fix(meetings): strip the chat marker from the translation source (#6763)

[chenmingwei23]

What is the problem?

In the Meetings app, typed/chat lines are dispatched with the internal [chat] marker prepended (handle_dispatch_text builds line = f"{k.CHAT_PREFIX} {transcript_text}" for typed input) so the meeting agents can tell typed input from speech. MeetingSession.broadcast passed that same prefixed line to TranslationQueue.enqueue, so the literal [chat] marker was persisted as the translation source and interpolated into the translation prompt. Flagged as a real, observable advisory by the Opus review of #5739.

Why this issue matters to the user

The TranslationSidebar's source column shows the raw internal marker in front of every typed line, and the marker is spent as translation-prompt tokens -- the model can echo it into the translated line. Separately, the code contradicted its own documentation: the in-code comment and docs/system-specs/modules/meetings.md both said translation happens from the dictionary-corrected text, but the code enqueued the raw line.

What changed

One call site, MeetingSession.broadcast (the app's only translations.enqueue caller): the translation source now strips a leading, whitespace-bounded CHAT_PREFIX from the RAW line first, then runs the same dictionary-correct/clamp/noise pipeline the agents get (_prepare_line). Stripping before the clamp means a max-length typed message is capped on its payload rather than losing its final characters to the marker (found by the GPT review lane on the first head).

• The strip is anchored: only a bare marker or [chat] ... counts as agent context. A mid-line occurrence (quoted speech) and marker-like speech ([chat]room availability...) reach translation verbatim.
• The agent-broadcast line and the stored transcript are byte-identical to before -- the marker still reaches the agents, whose prompts rely on it.
• Correcting the clean payload also makes the code match the pre-existing comment and module-spec claim that translation happens from the corrected text. One consequence, now documented in the spec: typed filler ("ok") falls under the same noise gate as spoken filler and is dropped from the translation panel only (agents and transcript still get it).
• Spec updated in the same commit (docs/system-specs/modules/meetings.md, Live translation section).

Tests

• test_broadcast_strips_chat_prefix_from_the_translation_source: a [chat]-prefixed dispatch produces a clean translation source while the note-taker queue keeps the prefixed line. Mutation-verified: red on base behavior, green with the fix.
• test_broadcast_strips_only_the_anchored_marker: mid-line marker survives, glued marker-like speech survives, a bare marker strips to empty (dropped by the queue's blank filter, no empty translation row).
• test_broadcast_translates_the_dictionary_corrected_speech_line: a speech line's translation source is the corrected text. Also red on base (base enqueued the raw line).
• test_a_max_length_typed_line_keeps_its_full_payload_in_translation: a MAX_TRANSCRIPT_CHARS typed payload survives intact -- red on the strip-after-clamp variant.
• Local gates: isort/flake8/mypy clean on touched files; black diff-gate green; meetings suites 338 passed; full backend pytest 7490 passed with 5 failures in test_artifact_source.py proven env-owned (identical failures on clean base ba65da1e8 -- host home-dir layout).
• Pre-push review fleet: two model-pinned reviewers, both NO BLOCKING; all four Low advisories dispositioned (anchored match + negative tests folded in; truncation-at-ceiling accepted as now-consistent-with-agents; dictionary-can-mangle-the-marker is pre-existing on base for the agent line and kept out of scope).

Any other suggestions on the work

Pre-existing, out of scope here: a domain-dictionary term with alias "chat" can rewrite the marker itself (e.g. [chat] -> [Chat]) before the strip runs -- on base the same rewrite already corrupts the marker on the agent line, which is the larger half of that problem. Tracked locally for its own issue; the fix belongs at _prepare_line (correct only the payload).

Closes #6763

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of 9ca42c758095237780550060f9d24279116bd9ae — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

One anchored strip at the single enqueue seam, agent/transcript bytes untouched, spec updated in-commit — real harm, minimal and proportionate fix.

[DESIGN-REVIEWED] 9ca42c7

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed 9ca42c758095237780550060f9d24279116bd9ae — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 9ca42c7

Verdict parsed from the review's SHA-scoped output markers for commit 9ca42c758095237780550060f9d24279116bd9ae.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 9ca42c758095237780550060f9d24279116bd9ae: <one-sentence reason>

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 9ca42c758095237780550060f9d24279116bd9ae and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 9ca42c7

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 9ca42c758095237780550060f9d24279116bd9ae: <one-sentence reason>

[github-actions]

First Principles Review (Fable 5) — ✅ PASS

Premise-level review of 9ca42c758095237780550060f9d24279116bd9ae — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

All claims verified. The change is fully declared, the single translations.enqueue caller is real (count: 1), no existing strip mechanism exists elsewhere (frontend and backend greps: prepend sites only), and the deferred root cause is named in the description itself. One useful depth note: _record_line already carries is_chat out-of-band (agents.py:313, stored as TRANSCRIPT_SOURCE_TYPED), so the marker-parsing heuristic exists only because chat-ness is erased into the string between the route and broadcast — the author defers exactly this, correctly, since fixing it changes agent-visible bytes they deliberately froze.

First-Principles-Verdict: PASS

A reported defect fixed at its mechanism, in the one call site that has it, with the residual cause named and deferred rather than hidden.

What this change ships

Intent: typed meeting-chat lines stop showing the internal [chat] marker in the translation panel and its prompt — a FIX.

1. Translation source column no longer shows [chat] before typed lines — justified (defect fix(meetings): chat prefix leaks into translation source and prompt for typed lines #6763, flagged in review of feat(meetings): translate the transcript line by line #5739).
2. Speech now translates from dictionary-corrected text, as the pre-existing comment and spec already claimed — justified (documented invariant).
3. Typed filler ("ok") no longer reaches the translation panel — declared consequence, spec updated same commit.
4. A bare [chat] line yields no empty translation row — justified edge of item 1 (queue blank filter, translate.py:184).
5. Spec section updated in the same commit — mandated by AGENTS.md.

Watch

The strip is a parse-back heuristic only because chat-ness is flattened into the string at _record_line (routes/agents.py:337) — yet that same function already carries is_chat out-of-band for the transcript (TRANSCRIPT_SOURCE_TYPED, line 327), and broadcast has exactly one caller (grep \.broadcast\(: agents.py:395) which holds the boolean. The description defers this cause ("the fix belongs at _prepare_line"); when the tracked follow-up runs, threading is_chat through the one call site is smaller than it looks and retires the heuristic, the dictionary-mangles-the-marker sibling, and the agent-line clamp charge together.

[FIRST-PRINCIPLES-REVIEWED] 9ca42c7

[chenmingwei23]

Disposition of the GPT 5.6 blocking finding on 23ea2a1 (max-length typed lines lose their final seven characters in translation): FIXED in 9ca42c7, using the finding's own remedy -- the marker is now stripped from the raw line before dictionary correction and the transcript-length cap, so the cap is charged against the payload. New regression test test_a_max_length_typed_line_keeps_its_full_payload_in_translation is red on the strip-after-clamp variant and green on this head. All review lanes re-run on the new head.

[chenmingwei23]

Disposition of the First Principles Watch note (thread is_chat out-of-band through broadcast's single caller instead of the parse-back heuristic): accepted-and-deferred into the same tracked follow-up as the dictionary-mangles-the-marker class -- the FP observation that one is_chat parameter retires the heuristic, the dictionary sibling, and the agent-line clamp charge together is recorded with it. Kept out of this PR because it changes agent-visible bytes this PR deliberately freezes.

Review-ready summary for head 9ca42c7: all CI checks green (flaky test_cli_logging shard rerun green, zero overlap with this diff), CodeQL success, PR Readiness success, and all five review lanes clean on this head (GPT + Opus no blocking, Design PASS, First Principles PASS, UX pass). One fix round: GPT's blocking max-length-tail-loss finding on the first head was fixed with its own remedy (strip before correct/clamp) plus a mutation-verified ceiling test. Merge is the maintainer's call.

[bolichen97]

Tier 1 auto-approve: fix (3 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: strips the [chat] agent-context marker from the translation source before enqueue so it is not translated or surfaced in the panel source column (clear root cause). Spec files changed as a ride-along (a minority of the diff on both file count and changed lines), not reviewed as a design decision: docs/system-specs/modules/meetings.md.

[iamwhatever]

Tier 1 auto-approve: fix (3 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: single root cause -- the CHAT_PREFIX marker is agent context rather than speech, so enqueueing the raw line sent it to translation and surfaced the literal marker in the source column; the fix strips the marker from the raw line before correction and the length cap. Spec files changed as a ride-along (a minority of the diff on both file count and changed lines), not reviewed as a design decision: docs/system-specs/modules/meetings.md.

[bolichen97]

Tier 1 auto-approve: fix (3 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: clear root cause -- MeetingSession.broadcast passed the [chat]-prefixed line to TranslationQueue.enqueue, persisting the internal marker into translations; the fix strips it at the enqueue boundary. Spec files changed as a ride-along (a minority of the diff on both file count and changed lines), not reviewed as a design decision: docs/system-specs/modules/meetings.md. [Note: the auto-merge arm is denied by this cron agent's permission posture; merge deferred to a human maintainer.]