feat(api): compress old tool_result content for small-context providers (#801)

* feat(api): compress old tool_result content for small-context providers

Adds a shim-layer pass that tiers tool_result content by age on
providers
  with small effective context windows (Copilot gpt-4o 128k, Mistral,
  Ollama). Recent turns remain full; mid-tier results are truncated to
2k
  chars; older results are replaced with a stub that preserves tool name
  and arguments so the model can re-invoke if needed.

  Tier sizes auto-tune via getEffectiveContextWindowSize, same
calculation
  used by auto-compact. Reuses COMPACTABLE_TOOLS and
  TOOL_RESULT_CLEARED_MESSAGE to complement (not duplicate)
microCompact.
  Configurable via /config toolHistoryCompressionEnabled.

  Addresses active-session context accumulation on Copilot where
  microCompact's time-based trigger never fires, which surfaces as
  "tools appearing in a loop" and prompt_too_long errors after ~15
turns.

* fix: config tool history

.gitignore

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

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/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,
@@ -484,13 +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

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/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.ts

@@ -46,6 +46,7 @@ import {
   type AnthropicUsage,
   type ShimCreateParams,
 } from './codexShim.js'
+import { compressToolHistory } from './compressToolHistory.js'
 import { fetchWithProxyRetry } from './fetchWithProxyRetry.js'
 import {
   getLocalProviderRetryBaseUrls,
@@ -1299,14 +1300,15 @@ 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,

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/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',