Filter interrupted recovery control text from visible transcript

[franksong2702]

Thinking Path

• Reproduced the reported shape from session e30f2cc1efe0: interrupted SSE recovery text can be replayed into the visible transcript instead of staying platform-only control state.
• Traced the lifecycle from backend synthetic run-journal payload (stale_interrupted_event) through static/messages.js SSE settle/error handlers and static/ui.js final transcript filtering.
• Tightened the first draft after review: the original guard only covered assistant content helpers, so the generic msgContent(m) path could still render the recovery text; it also missed the system-like/user-message leak shape.

Contract Routing

• Contract area: streaming / SSE recovery / transcript rendering.
• Mutated state layers: durable run journal payload metadata, frontend S.messages during done / apperror / _restoreSettledSession, and render-time visible transcript indexing.
• Invariant: recovery-control instructions are never user-visible transcript content, while real user/assistant text, tool calls, and settled transcript restoration remain replayable and idempotent.

What Changed

• api/run_journal.py
  • Marked stale_interrupted_event synthetic interrupted payloads with recovery_control: True.
• static/messages.js
  • Added strict recovery-control detection helpers.
  • Filters recovery-control rows from settled frontend state after done, recovery-control apperror, and _restoreSettledSession.
  • Treats backend recovery-control apperror as a restore signal instead of appending a visible synthetic assistant error.
• static/ui.js
  • Filters recovery-control messages before both visible transcript filtering and cached visible-index rebuild.
  • Handles system-like/user-message recovery prompts as control text while avoiding broad matching on ordinary user phrases.
• tests/test_live_stream_ux.py and tests/test_session_rotate_url_sync.py
  • Added regression assertions for backend tagging, frontend settle filtering, UI visible-index filtering, and non-broad phrase matching.
  • Updated the rotate-url static assertion window so the added settle filter does not create a false negative.

Why It Matters

• Prevents platform-injected recovery instructions from being presented as user or assistant transcript text.
• Keeps done, restore, and recovery-control apperror paths aligned with the same transcript visibility invariant.

Verification

• python -m pytest tests/test_live_stream_ux.py -q
• python -m pytest tests/test_live_stream_ux.py tests/test_static_js_runtime_lint.py tests/test_cancelled_turn_status.py -q
• python -m pytest tests/test_run_journal.py tests/test_run_journal_streaming_static.py tests/test_run_journal_routes.py tests/test_run_journal_frontend_static.py tests/test_issue765_streaming_persistence.py tests/test_stale_stream_pending_recovery.py tests/test_stale_stream_writeback.py -q
• python -m pytest tests/test_session_rotate_url_sync.py tests/test_live_stream_ux.py -q
• python -m pytest tests/test_live_stream_ux.py tests/test_session_rotate_url_sync.py tests/test_sprint9.py::test_no_duplicate_function_definitions -q
• python -m pytest tests/ -q --timeout=60 --shard-id=2 --num-shards=3
• git diff --check

Risks / Follow-ups

• Manual SSE reconnect validation is still useful in a stable UI session to confirm the user-facing status wording around restore.
• If future providers introduce new recovery-control text variants, they should be server-classified with recovery_control: True rather than added as broad frontend phrase matches.

AI usage

• Provider: OpenAI
• Model: GPT-5 via Codex
• Notable tools: local code inspection, targeted patching, pytest execution, GitHub CLI.

[franksong2702]

Cross reference: root cause and repro details tracked in issue #3320.

[nesquena-hermes]

Shipped in v0.51.202 (stage-batch14, via #3358). Thanks @franksong2702 — the recovery_control server marker + the narrow anchored-text fallback is the right approach for keeping interrupted-recovery control text out of the transcript.

One maintainer fix on the way in (flagged by both the Codex and Opus pre-release gates): _isRecoveryControlMessage / _streamRecoveryControlMessage also fell back to matching provider_details_label === 'interruption details'. But a genuine "Response interrupted" card (Stop button, real provider crash) carries that exact label — so on the next render/restore the filter would have dropped a real user-facing interruption from the transcript (the inverse of the data-loss class we just held #3300 for). I removed the provider_details_label heuristic from both predicates, keeping the explicit recovery_control marker + the two fully-anchored synthetic-text matches (pre-marker backward-compat). Added a node-driven regression test (revert-verified) asserting a label-only interruption card and a real user turn stay visible while marker + strict text are filtered. Co-authored-by preserved. 🎉