    // profile is Moonshot and was just applied.
     persisted: profile('openai', {
-      OPENAI_API_KEY: 'sk-persisted',
+      OPENAI_API_KEY: 'sk-stale',
       OPENAI_MODEL: 'Meta-Llama-3.1-70B-Instruct',
       OPENAI_BASE_URL: 'https://api.sambanova.ai/v1',
     }),
     processEnv,
   })
 
+  assert.equal(env, processEnv)
+  assert.equal(env.OPENAI_BASE_URL, 'https://api.moonshot.ai/v1')
+  assert.equal(env.OPENAI_MODEL, 'kimi-k2.6')
+  // Plural markers are retained — downstream code uses them to verify the
+  // env still belongs to the profile it was applied from.
+  assert.equal(env.CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED, '1')
+  assert.equal(env.CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED_ID, 'saved_moonshot')
+})
+
+test('buildStartupEnvFromProfile falls back to legacy file when plural system has not applied', async () => {
+  // Counter-example: first-run user with only the legacy file (no plural
+  // active profile yet). The legacy file is the correct source, so the
+  // load must proceed as before.
+  const processEnv = {
+    CLAUDE_CODE_USE_OPENAI: '1',
+  }
+
+  const env = await buildStartupEnvFromProfile({
+    persisted: profile('openai', {
+      OPENAI_API_KEY: 'sk-legacy',
+      OPENAI_MODEL: 'gpt-4o',
+      OPENAI_BASE_URL: 'https://api.openai.com/v1',
+    }),
+    processEnv,
+  })
+
   assert.notEqual(env, processEnv)
-  assert.equal(env.CLAUDE_CODE_USE_OPENAI, '1')
-  assert.equal(env.OPENAI_API_KEY, 'sk-persisted')
-  assert.equal(env.OPENAI_MODEL, 'Meta-Llama-3.1-70B-Instruct')
-  assert.equal(env.OPENAI_BASE_URL, 'https://api.sambanova.ai/v1')
-  assert.equal(env.CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED, undefined)
-  assert.equal(env.CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED_ID, undefined)
+  assert.equal(env.OPENAI_API_KEY, 'sk-legacy')
+  assert.equal(env.OPENAI_BASE_URL, 'https://api.openai.com/v1')
+  assert.equal(env.OPENAI_MODEL, 'gpt-4o')
 })
 
 test('buildStartupEnvFromProfile treats explicit falsey provider flags as user intent', async () => {

src/utils/providerProfile.ts

@@ -841,43 +841,35 @@ export async function buildStartupEnvFromProfile(options?: {
   const processEnv = options?.processEnv ?? process.env
   const persisted = options?.persisted ?? loadProfileFile()
 
-  // Saved /provider profiles should still win over provider-manager env that was
-  // auto-applied during startup. Only an explicit shell/flag provider selection
-  // should bypass the persisted startup profile.
-  //
   const profileManagedEnv = processEnv.CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED === '1'
 
-  // If the user explicitly selected a provider via env, allow it to bypass
-  // the persisted profile only when we can prove it was managed by the
-  // persisted profile env itself.
+  // The legacy single-profile file (~/.openclaude-profile.json) is a
+  // first-run / fallback mechanism. The newer plural provider-profile
+  // system (`/provider` presets + activeProviderProfileId in config) is
+  // applied earlier in the bootstrap via applyActiveProviderProfileFromConfig
+  // and signals completion with CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED=1.
   //
-  // Practically: on initial startup, provider routing env vars can already
-  // be present due to earlier auto-application steps. We should still apply
-  // the persisted profile rather than returning early.
+  // If the plural system has already set env, trust it — do NOT overlay the
+  // legacy file. addProviderProfile() does not sync the legacy file, so a
+  // stale legacy file (e.g. OpenAI defaults from an earlier manual setup)
+  // would otherwise overwrite the correct plural env and surface as the
+  // "banner shows gpt-4o / api.openai.com even though my saved profile is
+  // Moonshot" bug.
+  if (profileManagedEnv) {
+    return processEnv
+  }
 
   if (!persisted) {
     return processEnv
   }
 
-  const launchProcessEnv = profileManagedEnv
-    ? (() => {
-        const cleanedEnv = { ...processEnv }
-        for (const key of PROFILE_ENV_KEYS) {
-          delete cleanedEnv[key]
-        }
-        delete cleanedEnv.CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED
-        delete cleanedEnv.CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED_ID
-        return cleanedEnv
-      })()
-    : processEnv
-
   return buildLaunchEnv({
     profile: persisted.profile,
     persisted,
     goal:
       options?.goal ??
       normalizeRecommendationGoal(processEnv.OPENCLAUDE_PROFILE_GOAL),
-    processEnv: launchProcessEnv,
+    processEnv,
     getOllamaChatBaseUrl:
       options?.getOllamaChatBaseUrl ?? getOllamaChatBaseUrl,
     resolveOllamaDefaultModel: options?.resolveOllamaDefaultModel,

src/utils/providerProfiles.test.ts

@@ -256,6 +256,83 @@ describe('applyActiveProviderProfileFromConfig', () => {
     expect(process.env.OPENAI_MODEL).toBe('qwen2.5:3b')
   })
 
+  test('applies active profile when a bare CLAUDE_CODE_USE_OPENAI flag is stale (no BASE_URL/MODEL)', async () => {
+    // Regression: a leftover `CLAUDE_CODE_USE_OPENAI=1` in the shell with no
+    // paired OPENAI_BASE_URL / OPENAI_MODEL is not a real explicit selection
+    // — it's a stale export. The previous guard treated it as intent and
+    // skipped the saved profile, causing the startup banner to show hardcoded
+    // defaults (gpt-4o @ api.openai.com) instead of the user's active
+    // profile.
+    const { applyActiveProviderProfileFromConfig } =
+      await importFreshProviderProfileModules()
+    process.env.CLAUDE_CODE_USE_OPENAI = '1'
+    delete process.env.OPENAI_BASE_URL
+    delete process.env.OPENAI_API_BASE
+    delete process.env.OPENAI_MODEL
+
+    const applied = applyActiveProviderProfileFromConfig({
+      providerProfiles: [
+        buildProfile({
+          id: 'saved_moonshot',
+          baseUrl: 'https://api.moonshot.ai/v1',
+          model: 'kimi-k2.6',
+        }),
+      ],
+      activeProviderProfileId: 'saved_moonshot',
+    } as any)
+
+    expect(applied?.id).toBe('saved_moonshot')
+    expect(process.env.OPENAI_BASE_URL).toBe('https://api.moonshot.ai/v1')
+    expect(process.env.OPENAI_MODEL).toBe('kimi-k2.6')
+  })
+
+  test('still respects complete shell selection with USE flag + BASE_URL', async () => {
+    // Counter-example: when the user really did set both the flag AND a
+    // concrete BASE_URL, that IS explicit intent and wins over the saved
+    // profile. This preserves the original "explicit startup wins" semantic.
+    const { applyActiveProviderProfileFromConfig } =
+      await importFreshProviderProfileModules()
+    process.env.CLAUDE_CODE_USE_OPENAI = '1'
+    process.env.OPENAI_BASE_URL = 'http://192.168.1.1:8080/v1'
+    delete process.env.OPENAI_MODEL
+
+    const applied = applyActiveProviderProfileFromConfig({
+      providerProfiles: [
+        buildProfile({
+          id: 'saved_moonshot',
+          baseUrl: 'https://api.moonshot.ai/v1',
+          model: 'kimi-k2.6',
+        }),
+      ],
+      activeProviderProfileId: 'saved_moonshot',
+    } as any)
+
+    expect(applied).toBeUndefined()
+    expect(process.env.OPENAI_BASE_URL).toBe('http://192.168.1.1:8080/v1')
+  })
+
+  test('still respects complete shell selection with USE flag + MODEL', async () => {
+    const { applyActiveProviderProfileFromConfig } =
+      await importFreshProviderProfileModules()
+    process.env.CLAUDE_CODE_USE_OPENAI = '1'
+    process.env.OPENAI_MODEL = 'gpt-4o-mini'
+    delete process.env.OPENAI_BASE_URL
+
+    const applied = applyActiveProviderProfileFromConfig({
+      providerProfiles: [
+        buildProfile({
+          id: 'saved_moonshot',
+          baseUrl: 'https://api.moonshot.ai/v1',
+          model: 'kimi-k2.6',
+        }),
+      ],
+      activeProviderProfileId: 'saved_moonshot',
+    } as any)
+
+    expect(applied).toBeUndefined()
+    expect(process.env.OPENAI_MODEL).toBe('gpt-4o-mini')
+  })
+
   test('does not override explicit startup selection when profile marker is stale', async () => {
     const { applyActiveProviderProfileFromConfig } =
       await importFreshProviderProfileModules()
@@ -450,6 +527,18 @@ describe('getProviderPresetDefaults', () => {
     expect(defaults.baseUrl).toBe('http://localhost:11434/v1')
     expect(defaults.model).toBe('llama3.1:8b')
   })
+
+  test('atomic-chat preset defaults to a local Atomic Chat endpoint', async () => {
+    const { getProviderPresetDefaults } = await importFreshProviderProfileModules()
+    delete process.env.OPENAI_MODEL
+
+    const defaults = getProviderPresetDefaults('atomic-chat')
+
+    expect(defaults.provider).toBe('openai')
+    expect(defaults.name).toBe('Atomic Chat')
+    expect(defaults.baseUrl).toBe('http://127.0.0.1:1337/v1')
+    expect(defaults.requiresApiKey).toBe(false)
+  })
 })
 
 describe('setActiveProviderProfile', () => {

src/utils/providerProfiles.ts

@@ -33,6 +33,7 @@ export type ProviderPreset =
   | 'custom'
   | 'nvidia-nim'
   | 'minimax'
+  | 'atomic-chat'
 
 export type ProviderProfileInput = {
   provider?: ProviderProfile['provider']
@@ -285,6 +286,15 @@ export function getProviderPresetDefaults(
         apiKey: process.env.MINIMAX_API_KEY ?? '',
         requiresApiKey: true,
       }
+    case 'atomic-chat':
+      return {
+        provider: 'openai',
+        name: 'Atomic Chat',
+        baseUrl: 'http://127.0.0.1:1337/v1',
+        model: process.env.OPENAI_MODEL ?? 'local-model',
+        apiKey: '',
+        requiresApiKey: false,
+      }
     case 'ollama':
     default:
       return {
@@ -322,6 +332,58 @@ function hasProviderSelectionFlags(
   )
 }
 
+/**
+ * A "complete" explicit provider selection = a USE flag AND at least one
+ * concrete config value that tells us WHERE to route (a base URL) or WHAT
+ * to run (a model id). A bare `CLAUDE_CODE_USE_OPENAI=1` with nothing else
+ * is almost always a stale shell export from a previous session, not real
+ * intent — and if we respect it, we skip the user's saved active profile
+ * and fall back to hardcoded defaults (gpt-4o / api.openai.com), which is
+ * the exact bug users report as "my saved provider isn't picked up".
+ *
+ * Used to gate whether saved-profile env should override shell state at
+ * startup. The weaker `hasProviderSelectionFlags` is still used for the
+ * anthropic-profile conflict check (any flag is a conflict for
+ * first-party anthropic) and for alignment fingerprinting.
+ */
+function hasCompleteProviderSelection(
+  processEnv: NodeJS.ProcessEnv = process.env,
+): boolean {
+  if (!hasProviderSelectionFlags(processEnv)) return false
+  if (processEnv.CLAUDE_CODE_USE_OPENAI !== undefined) {
+    return (
+      trimOrUndefined(processEnv.OPENAI_BASE_URL) !== undefined ||
+      trimOrUndefined(processEnv.OPENAI_API_BASE) !== undefined ||
+      trimOrUndefined(processEnv.OPENAI_MODEL) !== undefined
+    )
+  }
+  if (processEnv.CLAUDE_CODE_USE_GEMINI !== undefined) {
+    return (
+      trimOrUndefined(processEnv.GEMINI_BASE_URL) !== undefined ||
+      trimOrUndefined(processEnv.GEMINI_MODEL) !== undefined ||
+      trimOrUndefined(processEnv.GEMINI_API_KEY) !== undefined ||
+      trimOrUndefined(processEnv.GOOGLE_API_KEY) !== undefined
+    )
+  }
+  if (processEnv.CLAUDE_CODE_USE_MISTRAL !== undefined) {
+    return (
+      trimOrUndefined(processEnv.MISTRAL_BASE_URL) !== undefined ||
+      trimOrUndefined(processEnv.MISTRAL_MODEL) !== undefined ||
+      trimOrUndefined(processEnv.MISTRAL_API_KEY) !== undefined
+    )
+  }
+  if (processEnv.CLAUDE_CODE_USE_GITHUB !== undefined) {
+    return (
+      trimOrUndefined(processEnv.GITHUB_TOKEN) !== undefined ||
+      trimOrUndefined(processEnv.GH_TOKEN) !== undefined ||
+      trimOrUndefined(processEnv.OPENAI_MODEL) !== undefined
+    )
+  }
+  // Bedrock / Vertex / Foundry signal cloud-provider routing in env; treat
+  // the flag alone as complete (these paths rely on ambient AWS/GCP creds).
+  return true
+}
+
 function hasConflictingProviderFlagsForProfile(
   processEnv: NodeJS.ProcessEnv,
   profile: ProviderProfile,
@@ -564,9 +626,15 @@ export function applyActiveProviderProfileFromConfig(
     processEnv[PROFILE_ENV_APPLIED_FLAG] === '1' &&
     trimOrUndefined(processEnv[PROFILE_ENV_APPLIED_ID]) === activeProfile.id
 
-  if (!options?.force && (hasProviderSelectionFlags(processEnv) || processEnv[PROFILE_ENV_APPLIED_FLAG] === '1')) {
+  if (!options?.force && (hasCompleteProviderSelection(processEnv) || processEnv[PROFILE_ENV_APPLIED_FLAG] === '1')) {
     // Respect explicit startup provider intent. Auto-heal only when this
     // exact active profile previously applied the current env.
+    // NOTE: we gate on hasCompleteProviderSelection (flag + concrete config)
+    // rather than hasProviderSelectionFlags alone. A bare CLAUDE_CODE_USE_*=1
+    // with no BASE_URL/MODEL is almost always a stale shell export, not
+    // intent — respecting it would skip the saved profile and fall through
+    // to hardcoded provider defaults, which surfaces as "my saved provider
+    // isn't being picked up at startup".
     if (!isCurrentEnvProfileManaged) {
       return undefined
     }

src/utils/sandbox/sandbox-adapter.ts

@@ -456,10 +456,19 @@ const checkDependencies = memoize((): SandboxDependencyCheck => {
   })
 })
 
+/**
+ * Read sandbox.enabled only from trusted settings sources.
+ * projectSettings is intentionally excluded — a malicious repo could
+ * otherwise disable the sandbox via .claude/settings.json.
+ */
 function getSandboxEnabledSetting(): boolean {
   try {
-    const settings = getSettings_DEPRECATED()
-    return settings?.sandbox?.enabled ?? false
+    return !!(
+      getSettingsForSource('userSettings')?.sandbox?.enabled ||
+      getSettingsForSource('localSettings')?.sandbox?.enabled ||
+      getSettingsForSource('flagSettings')?.sandbox?.enabled ||
+      getSettingsForSource('policySettings')?.sandbox?.enabled
+    )
   } catch (error) {
     logForDebugging(`Failed to get settings for sandbox check: ${error}`)
     return false

src/utils/settings/settings.ts

@@ -300,9 +300,9 @@ export function getRelativeSettingsFilePathForSource(
 ): string {
   switch (source) {
     case 'projectSettings':
-      return join('.openclaude', 'settings.json')
+      return '.openclaude/settings.json'
     case 'localSettings':
-      return join('.openclaude', 'settings.local.json')
+      return '.openclaude/settings.local.json'
   }
 }
 

src/utils/swarm/permissionSync.ts

@@ -207,6 +207,10 @@ export function createPermissionRequest(params: {
 }
 
 /**
+ * @deprecated Use sendPermissionRequestViaMailbox() instead. This file-based
+ * approach writes to an unauthenticated directory where any local process can
+ * forge requests. Retained for backward compatibility but no longer called.
+ *
  * Write a permission request to the pending directory with file locking
  * Called by worker agents when they need permission approval from the leader
  *
@@ -250,6 +254,10 @@ export async function writePermissionRequest(
 }
 
 /**
+ * @deprecated No longer called — permission requests are sent via mailbox.
+ * The pending directory is an unauthenticated channel. Retained for backward
+ * compatibility.
+ *
  * Read all pending permission requests for a team
  * Called by the team leader to see what requests need attention
  */
@@ -312,6 +320,11 @@ export async function readPendingPermissions(
 }
 
 /**
+ * @deprecated No longer called — permission responses are delivered via mailbox
+ * (processMailboxPermissionResponse). The resolved directory is an unauthenticated
+ * channel where any local process can forge approvals. Retained for backward
+ * compatibility.
+ *
  * Read a resolved permission request by ID
  * Called by workers to check if their request has been resolved
  *
@@ -352,6 +365,10 @@ export async function readResolvedPermission(
 }
 
 /**
+ * @deprecated Use sendPermissionResponseViaMailbox() instead. This file-based
+ * approach writes to an unauthenticated directory where any local process can
+ * forge approvals. Retained for backward compatibility but no longer called.
+ *
  * Resolve a permission request
  * Called by the team leader (or worker in self-resolution cases)
  *
@@ -536,6 +553,10 @@ export type PermissionResponse = {
 }
 
 /**
+ * @deprecated Use processMailboxPermissionResponse() via useInboxPoller instead.
+ * File-based polling reads from an unauthenticated directory where any local
+ * process can forge approval files. Retained for backward compatibility.
+ *
  * Poll for a permission response (worker-side convenience function)
  * Converts the resolved request into a simpler response format
  *
@@ -564,6 +585,9 @@ export async function pollForResponse(
 }
 
 /**
+ * @deprecated File-based response cleanup is no longer needed — responses are
+ * delivered via mailbox. Retained for backward compatibility.
+ *
  * Remove a worker's response after processing
  * This is an alias for deleteResolvedPermission for backward compatibility
  */
@@ -601,6 +625,9 @@ export function isSwarmWorker(): boolean {
 }
 
 /**
+ * @deprecated File-based resolved permissions are no longer written. Responses
+ * are delivered via mailbox. Retained for backward compatibility.
+ *
  * Delete a resolved permission file
  * Called after a worker has processed the resolution
  */
@@ -635,8 +662,8 @@ export async function deleteResolvedPermission(
 }
 
 /**
- * Submit a permission request (alias for writePermissionRequest)
- * Provided for backward compatibility with worker integration code
+ * @deprecated Alias for writePermissionRequest, which is itself deprecated.
+ * Use sendPermissionRequestViaMailbox() instead.
  */
 export const submitPermissionRequest = writePermissionRequest
 

src/utils/thinkingTokenExtractor.test.ts

@@ -0,0 +1,106 @@
+import { describe, expect, it } from 'bun:test'
+import { ThinkingTokenAnalyzer } from './thinkingTokenExtractor.js'
+
+describe('ThinkingTokenAnalyzer', () => {
+  describe('extract', () => {
+    it('extracts thinking and output separately', () => {
+      const message = {
+        type: 'assistant',
+        message: {
+          content: [
+            { type: 'thinking', thinking: 'Let me think about this...' },
+            { type: 'text', text: 'Here is my answer.' },
+          ],
+        },
+      } as any
+
+      const result = ThinkingTokenAnalyzer.extract(message)
+
+      expect(result.thinking).toBeGreaterThan(0)
+      expect(result.output).toBeGreaterThan(0)
+      expect(result.total).toBe(result.thinking + result.output)
+    })
+
+    it('handles no thinking', () => {
+      const message = {
+        type: 'assistant',
+        message: {
+          content: [{ type: 'text', text: 'Hello world' }],
+        },
+      } as any
+
+      const result = ThinkingTokenAnalyzer.extract(message)
+
+      expect(result.thinking).toBe(0)
+      expect(result.output).toBeGreaterThan(0)
+    })
+
+    it('handles redacted thinking', () => {
+      const message = {
+        type: 'assistant',
+        message: {
+          content: [
+            { type: 'redacted_thinking', data: '[thinking hidden]' },
+            { type: 'text', text: 'Answer here.' },
+          ],
+        },
+      } as any
+
+      const result = ThinkingTokenAnalyzer.extract(message)
+
+      expect(result.thinking).toBeGreaterThan(0)
+      expect(result.output).toBeGreaterThan(0)
+    })
+  })
+
+  describe('analyze', () => {
+    it('calculates percentages', () => {
+      const message = {
+        type: 'assistant',
+        message: {
+          content: [
+            { type: 'thinking', thinking: 'Thinking1 Thinking2 Thinking3' },
+            { type: 'text', text: 'Output1 Output2' },
+          ],
+        },
+      } as any
+
+      const analysis = ThinkingTokenAnalyzer.analyze(message)
+
+      expect(analysis.hasThinking).toBe(true)
+      expect(analysis.thinkingPercentage).toBeGreaterThan(0)
+      expect(analysis.outputPercentage).toBeGreaterThan(0)
+      expect(analysis.reasoningComplexity).toBeTruthy()
+    })
+  })
+
+  describe('hasSignificantThinking', () => {
+    it('detects significant thinking', () => {
+      const message = {
+        type: 'assistant',
+        message: {
+          content: [
+            { type: 'thinking', thinking: 'x'.repeat(500) },
+            { type: 'text', text: 'short' },
+          ],
+        },
+      } as any
+
+      expect(ThinkingTokenAnalyzer.hasSignificantThinking(message, 20)).toBe(true)
+    })
+
+    it('rejects minimal thinking', () => {
+      const message = {
+        type: 'assistant',
+        message: {
+          content: [
+            { type: 'thinking', thinking: 'a' },
+            { type: 'text', text: 'much longer output text here with more content' },
+          ],
+        },
+      } as any
+
+      expect(ThinkingTokenAnalyzer.hasSignificantThinking(message, 20)).toBe(false)
+    })
+  })
+})

src/utils/thinkingTokenExtractor.ts

@@ -0,0 +1,192 @@
+/**
+ * Thinking Token Extractor - Production-grade thinking token analysis
+ * 
+ * Extracts and analyzes thinking tokens from assistant messages.
+ * Provides detailed breakdown, statistics, and insights.
+ */
+
+import { roughTokenCountEstimation } from '../services/tokenEstimation.js'
+import { jsonStringify } from './slowOperations.js'
+import type { AssistantMessage, Message } from '../types/message.js'
+
+export interface ThinkingBlock {
+  type: 'thinking' | 'redacted_thinking'
+  content: string
+  tokens: number
+}
+
+export interface OutputBlock {
+  type: 'text' | 'tool_use'
+  content: string
+  tokens: number
+}
+
+export interface ThinkingTokenBreakdown {
+  thinking: number
+  output: number
+  total: number
+  thinkingBlocks: ThinkingBlock[]
+  outputBlocks: OutputBlock[]
+}
+
+export interface ThinkingAnalysis {
+  hasThinking: boolean
+  thinkingPercentage: number
+  outputPercentage: number
+  blockCount: number
+  avgThinkingBlockSize: number
+  avgOutputBlockSize: number
+  totalTextLength: number
+  reasoningComplexity: 'low' | 'medium' | 'high'
+}
+
+export class ThinkingTokenAnalyzer {
+  /**
+   * Extract detailed thinking vs output breakdown
+   */
+  static extract(message: AssistantMessage): ThinkingTokenBreakdown {
+    const thinkingBlocks: ThinkingBlock[] = []
+    const outputBlocks: OutputBlock[] = []
+    let thinking = 0
+    let output = 0
+
+    for (const block of message.message.content) {
+      if (block.type === 'thinking') {
+        const tokens = roughTokenCountEstimation(block.thinking)
+        thinking += tokens
+        thinkingBlocks.push({
+          type: 'thinking',
+          content: block.thinking,
+          tokens,
+        })
+      } else if (block.type === 'redacted_thinking') {
+        const tokens = roughTokenCountEstimation(block.data)
+        thinking += tokens
+        thinkingBlocks.push({
+          type: 'redacted_thinking',
+          content: block.data,
+          tokens,
+        })
+      } else if (block.type === 'text') {
+        const tokens = roughTokenCountEstimation(block.text)
+        output += tokens
+        outputBlocks.push({
+          type: 'text',
+          content: block.text,
+          tokens,
+        })
+      } else if (block.type === 'tool_use') {
+        const content = jsonStringify(block.input)
+        const tokens = roughTokenCountEstimation(content)
+        output += tokens
+        outputBlocks.push({
+          type: 'tool_use',
+          content,
+          tokens,
+        })
+      }
+    }
+
+    return {
+      thinking,
+      output,
+      total: thinking + output,
+      thinkingBlocks,
+      outputBlocks,
+    }
+  }
+
+  /**
+   * Simple extraction for quick use
+   */
+  static extractSimple(message: AssistantMessage): ThinkingTokenBreakdown {
+    return this.extract(message)
+  }
+
+  /**
+   * Analyze thinking patterns and provide insights
+   */
+  static analyze(message: AssistantMessage): ThinkingAnalysis {
+    const breakdown = this.extract(message)
+    const { thinking, output, total, thinkingBlocks, outputBlocks } = breakdown
+
+    const hasThinking = thinking > 0
+    const thinkingPercentage = total > 0 ? (thinking / total) * 100 : 0
+    const outputPercentage = total > 0 ? (output / total) * 100 : 0
+
+    const avgThinkingBlockSize = thinkingBlocks.length > 0
+      ? thinkingBlocks.reduce((sum, b) => sum + b.tokens, 0) / thinkingBlocks.length
+      : 0
+
+    const avgOutputBlockSize = outputBlocks.length > 0
+      ? outputBlocks.reduce((sum, b) => sum + b.tokens, 0) / outputBlocks.length
+      : 0
+
+    const totalTextLength = [...thinkingBlocks, ...outputBlocks].reduce(
+      (sum, b) => sum + b.content.length,
+      0,
+    )
+
+    // Complexity based on thinking percentage and block count
+    let reasoningComplexity: 'low' | 'medium' | 'high' = 'low'
+    if (thinkingPercentage > 30 || thinkingBlocks.length > 5) {
+      reasoningComplexity = 'high'
+    } else if (thinkingPercentage > 10 || thinkingBlocks.length > 2) {
+      reasoningComplexity = 'medium'
+    }
+
+    return {
+      hasThinking,
+      thinkingPercentage: Math.round(thinkingPercentage * 10) / 10,
+      outputPercentage: Math.round(outputPercentage * 10) / 10,
+      blockCount: thinkingBlocks.length + outputBlocks.length,
+      avgThinkingBlockSize: Math.round(avgThinkingBlockSize),
+      avgOutputBlockSize: Math.round(avgOutputBlockSize),
+      totalTextLength,
+      reasoningComplexity,
+    }
+  }
+
+  /**
+   * Check if message has significant thinking
+   */
+  static hasSignificantThinking(
+    message: AssistantMessage,
+    thresholdPercent = 20,
+  ): boolean {
+    const analysis = this.analyze(message)
+    return analysis.thinkingPercentage >= thresholdPercent
+  }
+
+  /**
+   * Get thinking-only messages from an array
+   */
+  static filterThinkingMessages(messages: Message[]): AssistantMessage[] {
+    return messages
+      .filter((m): m is AssistantMessage => m.type === 'assistant')
+      .filter(m => this.hasSignificantThinking(m))
+  }
+
+  /**
+   * Calculate total thinking tokens across messages
+   */
+  static totalThinkingTokens(messages: Message[]): number {
+    return messages
+      .filter((m): m is AssistantMessage => m.type === 'assistant')
+      .reduce((sum, m) => sum + this.extract(m).thinking, 0)
+  }
+}
+
+/**
+ * Legacy export for backward compatibility
+ */
+export function extractThinkingTokens(
+  message: AssistantMessage,
+): { thinking: number; output: number; total: number } {
+  const result = ThinkingTokenAnalyzer.extract(message)
+  return {
+    thinking: result.thinking,
+    output: result.output,
+    total: result.total,
+  }
+}

src/utils/thinkingTokens.test.ts

@@ -0,0 +1,69 @@
+import { describe, expect, it } from 'bun:test'
+import { extractThinkingTokens } from './tokens.js'
+
+describe('extractThinkingTokens', () => {
+  it('extracts thinking and output separately', () => {
+    const message = {
+      type: 'assistant',
+      message: {
+        content: [
+          { type: 'thinking', thinking: 'Let me think about this...' },
+          { type: 'text', text: 'Here is my answer.' },
+        ],
+      },
+    } as any
+
+    const result = extractThinkingTokens(message)
+
+    expect(result.thinking).toBeGreaterThan(0)
+    expect(result.output).toBeGreaterThan(0)
+    expect(result.total).toBe(result.thinking + result.output)
+  })
+
+  it('handles no thinking', () => {
+    const message = {
+      type: 'assistant',
+      message: {
+        content: [{ type: 'text', text: 'Hello world' }],
+      },
+    } as any
+
+    const result = extractThinkingTokens(message)
+
+    expect(result.thinking).toBe(0)
+    expect(result.output).toBeGreaterThan(0)
+  })
+
+  it('handles redacted thinking', () => {
+    const message = {
+      type: 'assistant',
+      message: {
+        content: [
+          { type: 'redacted_thinking', data: '[thinking hidden]' },
+          { type: 'text', text: 'Answer here.' },
+        ],
+      },
+    } as any
+
+    const result = extractThinkingTokens(message)
+
+    expect(result.thinking).toBeGreaterThan(0)
+    expect(result.output).toBeGreaterThan(0)
+  })
+
+  it('handles tool use', () => {
+    const message = {
+      type: 'assistant',
+      message: {
+        content: [
+          { type: 'tool_use', id: 'tool_1', name: 'bash', input: { cmd: 'echo test' } },
+          { type: 'text', text: 'Ran command.' },
+        ],
+      },
+    } as any
+
+    const result = extractThinkingTokens(message)
+
+    expect(result.output).toBeGreaterThan(0)
+  })
+})

src/utils/tokenAnalytics.test.ts

@@ -0,0 +1,84 @@
+import { describe, expect, it, beforeEach } from 'bun:test'
+import { TokenUsageTracker } from './tokenAnalytics.js'
+
+describe('TokenUsageTracker', () => {
+  let tracker: TokenUsageTracker
+
+  beforeEach(() => {
+    tracker = new TokenUsageTracker(100)
+  })
+
+  it('records token usage', () => {
+    tracker.record({
+      input_tokens: 1000,
+      output_tokens: 500,
+      cache_read_input_tokens: 200,
+      cache_creation_input_tokens: 100,
+      model: 'claude-sonnet-4-5-20250514',
+    })
+
+    expect(tracker.size).toBe(1)
+  })
+
+  it('calculates analytics', () => {
+    tracker.record({
+      input_tokens: 1000,
+      output_tokens: 500,
+      model: 'claude-sonnet-4-5-20250514',
+    })
+
+    tracker.record({
+      input_tokens: 2000,
+      output_tokens: 300,
+      model: 'claude-sonnet-4-5-20250514',
+    })
+
+    const analytics = tracker.getAnalytics()
+
+    expect(analytics.totalRequests).toBe(2)
+    expect(analytics.totalInputTokens).toBe(3000)
+    expect(analytics.totalOutputTokens).toBe(800)
+    expect(analytics.averageInputPerRequest).toBe(1500)
+    expect(analytics.averageOutputPerRequest).toBe(400)
+  })
+
+  it('tracks cache hit rate', () => {
+    tracker.record({
+      input_tokens: 1000,
+      output_tokens: 500,
+      cache_read_input_tokens: 500, // 33% cache
+      model: 'claude-sonnet-4-5-20250514',
+    })
+
+    const analytics = tracker.getAnalytics()
+
+    expect(analytics.cacheHitRate).toBeGreaterThan(0)
+  })
+
+  it('tracks most used model', () => {
+    tracker.record({ input_tokens: 1000, output_tokens: 100, model: 'sonnet' })
+    tracker.record({ input_tokens: 1000, output_tokens: 100, model: 'sonnet' })
+    tracker.record({ input_tokens: 1000, output_tokens: 100, model: 'opus' })
+
+    expect(tracker.getAnalytics().mostUsedModel).toBe('sonnet')
+  })
+
+  it('respects max entries limit', () => {
+    const smallTracker = new TokenUsageTracker(3)
+
+    smallTracker.record({ input_tokens: 1, output_tokens: 1, model: 'a' })
+    smallTracker.record({ input_tokens: 2, output_tokens: 2, model: 'b' })
+    smallTracker.record({ input_tokens: 3, output_tokens: 3, model: 'c' })
+    smallTracker.record({ input_tokens: 4, output_tokens: 4, model: 'd' })
+    smallTracker.record({ input_tokens: 5, output_tokens: 5, model: 'e' })
+
+    expect(smallTracker.size).toBe(3)
+  })
+
+it('clears history', () => {
+      tracker.record({ input_tokens: 1000, output_tokens: 100, model: 'test' })
+      tracker.clear()
+
+      expect(tracker.size).toBe(0)
+    })
+})

src/utils/tokenAnalytics.ts

@@ -0,0 +1,211 @@
+/**
+ * Token Analytics - Historical token usage tracking and analysis
+ * 
+ * Tracks token usage patterns over time for cost optimization
+ * and capacity planning.
+ */
+
+import type { BetaUsage as Usage } from '@anthropic-ai/sdk/resources/beta/messages/messages.mjs'
+
+export interface TokenUsageEntry {
+  timestamp: number
+  inputTokens: number
+  outputTokens: number
+  cacheReadTokens: number
+  cacheCreationTokens: number
+  model: string
+}
+
+export interface TokenAnalytics {
+  totalRequests: number
+  totalInputTokens: number
+  totalOutputTokens: number
+  totalCacheRead: number
+  totalCacheCreation: number
+  averageInputPerRequest: number
+  averageOutputPerRequest: number
+  cacheHitRate: number
+  mostUsedModel: string
+  requestsLastHour: number
+  requestsLastDay: number
+}
+
+/**
+ * Historical Token Analytics Tracker
+ * 
+ * Tracks token usage patterns over time for analytics,
+ * cost optimization, and capacity planning.
+ */
+export class TokenUsageTracker {
+  private history: TokenUsageEntry[] = []
+  private readonly maxEntries: number
+
+  constructor(maxEntries = 1000) {
+    this.maxEntries = maxEntries
+  }
+
+  /**
+   * Record a token usage event from API response.
+   */
+  record(usage: {
+    input_tokens: number
+    output_tokens: number
+    cache_read_input_tokens?: number
+    cache_creation_input_tokens?: number
+    model: string
+  }): void {
+    const entry: TokenUsageEntry = {
+      timestamp: Date.now(),
+      inputTokens: usage.input_tokens,
+      outputTokens: usage.output_tokens,
+      cacheReadTokens: usage.cache_read_input_tokens ?? 0,
+      cacheCreationTokens: usage.cache_creation_input_tokens ?? 0,
+      model: usage.model,
+    }
+
+    this.history.push(entry)
+
+    if (this.history.length > this.maxEntries) {
+      this.history = this.history.slice(-this.maxEntries)
+    }
+  }
+
+  /**
+   * Get analytics summary for all recorded usage.
+   */
+  getAnalytics(): TokenAnalytics {
+    if (this.history.length === 0) {
+      return {
+        totalRequests: 0,
+        totalInputTokens: 0,
+        totalOutputTokens: 0,
+        totalCacheRead: 0,
+        totalCacheCreation: 0,
+        averageInputPerRequest: 0,
+        averageOutputPerRequest: 0,
+        cacheHitRate: 0,
+        mostUsedModel: 'unknown',
+        requestsLastHour: 0,
+        requestsLastDay: 0,
+      }
+    }
+
+    const now = Date.now()
+    const hourAgo = now - 60 * 60 * 1000
+    const dayAgo = now - 24 * 60 * 60 * 1000
+
+    let totalInput = 0
+    let totalOutput = 0
+    let totalCacheRead = 0
+    let totalCacheCreation = 0
+    const modelCounts = new Map<string, number>()
+    let requestsLastHour = 0
+    let requestsLastDay = 0
+
+    for (const entry of this.history) {
+      totalInput += entry.inputTokens
+      totalOutput += entry.outputTokens
+      totalCacheRead += entry.cacheReadTokens
+      totalCacheCreation += entry.cacheCreationTokens
+
+      modelCounts.set(entry.model, (modelCounts.get(entry.model) ?? 0) + 1)
+
+      if (entry.timestamp >= hourAgo) requestsLastHour++
+      if (entry.timestamp >= dayAgo) requestsLastDay++
+    }
+
+    let mostUsedModel = 'unknown'
+    let maxCount = 0
+    for (const [model, count] of modelCounts) {
+      if (count > maxCount) {
+        maxCount = count
+        mostUsedModel = model
+      }
+    }
+
+    const totalRequests = this.history.length
+    const totalCache = totalCacheRead + totalCacheCreation
+    const totalTokens = totalInput + totalOutput + totalCache
+    const cacheHitRate = totalTokens > 0 ? (totalCacheRead / totalTokens) * 100 : 0
+
+    return {
+      totalRequests,
+      totalInputTokens: totalInput,
+      totalOutputTokens: totalOutput,
+      totalCacheRead,
+      totalCacheCreation,
+      averageInputPerRequest: Math.round(totalInput / totalRequests),
+      averageOutputPerRequest: Math.round(totalOutput / totalRequests),
+      cacheHitRate: Math.round(cacheHitRate),
+      mostUsedModel,
+      requestsLastHour,
+      requestsLastDay,
+    }
+  }
+
+  /**
+   * Get recent entries within time window.
+   */
+  getRecent(windowMs: number): TokenUsageEntry[] {
+    const cutoff = Date.now() - windowMs
+    return this.history.filter(e => e.timestamp >= cutoff)
+  }
+
+  /**
+   * Get entries for a specific model
+   */
+  getByModel(model: string): TokenUsageEntry[] {
+    return this.history.filter(e => e.model === model)
+  }
+
+  /**
+   * Calculate cost estimate (approximate)
+   */
+  estimateCost(): { input: number; output: number; cache: number } {
+    const analytics = this.getAnalytics()
+    
+    // Approximate pricing (adjust as needed)
+    const inputCost = analytics.totalInputTokens * 0.00015
+    const outputCost = analytics.totalOutputTokens * 0.0006
+    const cacheCost = analytics.totalCacheRead * 0.000075
+    
+    return {
+      input: Math.round(inputCost * 100) / 100,
+      output: Math.round(outputCost * 100) / 100,
+      cache: Math.round(cacheCost * 100) / 100,
+    }
+  }
+
+  /**
+   * Clear history.
+   */
+  clear(): void {
+    this.history = []
+  }
+
+  /**
+   * Get history size.
+   */
+  get size(): number {
+    return this.history.length
+  }
+
+  /**
+   * Export history as JSON
+   */
+  export(): string {
+    return JSON.stringify(this.history, null, 2)
+  }
+
+  /**
+   * Import history from JSON
+   */
+  import(json: string): void {
+    try {
+      const entries = JSON.parse(json) as TokenUsageEntry[]
+      this.history = entries.slice(-this.maxEntries)
+    } catch {
+      // Invalid JSON, ignore
+    }
+  }
+}

src/utils/tokens.ts

@@ -1,5 +1,5 @@
 import type { BetaUsage as Usage } from '@anthropic-ai/sdk/resources/beta/messages/messages.mjs'
-import { roughTokenCountEstimationForMessages } from '../services/tokenEstimation.js'
+import { roughTokenCountEstimation, roughTokenCountEstimationForMessages } from '../services/tokenEstimation.js'
 import type { AssistantMessage, Message } from '../types/message.js'
 import { SYNTHETIC_MESSAGES, SYNTHETIC_MODEL } from './messages.js'
 import { jsonStringify } from './slowOperations.js'
@@ -198,6 +198,198 @@ export function getAssistantMessageContentLength(
   return contentLength
 }
 
+/**
+ * Extract thinking tokens from an assistant message.
+ * Returns breakdown of thinking vs output tokens.
+ */
+export function extractThinkingTokens(
+  message: AssistantMessage,
+): { thinking: number; output: number; total: number } {
+  let thinking = 0
+  let output = 0
+
+  for (const block of message.message.content) {
+    if (block.type === 'thinking') {
+      thinking += roughTokenCountEstimation(block.thinking)
+    } else if (block.type === 'redacted_thinking') {
+      thinking += roughTokenCountEstimation(block.data)
+    } else if (block.type === 'text') {
+      output += roughTokenCountEstimation(block.text)
+    } else if (block.type === 'tool_use') {
+      output += roughTokenCountEstimation(jsonStringify(block.input))
+    }
+  }
+
+  return { thinking, output, total: thinking + output }
+}
+
+/**
+ * Token usage history entry for tracking patterns over time.
+ */
+export interface TokenUsageEntry {
+  timestamp: number
+  inputTokens: number
+  outputTokens: number
+  cacheReadTokens: number
+  cacheCreationTokens: number
+  model: string
+}
+
+/**
+ * Token analytics summary from historical data.
+ */
+export interface TokenAnalytics {
+  totalRequests: number
+  totalInputTokens: number
+  totalOutputTokens: number
+  totalCacheRead: number
+  totalCacheCreation: number
+  averageInputPerRequest: number
+  averageOutputPerRequest: number
+  cacheHitRate: number
+  mostUsedModel: string
+  requestsLastHour: number
+  requestsLastDay: number
+}
+
+/**
+ * Historical Token Analytics Tracker
+ * 
+ * Tracks token usage patterns over time for analytics,
+ * cost optimization, and capacity planning.
+ */
+export class TokenUsageTracker {
+  private history: TokenUsageEntry[] = []
+  private readonly maxEntries: number
+
+  constructor(maxEntries = 1000) {
+    this.maxEntries = maxEntries
+  }
+
+  /**
+   * Record a token usage event from API response.
+   */
+  record(usage: {
+    input_tokens: number
+    output_tokens: number
+    cache_read_input_tokens?: number
+    cache_creation_input_tokens?: number
+    model: string
+  }): void {
+    const entry: TokenUsageEntry = {
+      timestamp: Date.now(),
+      inputTokens: usage.input_tokens,
+      outputTokens: usage.output_tokens,
+      cacheReadTokens: usage.cache_read_input_tokens ?? 0,
+      cacheCreationTokens: usage.cache_creation_input_tokens ?? 0,
+      model: usage.model,
+    }
+
+    this.history.push(entry)
+
+    // Trim old entries
+    if (this.history.length > this.maxEntries) {
+      this.history = this.history.slice(-this.maxEntries)
+    }
+  }
+
+  /**
+   * Get analytics summary for all recorded usage.
+   */
+  getAnalytics(): TokenAnalytics {
+    if (this.history.length === 0) {
+      return {
+        totalRequests: 0,
+        totalInputTokens: 0,
+        totalOutputTokens: 0,
+        totalCacheRead: 0,
+        totalCacheCreation: 0,
+        averageInputPerRequest: 0,
+        averageOutputPerRequest: 0,
+        cacheHitRate: 0,
+        mostUsedModel: 'unknown',
+        requestsLastHour: 0,
+        requestsLastDay: 0,
+      }
+    }
+
+    const now = Date.now()
+    const hourAgo = now - 60 * 60 * 1000
+    const dayAgo = now - 24 * 60 * 60 * 1000
+
+    let totalInput = 0
+    let totalOutput = 0
+    let totalCacheRead = 0
+    let totalCacheCreation = 0
+    let modelCounts = new Map<string, number>()
+    let requestsLastHour = 0
+    let requestsLastDay = 0
+
+    for (const entry of this.history) {
+      totalInput += entry.inputTokens
+      totalOutput += entry.outputTokens
+      totalCacheRead += entry.cacheReadTokens
+      totalCacheCreation += entry.cacheCreationTokens
+
+      modelCounts.set(entry.model, (modelCounts.get(entry.model) ?? 0) + 1)
+
+      if (entry.timestamp >= hourAgo) requestsLastHour++
+      if (entry.timestamp >= dayAgo) requestsLastDay++
+    }
+
+    // Find most used model
+    let mostUsedModel = 'unknown'
+    let maxCount = 0
+    for (const [model, count] of modelCounts) {
+      if (count > maxCount) {
+        maxCount = count
+        mostUsedModel = model
+      }
+    }
+
+    const totalRequests = this.history.length
+    const totalCache = totalCacheRead + totalCacheCreation
+    const totalTokens = totalInput + totalOutput + totalCache
+    const cacheHitRate = totalTokens > 0 ? (totalCacheRead / totalTokens) * 100 : 0
+
+    return {
+      totalRequests,
+      totalInputTokens: totalInput,
+      totalOutputTokens: totalOutput,
+      totalCacheRead,
+      totalCacheCreation,
+      averageInputPerRequest: Math.round(totalInput / totalRequests),
+      averageOutputPerRequest: Math.round(totalOutput / totalRequests),
+      cacheHitRate: Math.round(cacheHitRate),
+      mostUsedModel,
+      requestsLastHour,
+      requestsLastDay,
+    }
+  }
+
+  /**
+   * Get recent entries within time window.
+   */
+  getRecent(windowMs: number): TokenUsageEntry[] {
+    const cutoff = Date.now() - windowMs
+    return this.history.filter(e => e.timestamp >= cutoff)
+  }
+
+  /**
+   * Clear history.
+   */
+  clear(): void {
+    this.history = []
+  }
+
+  /**
+   * Get history size.
+   */
+  get size(): number {
+    return this.history.length
+  }
+}
+
 /**
  * Get the current context window size in tokens.
  *

src/utils/urlRedaction.test.ts

@@ -0,0 +1,38 @@
+import { describe, expect, test } from 'bun:test'
+
+import { redactUrlForDisplay } from './urlRedaction.ts'
+
+describe('redactUrlForDisplay', () => {
+  test('redacts credentials and sensitive query params for valid URLs', () => {
+    const redacted = redactUrlForDisplay(
+      'http://user:pass@localhost:11434/v1?api_key=secret&foo=bar',
+    )
+
+    expect(redacted).toBe(
+      'http://redacted:redacted@localhost:11434/v1?api_key=redacted&foo=bar',
+    )
+  })
+
+  test('redacts token-like query parameter names', () => {
+    const redacted = redactUrlForDisplay(
+      'https://example.com/v1?x_access_token=abc123&model=qwen2.5-coder',
+    )
+
+    expect(redacted).toBe(
+      'https://example.com/v1?x_access_token=redacted&model=qwen2.5-coder',
+    )
+  })
+
+  test('falls back to regex redaction for malformed URLs', () => {
+    const redacted = redactUrlForDisplay(
+      '//user:pass@localhost:11434?token=abc&mode=test',
+    )
+
+    expect(redacted).toBe('//redacted@localhost:11434?token=redacted&mode=test')
+  })
+
+  test('keeps non-sensitive URLs unchanged', () => {
+    const url = 'http://localhost:11434/v1?model=llama3.1:8b'
+    expect(redactUrlForDisplay(url)).toBe(url)
+  })
+})

src/utils/urlRedaction.ts

@@ -0,0 +1,48 @@
+const SENSITIVE_URL_QUERY_PARAM_TOKENS = [
+  'api_key',
+  'apikey',
+  'key',
+  'token',
+  'access_token',
+  'refresh_token',
+  'signature',
+  'sig',
+  'secret',
+  'password',
+  'passwd',
+  'pwd',
+  'auth',
+  'authorization',
+]
+
+function shouldRedactUrlQueryParam(name: string): boolean {
+  const lower = name.toLowerCase()
+  return SENSITIVE_URL_QUERY_PARAM_TOKENS.some(token => lower.includes(token))
+}
+
+export function redactUrlForDisplay(rawUrl: string): string {
+  try {
+    const parsed = new URL(rawUrl)
+    if (parsed.username) {
+      parsed.username = 'redacted'
+    }
+    if (parsed.password) {
+      parsed.password = 'redacted'
+    }
+
+    for (const key of parsed.searchParams.keys()) {
+      if (shouldRedactUrlQueryParam(key)) {
+        parsed.searchParams.set(key, 'redacted')
+      }
+    }
+
+    return parsed.toString()
+  } catch {
+    return rawUrl
+      .replace(/\/\/[^/@\s]+(?::[^/@\s]*)?@/g, '//redacted@')
+      .replace(
+        /([?&](?:token|access_token|refresh_token|api_key|apikey|key|password|passwd|pwd|auth|authorization|signature|sig|secret)=)[^&#]*/gi,
+        '$1redacted',
+      )
+  }
+}