fix(sse): pass Claude passthrough thinking blocks through unchanged

[havockdev]

Problem

On the Anthropic-native Claude OAuth passthrough, every multi-turn request that carries extended thinking fails with:

[400]: messages.N.content.M: `thinking` or `redacted_thinking` blocks in the latest
assistant message cannot be modified. These blocks must remain as they were in the
original response.

It is especially visible on long Claude Code tool-loops (the conversation accumulates many thinking turns, so it reproduces on almost every follow-up request).

Root cause

redactPassthroughThinkingSignatures (open-sse/handlers/chatCore.ts) rewrites every assistant thinking/redacted_thinking block to redacted_thinking{data:<synthetic>} before forwarding to Anthropic.

The Messages API validates submitted thinking blocks against the original response and rejects any modification — so converting them is exactly what triggers the cannot be modified 400.

The rewrite was added to avoid a presumed 400 "Invalid signature in thinking block" (#2454) on a token/model switch, on the assumption that the signature is bound to the auth token that produced it. In practice the thinking-block signature is validated server-side by Anthropic and stays valid on replay — including under a different OAuth token, which I verified by preserving the blocks across a mid-conversation account switch with zero Invalid signature responses. So the redaction is both unnecessary and the cause of this regression.

Fix

Pass the thinking/redacted_thinking blocks through unchanged (the function now returns messages as-is). The signature parameter is kept for call-site compatibility, and the doc-comment is updated to explain the two errors and why the redaction was removed.

Testing

• Updated tests/unit/claude-passthrough-thinking-2454.test.ts to assert the blocks pass through unchanged (historical and latest), pre-existing redacted_thinking is not re-stamped, and non-array input passes through. Run:

  node --import tsx/esm --test tests/unit/claude-passthrough-thinking-2454.test.ts
  

  → 3/3 pass.
• Validated in production: multi-turn Claude Code sessions with extended thinking (long tool-loops) no longer return the 400, including across a mid-conversation OAuth account switch with no Invalid signature.

Checklist

• Changed unit test included (tests/unit/claude-passthrough-thinking-2454.test.ts)
• prettier --check clean on changed files; eslint clean on changed files
• No hardcoded secrets / no raw error bodies touched
• CHANGELOG updated under [Unreleased]
• No Co-Authored-By trailers

[gemini-code-assist]

Code Review

This pull request stops rewriting assistant thinking blocks to redacted_thinking in the Claude OAuth passthrough, resolving a 400 error where the Messages API rejected modified thinking blocks. The redactPassthroughThinkingSignatures function is now a no-op that returns the messages unchanged, and unit tests have been updated accordingly. The reviewer suggested marking this now-obsolete function as @deprecated in its JSDoc to signal that it should be refactored out of call sites.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Since redactPassthroughThinkingSignatures is now a no-op that returns the input messages unchanged, it is highly recommended to mark this function as @deprecated in the JSDoc. This informs other developers and IDEs that the function is obsolete and should be refactored out of call sites in future cleanups.

Suggested change

	*/
	*\n * @deprecated This function is now a no-op and returns the messages unchanged.\n * Callers should be refactored to avoid calling this function.\n */

[diegosouzapw]

✅ Live validation (Rule #18) — confirmed against the real Anthropic Messages API

Validated the core claim directly against api.anthropic.com/v1/messages (model claude-sonnet-4-6, extended thinking enabled):

1. Generate a real thinking block → 200, block carries a 332-char signature.
2. Replay the thinking block VERBATIM (this PR's behavior — pass blocks through unchanged) → 200 ACCEPTED, and crucially no Invalid signature in thinking block (fix: Sanitize Claude Thinking Block Signatures & Beta Headers in Proxy Passthrough #2454). The signature validates on replay, refuting the original premise that it is auth-token-bound.
3. Replay with the block rewritten to redacted_thinking (the current release behavior, redactPassthroughThinkingSignatures → synthetic data) → 400 REJECTED: messages.N.content.0: Invalid \data` in `redacted_thinking` block`.

Result: the verbatim passthrough this PR restores is accepted; the redaction it removes is rejected by Anthropic. The fix resolves a real multi-turn extended-thinking breakage and does not reintroduce #2454.

Minor note: the observed 400 string was Invalid \data` in `redacted_thinking`rather than the exactcannot be modified wording in the CHANGELOG — same root cause (the redaction is rejected), just a different Anthropic error message for the synthetic-data` case. Optional CHANGELOG wording tweak, non-blocking.

Merging into release/v3.8.24. Thanks @havockdev! 🙌