fix(kosong): preserve unsigned thinking in anthropic history serialization - #222
Conversation
…ation When converting assistant history to the Anthropic wire format, convertMessage() dropped any thinking block that had no signature. That was meant to satisfy api.anthropic.com (which requires a valid signature on thinking blocks), but it broke Anthropic-compatible backends. Kimi's Anthropic-protocol endpoint streams thinking without a signature_delta, yet requires the thinking to be present on a tool-call turn — once it was dropped, the next request failed with "thinking is enabled but reasoning_content is missing in assistant tool call message", making multi-step tool use unusable on those backends. Preserve unsigned thinking instead, emitting it without a `signature` field. The two backends are partitioned by signature presence: api.anthropic.com always supplies a signature (its history takes the signed branch unchanged), while Kimi never does (its thinking is now kept). Empty-and-unsigned parts carry nothing and are still skipped.
🦋 Changeset detectedLatest commit: f50169b The changes in this PR will be included in the next version bump. This PR includes changesets to release 1 package
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
commit: |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f50169b367
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| } else if (part.think !== '') { | ||
| blocks.push({ type: 'thinking', thinking: part.think } as unknown as ThinkingBlockParam); |
There was a problem hiding this comment.
When this provider targets api.anthropic.com and the caller supplies history containing an unsigned ThinkPart (for example after restoring or switching from an Anthropic-compatible backend that does not emit signature_delta), this now serializes a thinking block with no signature. Anthropic's extended-thinking flow returns a signature/signature_delta and requires the complete original thinking block to be passed back unchanged (https://platform.claude.com/docs/en/build-with-claude/extended-thinking), so these requests can be rejected where the old code skipped the unsigned block. Please keep the unsigned preservation behind a backend/capability check or otherwise avoid sending unsigned thinking to the official Anthropic API.
Useful? React with 👍 / 👎.
1 participant
Problem
Multi-step tool use was broken against Anthropic-compatible backends (e.g. Kimi's
Anthropic-protocol endpoint). After the first tool call, the follow-up request failed with:
The Anthropic provider's
convertMessage()dropped any assistant thinking block that hadno signature. That guard was meant for api.anthropic.com (which rejects thinking blocks with
an invalid/empty signature), but it also discarded the thinking streamed by backends that
don't emit a
signature_delta. Those backends require the thinking to remain present on atool-call turn, so dropping it made every multi-step turn fail.
What changed
convertMessage()now preserves unsigned thinking instead of dropping it, emitting itwithout a
signaturefield. The two backends are cleanly partitioned by signature presence,so no backend switch is needed:
fixing the "reasoning_content is missing" failure.
Empty-and-unsigned parts carry no information and are still skipped. The streaming merge
logic (
mergeInPlace) was checked and is correct, so it is left untouched -- this is purelya history-serialization fix.
Verified against both real endpoints: the Kimi Anthropic endpoint now accepts the multi-step
request, while api.anthropic.com behaves exactly as before (it requires a valid signature,
which it always provides). Added regression tests covering unsigned thinking both standalone
and immediately before a
tool_useblock.