fix(responses): carry server-tool work through the image loop

[orangeboyChen]

Fixes four logic bugs in the image_generation_call path added by #156. All were independently reproduced with probe tests before fixing, and each fix has a regression test that I confirmed fails without the change.

1. Server-tool executions were dropped (the search vanished)

getServerToolExecutions reads a WeakMap keyed on Response identity (web-search-loop.ts:785), but the image loop rebuilds the response before the caller reads it. So getServerToolExecutions(imageResponse) at responses.ts always returned [].

A turn that searched and drew a picture reported the image and silently dropped the web_search_call. The search ran and was billed; the client was never told. Measured 1 execution in → 0 out.

Fix: executions are read off the upstream response inside the loop and carried on ImageGenerationLoopResult.serverToolExecutions.

2. Rebuilt responses kept the upstream's framing headers

rebuildResponse copied response.headers verbatim while swapping in a re-serialized body of a different length. Measured: content-length: 51 on a 217-byte body, plus content-encoding: gzip on plaintext. The former truncates, the latter makes the client try to decompress plaintext.

Fix: drop content-length, content-encoding, and transfer-encoding.

3. Cap-exit reported a dangling call and repeated the prose

When the 3-iteration cap ends the loop, the closing hop's calls have already been executed. The payload still carried them, so the client got both an image_generation_call and a function_call for the same work — and the closing hop's prose was appended twice ("prose 3\n\nprose 3").

That dangling tool_call also persisted into the stored transcript with no matching tool message, which most upstreams reject on the next turn.

Fix: clearClosingHop empties the closing hop before folding.

4. Buffered streaming emitted no text deltas

The replay shipped text only inside response.completed. A client that renders as it reads subscribes to output_text.delta and saw nothing until the turn ended.

This was broader than it looks: buffering is decided by the tool declaration, not by whether the model used it. Verified — a plain "tell me about cats" with the tool merely declared also lost its deltas.

Fix: the replay emits the message-item sequence the live path does (output_item.added → output_text.delta → output_text.done → output_item.done). No content_part events, matching the live path.

Also removes the no-op prepared.defaults.tools = prepared.defaults.tools.

Verification

• 730 → 738 tests; full suite green
• tsc --noEmit and eslint --max-warnings=0 clean
• Each new test confirmed failing on the pre-fix code (stash the two source files → 5 failures, restore → all pass)

Not changed

Two findings from the earlier review I deliberately left alone:

• revised_prompt carries the original prompt rather than a model-revised one. The value is still useful and renaming it would break clients for no gain.
• A URL-only result reports completed with result: null, so the image is unreachable. Fixing it needs a non-standard field on image_generation_call, and existing tests pin the current shape — better as its own change.

Both are worth tracking separately if you want them.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 838218e150

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".

[codecov]

Codecov Report

❌ Patch coverage is 96.07843% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.57%. Comparing base (ce18a85) to head (21d2b8f).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #159      +/-   ##
==========================================
+ Coverage   95.48%   95.57%   +0.08%     
==========================================
  Files          37       37              
  Lines        6650     6738      +88     
  Branches     1912     1944      +32     
==========================================
+ Hits         6350     6440      +90     
+ Misses        300      298       -2     

Flag	Coverage Δ
unittests	95.57% <96.07%> (+0.08%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.