fix(server): plain-text call:verb spans must survive emit_finish malformed-parse + responses .done

[easel]

Summary

Two test_server_unit failures (PR Luce-Org#285's feat/lucebox-docker) traced back to commit acf718b ("detect call:verb{} in streaming emitter"). That commit added Pattern B to find_tool_start() so plain-text call:<verb>{ openers route into StreamMode::TOOL_BUFFER, but two downstream behaviors in emit_finish() were never updated for the plain-text case. The failures surfaced once CI re-enabled the suite — pre-existing feat bug, not a recent-merge regression.

Failing tests:

• test_emitter_content_mode_malformed_call_dropped (test_server_unit.cpp:1029)
• test_emitter_content_mode_responses_done_uses_pre_strip_text (test_server_unit.cpp:1157)

Pattern A vs Pattern B distinction

find_tool_start() now reports is_plain_text (out-param). The CONTENT state machine records it on the TOOL_BUFFER transition as tool_open_is_plain_text_. At emit_finish():

• Pattern A (XML envelope: <tool_call> / <function= / <tool_code>):
  • malformed parse → drop buffer (protocol artifact, not prose; test_emitter_does_not_leak_malformed_tool_xml stays green).
  • responses_streamed_text excludes tool_buffer_.
• Pattern B (plain-text call:<verb>{ opener):
  • malformed parse → flush tool_buffer_ back to accumulated_content_ AND emit it as a content delta. The literal span stays caller-visible.
  • responses_streamed_text = accumulated_content_ + tool_buffer_ so .done payloads carry the raw call span.

Intentional accessor divergence

For a successful Pattern B hoist, two accessors now diverge by design:

• accumulated_text() → stripped (call span replaced by cleaned_text). Used by OpenAI Chat / Anthropic non-streaming response builders that would otherwise duplicate the call as both literal text AND a tool_use block. test_emitter_content_mode_strips_call_span_from_accumulated_text requires this stripped form.
• responses_streamed_text (local to emit_finish) → pre-strip (includes the raw call: text). Used only inside the RESPONSES branch for .done / .completed events so streaming clients' accumulated .delta buffer matches the server's final payload.

Files changed

• server/src/server/sse_emitter.h: adds tool_open_is_plain_text_ member (+15 lines, mostly docstring)
• server/src/server/sse_emitter.cpp: extends find_tool_start() signature, sets the flag at the TOOL_BUFFER transition, splits the malformed-parse branch on the flag, conditionally appends tool_buffer_ to responses_streamed_text (+68/-16)

Test plan

• Reproduce: ./test_server_unit reports 2001 assertions / 2 failures on feat/lucebox-docker (e63f4e0).
• After fix: ./test_server_unit reports 2001 assertions / 0 failures.
• Pattern A malformed XML still suppressed: test_emitter_does_not_leak_malformed_tool_xml passes.
• Pattern B success path still strips accumulated_text(): test_emitter_content_mode_strips_call_span_from_accumulated_text passes.
• Pattern B malformed flush: test_emitter_content_mode_malformed_call_dropped passes.
• Responses .done carries raw call span: test_emitter_content_mode_responses_done_uses_pre_strip_text passes.

Origin of regression: acf718b.

Built with lucebox-hub:build-env container (matches PR Luce-Org#329/Luce-Org#331/Luce-Org#326 pattern); CPU-only — no GPU needed to exercise the host-side emitter.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c85c3b5034

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".

[chatgpt-codex-connector]

Avoid adding unstreamed tool buffer to Responses text

When a plain-text call: opener is detected before it has been flushed past the holdback (for example Looking up: + call:get_weather{...}), the emitter switches to TOOL_BUFFER, so the raw call span is never sent as response.output_text.delta; on successful parse, only parsed.cleaned_text is emitted as content before the final events. Appending tool_buffer_ here therefore makes response.output_text.done / response.completed.output_text include text that streaming clients never received in deltas, recreating the mismatch this snapshot is meant to avoid for the TOOL_BUFFER path.

Useful? React with 👍 / 👎.