🤖 fix: drop orphaned tool calls before sending

[ammar-agent]

Summary

• drop orphaned tool-call/tool-result pairs before provider transforms
• wire the self-healing pass into the model message pipeline
• add unit coverage for the orphaned tool cleanup

Rationale

• providers (e.g., OpenAI Responses) reject tool calls without matching outputs; interrupted or corrupted histories can leave orphans that brick retries
• request-time pruning unblocks sends without mutating persisted history

Why ignoreIncompleteToolCalls is insufficient

• ignoreIncompleteToolCalls only removes explicitly incomplete tool-call parts (e.g., input-available during streaming) during conversion
• the failures here come from completed tool-call parts whose matching tool-result messages are missing after crashes/truncation
• the SDK does not validate cross-message pairing, so the orphan survives conversion and only fails at provider validation time

Regression Risk

• Low: only removes unmatched tool-call/tool-result parts; valid pairs (including tool results embedded in assistant messages) are preserved
• Worst case, we hide a missing result, but that request would be invalid anyway for strict providers

Testing

• bun test src/browser/utils/messages/modelMessageTransform.test.ts
• make typecheck
• make static-check (fails locally: zizmor 403 from GitHub advisories)

---

Generated with mux • Model: openai:gpt-5.2-codex • Thinking: xhigh • Cost: $6.19

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 95e69119fb

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ammar-agent]

@codex review

Updated to treat tool results embedded in assistant messages as valid pairings and to drop orphan tool results in assistant content. Added coverage for both cases.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".