fix: strip Anthropic-specific params from 3P provider paths

Three silent failure modes affecting all third-party provider users:

1. Thinking blocks serialized as <thinking> text corrupt multi-turn
   context — strip them instead of converting to raw text tags.

2. Unknown models fall through to 200k context window default, so
   auto-compact never triggers — use conservative 8k for unknown
   3P models with a warning log.

3. Session resume with thinking blocks causes 400 or context corruption
   on 3P providers — strip thinking/redacted_thinking content blocks
   from deserialized messages when resuming against a non-Anthropic
   provider.

Addresses findings 2, 3, and 5 from #248.

src/services/api/openaiShim.test.ts

@@ -313,6 +313,57 @@ test('preserves Gemini tool call extra_content from streaming chunks', async ()
   })
 })
 
+test('strips thinking blocks from assistant messages instead of leaking them as text', 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: 10, completion_tokens: 1, total_tokens: 11 },
+      }),
+      { headers: { 'Content-Type': 'application/json' } },
+    )
+  }) as FetchType
+
+  const client = createOpenAIShimClient({}) as OpenAIShimClient
+
+  await client.beta.messages.create({
+    model: 'gpt-4o',
+    system: 'test',
+    messages: [
+      { role: 'user', content: 'hello' },
+      {
+        role: 'assistant',
+        content: [
+          { type: 'thinking', thinking: 'secret reasoning' },
+          { type: 'text', text: 'visible reply' },
+        ],
+      },
+      { role: 'user', content: 'follow up' },
+    ],
+    max_tokens: 64,
+    stream: false,
+  })
+
+  const msgs = requestBody?.messages as Array<{ role: string; content: string }>
+  const assistantMsg = msgs.find(m => m.role === 'assistant')
+
+  // The assistant message should contain only the visible text,
+  // not <thinking>secret reasoning</thinking>
+  expect(assistantMsg?.content).toBe('visible reply')
+  expect(assistantMsg?.content).not.toContain('thinking')
+})
+
 test('sanitizes malformed MCP tool schemas before sending them to OpenAI', async () => {
   let requestBody: Record<string, unknown> | undefined
 

src/services/api/openaiShim.ts

@@ -139,10 +139,12 @@ function convertContentBlocks(
         // handled separately
         break
       case 'thinking':
-        // Append thinking as text with a marker for models that support reasoning
-        if (block.thinking) {
-          parts.push({ type: 'text', text: `<thinking>${block.thinking}</thinking>` })
-        }
+      case 'redacted_thinking':
+        // Strip thinking blocks for OpenAI-compatible providers.
+        // These are Anthropic-specific content types that 3P providers
+        // don't understand. Serializing them as <thinking> text corrupts
+        // multi-turn context: the model sees the tags as part of its
+        // previous reply and may mimic or misattribute them.
         break
       default:
         if (block.text) {

src/utils/context.ts

@@ -72,16 +72,23 @@ export function getContextWindowForModel(
     return 1_000_000
   }
 
-  // OpenAI-compatible provider — use known context windows for the model
-  if (
+  // OpenAI-compatible provider — use known context windows for the model.
+  // Unknown models get a conservative 8k default so auto-compact triggers
+  // before hitting a hard context_window_exceeded error (issue #248 finding 3).
+  const isOpenAIProvider =
     isEnvTruthy(process.env.CLAUDE_CODE_USE_OPENAI) ||
     isEnvTruthy(process.env.CLAUDE_CODE_USE_GEMINI) ||
     isEnvTruthy(process.env.CLAUDE_CODE_USE_GITHUB)
-  ) {
+  if (isOpenAIProvider) {
     const openaiWindow = getOpenAIContextWindow(model)
     if (openaiWindow !== undefined) {
       return openaiWindow
     }
+    console.error(
+      `[context] Warning: model "${model}" not in context window table — using conservative 8k default. ` +
+      'Add it to src/utils/model/openaiContextWindows.ts for accurate compaction.',
+    )
+    return 8_000
   }
 
   const cap = getModelCapability(model)

src/utils/conversationRecovery.ts

@@ -24,6 +24,7 @@ import {
   type FileHistorySnapshot,
 } from './fileHistory.js'
 import { logError } from './log.js'
+import { getAPIProvider } from './model/providers.js'
 import {
   createAssistantMessage,
   createUserMessage,
@@ -145,6 +146,25 @@ export type DeserializeResult = {
   turnInterruptionState: TurnInterruptionState
 }
 
+/**
+ * Remove thinking/redacted_thinking content blocks from assistant messages.
+ * Messages that become empty after stripping are removed entirely.
+ */
+function stripThinkingBlocks(messages: NormalizedMessage[]): NormalizedMessage[] {
+  return messages.reduce<NormalizedMessage[]>((acc, msg) => {
+    if (msg.type !== 'assistant' || !Array.isArray(msg.message?.content)) {
+      acc.push(msg)
+      return acc
+    }
+    const filtered = msg.message.content.filter(
+      (block: { type?: string }) => block.type !== 'thinking' && block.type !== 'redacted_thinking',
+    )
+    if (filtered.length === 0) return acc
+    acc.push({ ...msg, message: { ...msg.message, content: filtered } })
+    return acc
+  }, [])
+}
+
 /**
  * Deserializes messages from a log file into the format expected by the REPL.
  * Filters unresolved tool uses, orphaned thinking messages, and appends a
@@ -195,10 +215,19 @@ export function deserializeMessagesWithInterruptDetection(
       filteredToolUses,
     ) as NormalizedMessage[]
 
+    // Strip thinking/redacted_thinking content blocks from assistant messages
+    // when resuming against a 3P provider. These Anthropic-specific blocks cause
+    // 400 errors or context corruption on OpenAI-compatible providers (issue #248 finding 5).
+    const provider = getAPIProvider()
+    const isThirdPartyProvider = provider !== 'firstParty' && provider !== 'bedrock' && provider !== 'vertex' && provider !== 'foundry'
+    const thinkingStripped = isThirdPartyProvider
+      ? stripThinkingBlocks(filteredThinking)
+      : filteredThinking
+
     // Filter out assistant messages with only whitespace text content.
     // This can happen when model outputs "\n\n" before thinking, user cancels mid-stream.
     const filteredMessages = filterWhitespaceOnlyAssistantMessages(
-      filteredThinking,
+      thinkingStripped,
     ) as NormalizedMessage[]
 
     const internalState = detectTurnInterruption(filteredMessages)