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

[auriti]

Summary

Fixes 3 of 5 findings from #248 — silent failure modes affecting all third-party provider users. Findings 1 and 4 were already fixed upstream.

Root Causes

Finding	Severity	Root Cause
#2 Thinking blocks as text	HIGH	convertContentBlocks() serialized thinking blocks as <thinking>...</thinking> raw text, corrupting multi-turn context on 3P providers
#3 Unknown model 200k default	HIGH	Models not in the lookup table got 200k default — auto-compact never triggered, causing hard context_window_exceeded errors
#5 Session resume thinking blocks	MEDIUM	deserializeMessages() restored thinking blocks from Anthropic sessions, causing 400 errors when resumed against 3P providers

Changes

File	Change
src/services/api/openaiShim.ts	Strip thinking/redacted_thinking blocks instead of converting to text tags
src/utils/context.ts	Use conservative 8k default for unknown 3P models + warning log
src/utils/conversationRecovery.ts	Strip thinking blocks from deserialized messages for non-Anthropic providers
src/services/api/openaiShim.test.ts	Regression test: thinking blocks must not leak as text in assistant messages

Issues addressed

• bug: REPL session management sends Anthropic-only parameters to 3P providers #248 (findings 2, 3, 5)
• Error 500 using model Devstral-Small #202 (related: Devstral Jinja template error from role alternation, compounded by thinking block leaks)

Test plan

• bun test src/services/api/openaiShim.test.ts — 5/5 pass (includes new regression test)
• bun test src/utils/context.test.ts — 6/6 pass
• Manual: resume an Anthropic session against an OpenAI provider — no 400 errors
• Manual: use an unknown model — verify 8k default + warning in stderr

[gnanam1990]

Review: Request Changes

The core fixes are correct and well-scoped — stripping thinking blocks instead of serializing them as text is the right call. A few issues need addressing before merge.

---

🔴 Issue 1 — Fragile opt-out provider check (conversationRecovery.ts)

const isThirdPartyProvider = provider !== 'firstParty' && provider !== 'bedrock' && provider !== 'vertex' && provider !== 'foundry'

This is an opt-out blocklist. Any new Anthropic-compatible provider added in the future will incorrectly strip thinking blocks by default, causing silent data loss on resume.

Suggestion: Flip to a positive allowlist:

const isThirdPartyProvider = provider === 'openai' || provider === 'gemini' || provider === 'github' || ...

Or better, add a dedicated isOpenAICompatibleProvider() utility that can be maintained in one place.

---

🔴 Issue 2 — No test coverage for stripThinkingBlocks in session resume

The conversationRecovery.ts change is the riskiest part of this PR — session resume is a complex flow and a regression here is hard to catch manually. The openaiShim.test.ts regression test is great, but this path has zero coverage.

Suggestion: Add a unit test for deserializeMessagesWithInterruptDetection (or stripThinkingBlocks directly) that verifies:

• Thinking blocks are stripped for 3P providers
• Text blocks are preserved
• Messages that become empty after stripping are dropped without breaking role alternation
• Thinking blocks are NOT stripped for firstParty/bedrock/vertex

---

🟡 Minor — console.error bypasses project logging (context.ts:84)

console.error(`[context] Warning: model "${model}" not in context window table...`)

The project uses a logError utility elsewhere. This should use the same abstraction for consistency and to ensure log output follows the same format/routing.

---

🟡 Minor — Silent message drop in stripThinkingBlocks

If an assistant message contains only thinking blocks, it is silently dropped. This could cause role-alternation issues in edge cases. Worth adding a logError or debug trace when a message is removed.

---

Overall this is a solid fix for real bugs. Addressing issues 1 and 2 before merge would make it safe to land.

[auriti]

Updated this branch on latest main and addressed the requested changes.

What changed:

• replaced the fragile opt-out provider check with a positive helper (isOpenAICompatibleProvider) so thinking stripping only applies to OpenAI-compatible providers
• added focused resume-path coverage for deserializeMessagesWithInterruptDetection(...)
  • strips thinking blocks for OpenAI-compatible providers
  • preserves visible text
  • does not apply the 3P stripping path to Anthropic-compatible providers
  • stays consistent with the existing orphan-thinking filtering behavior
• replaced the console.error warning path in context.ts with logError
• added a debug log when a resume assistant message is removed because it contains only thinking blocks

Validation run locally:

• bun test src/utils/conversationRecovery.test.ts src/services/api/openaiShim.test.ts src/utils/context.test.ts
• bun run smoke

[auriti]

Follow-up pushed for the provider-sensitive resume test placement. The latest smoke-and-tests run is now green.

[kevincodex1]

@auriti please fix conflicts bro