fix(openai): replay reasoning signatures on function calls

[justrach]

Some providers that serve reasoning models over the Responses protocol return the model's opaque reasoning state as encrypted_content on function_call output items, not only on reasoning items. A stream can send this value only on response.output_item.done, after response.function_call_arguments.done. In a stateless (store: false) tool loop, the client must send the value back unchanged on the next request. Before this change it was lost in four places:

• The item schema removed the field during decoding.
• Streamed calls were emitted before the completed item arrived.
• The tool-call part metadata did not keep the value.
• The replayed function_call input item did not include it.

Without the value, the model can lose its reasoning state between turns, and a later turn can complete with an empty output.

Changes:

• OpenAiSchema: the function_call item accepts an optional encrypted_content (string or null). One schema covers request input and response output, so the field survives response decoding, stream event decoding and request encoding.
• OpenAiLanguageModel: generated responses keep the value as metadata.openai.encryptedContent on the tool-call part. Reasoning parts already use this metadata key.
• OpenAiLanguageModel: streaming records the arguments when response.function_call_arguments.done arrives. It emits the call when response.output_item.done arrives, with that item's encrypted_content. The value from response.output_item.added is the fallback.
• OpenAiLanguageModel: if a call's item never completes, the call is emitted at response.completed, response.incomplete or response.failed. These calls are emitted in output order, before the finish part. Without this, a stream that never completes the item would lose calls that the old argument-completion path emitted.
• OpenAiLanguageModel: prepareMessages sends encryptedContent from tool-call part options as encrypted_content on the function_call input item. Reasoning items keep their existing replay and stay in the received order relative to function calls.
• ToolCallPartOptions / ToolCallPartMetadata: declare encryptedContent?: string | null. The OpenAI-compatible package declares the same field. Both packages augment the same Effect module interfaces, so their declarations must stay identical.
• CLAVIA_PATCHES.md: add a "Function call signatures" row and section.

An explicit include: ["reasoning.encrypted_content"] already reaches the request for model ids that are not in the built-in reasoning-model list. The new tests check this together with the replay. OpenAI requests are unchanged:

• Function calls without encrypted_content replay without the key.
• Inferred include values are unchanged.
• In a normal stream, the call is emitted when the next event (item completion) arrives instead of on argument completion. The emitted parts and their order are unchanged.

Verification:

• pnpm test --run packages/ai/clavia/test/LanguageModel.test.ts packages/ai/clavia/test/ProviderStreams.test.ts packages/ai/openai/test/OpenAiLanguageModel.test.ts packages/ai/anthropic/test/AnthropicLanguageModel.test.ts packages/ai/openai-compat/test/OpenAiLanguageModel.test.ts: 5 files, 202 tests passed.
• pnpm test --run packages/ai/openai/test packages/ai/openai-compat/test packages/ai/clavia/test: 10 files, 255 tests passed.
• New tests:
  • Non-streamed response for a model id outside the reasoning-model list, with function call and reasoning signatures. The test checks the explicit include, the tool-call metadata, and that the replayed reasoning and function_call items carry their encrypted content in order.
  • Unchanged OpenAI requests for gpt-4o-mini and gpt-5 with store: false.
  • Streamed SSE through the real client, with interleaved reasoning and function call items. One call's signature is only on the completed item, which arrives after argument completion. After a JSON round trip of the prompt, all items replay in received order.
  • A stream with no item completion: the calls are emitted at the terminal event, in order, with the added item's signature, before a tool-calls finish.
  • The existing tests still pass, including the argument-completion-only test and the duplicate-call fixture.
• The streamed signature test fails on the previous emission path.
• pnpm lint-fix, pnpm lint and pnpm check pass.
• node scripts/clavia-release.mjs pack and check pass. The packed consumer runs with skipLibCheck: false (6 files, 101 tests) and resolves a single effect@4.0.0-rc.115.
• pnpm jsdocs --check still reports three existing issues, all in files this PR does not touch (packages/ai/clavia/src/ToolResult.ts and packages/effect/src/unstable/ai/McpSchema.ts).

Release note: publish @tardie/ai-openai and @tardie/ai-openai-compat together, because their module augmentations must match.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-15T05:38:14.084957Z	2f6d65a	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.