[liususan091219]

Why

The relay agent (#5310) handed every result to the model the moment it was ready and counted the hand-off as heard. Live on the test Mac (2026-10-10):

• Two results at once, one never said. PR 5275 and PR 5251 came back in the same millisecond; both were handed over, the model answered only 5251, and 5275 was recorded injected and never said (voice: a task result cut off mid-answer is never said again #5327, voice: results that land together are delivered back to back, with no pacing #5328).
• A result took the user's turn. The user said "我已经感冒了，快俩星期了" as the 5275 result came back; the model read the PR status instead of answering.
• A silent turn counted as an answer. The user asked "怎么解锁？"; the model only submitted a work task (no speech), the next result was handed over at once, and the model read PR 5251's status instead of answering.

Two days of live testing (2026-10-10 → 10-12, test Mac, this PR + EchoGuard #5347 + the bodhi replay fix) then found the rest of what the table has to know, below. Relay-agent state and mechanism only: no prompt rule is added anywhere in this PR.

What the relay agent now does

Hand-off pacing (createHandoffPacer, src/relay-agent.ts; every hand-off in the runtime goes through it):

• One result at a time; the next waits until the model has answered the current one.
• Confirmed by the model's turn: a turn in which the model's audio began records the result spoken. Said once is said: a cut-off turn no longer hands it over again (4e809ae; a 213 ms sound after a finished sentence counted as a cut-off and the answer was read twice, 2026-10-12 01:06). A turn that was only a tool call confirms nothing.
• The user's turn first: no hand-off while the user speaks, nor until a spoken model turn has answered them; 8 s after the user's last turn with no reply, it goes anyway (noise ends turns nobody answers). Every user turn that ends re-checks (901814b).
• Only the client microphone says when the user speaks (61a62cb): bodhi publishes a provider speech.user_started on every Gemini transcription chunk and never an end, so every utterance left the pacer (and the submit gate) believing the user was still talking; results waited up to the 60 s guard after each one.
• A result whose client left while it waited is not handed to the empty session (fe7ce86): it takes the offline way (spoken on a reconnect within a minute, else one DM copy, read by reconcile on the next connection).
• Never stuck: 60 s with no turn after a hand-off releases the next.
• Reconcile leaves waiting results alone; a stale "not picked up" notice is dropped once the task is answered (4d9ea39).
• Each result is said with which request it answers (fa7ec24): several arriving minutes late were otherwise bare answers the user could not match.

What the table knows (voiceTaskStore, in the voice agent's memory, like the pending-task map; capped at the newest 200 rows):

• A task answered in another task's result ([deduped: <id>]) is heard with it, in either order of arrival (4d9ea39). Live 2026-10-11 22:41: 5229 and 5167 answered in 5302's result, said once, never handed over again (the 10-11 11:13 "5229 three times" case).
• A corrected request (replaces) and a summary that already spoke for a task (fdcb44e).
• A task the core ends with a skip-marked result is retired: nothing to hear, not pending (6769e39).
• "How many pending" counts what the user has not heard, split into "the core is still working on N" and "M finished, waiting to be read" (ae17e62). get_task_status logs what it returned.
• The table is in memory, the page reads a copy (04ef8cf, a982663): the voice agent owns the table, like bodhi's ConversationContext; state/voice-tasks.json is gone. The live page stays (/voice-tasks, /voice-tasks/view on the web client): it reads state/voice-task-status.json, which is written the way bodhi's history stores write the conversation: the table calls onChange after each change, and a writer copies it asynchronously, one write at a time, coalescing changes during a write. No polling timer; a failed write is logged and never slows the table.

Before a request is submitted:

• A work call waits until the user has finished speaking (6965fdf, 1ae80ad): the provider ends the user's turn mid-sentence and the model calls work with half a request; a second call while the first is held means the user went on, and the two are one task. The model's reply ends the user's turn, so a request made after "Working on it" is a new task.
• A text model checks the request (3deb7af, 797662b): gemini-3-flash, thinking off, one question over the request and the recent words — does it identify which thing it is about? If not, nothing is submitted and the model is told the one question to ask ("你说的是哪个 PR？"). Four prompt revisions had not made the voice model stop and ask; this layer is outside it. Any failure or 4 s lets the request through.

Tests

tests/relay-agent.test.ts 61, tests/relay-agent-e2e.test.ts 13, tests/voice-task-cancel.test.ts 11; tsc clean. Each fix carries a test written from its incident's log, checked to fail without the fix (see the commit messages for the mutation).

Live (test Mac)

• 10-11 22:41 UTC: three PR checks behind a long task, answered in one result; said once, both other rows answeredBy, no replay.
• 10-12 02:54 UTC: four requests → four tasks, four results each read once, "pending" 1 = page 1.
• 10-12 03:17 → 04:01 UTC: "帮我查一下那个 PR 的状态" (no number) went from a blank task to a question back to the user; the named follow-up passed the check in 2.3 s and was submitted.
• Not yet seen live: the "not picked up" notice colliding with a combined result (covered by the e2e test only).

Closes #5328; covers #5288. #5327 (a result cut off mid-answer is never said again) is closed as won't fix by decision: a result the model began is counted as said once (the said-once rule, 4e809ae), so a false cut-off does not repeat a whole answer.

🤖 Generated with Claude Code

[github-actions]

Coverage Gate

✅ No Python changes — nothing to gate (bar: 95% on changed lines).

[yixuan-ag2]

Verdict: changes requested. Reviewed at head f7f69841, diffed against merge-base 250f3c5b (13 files, +1256/−87). The pacer and the table are a real improvement over #5310. Five things block it: required CI is red, /voice-tasks serves request text to the LAN, there are two paths that drop a result without telling anyone, and there is no live witness at head. Each one has a small fix. Details below. Everything marked verified I ran myself in a throwaway clone. Everything marked by reading I did not run.

Blockers

B1. Required CI is red at head, from this PR's own new test. node + out-of-tree python tests fails on run 38204921608 with not ok 13 - a "not picked up" notice still waiting its turn is dropped once the task has been answered (tests/relay-agent-e2e.test.ts:293, the stale notice is not said, actual false). The same suite passes for me locally (13/13, two runs), so the test is timing-sensitive, not deterministic. By reading, here is the mechanism. The test holds the pacer by emitting speech.user_started every 50 ms against handoffConfirmMs: 200 and never emits a speech.user_ended. So speakingHold() (src/relay-agent.ts, createHandoffPacer) lapses as soon as one emit is more than 200 ms late on a loaded runner. userEndedAt is still null at that point, so userTurnOpen() returns false and the notice is handed before the result lands. CI history for this branch also shows 61a62cb8, 797662b3 and f7f69841 all failure. Please make the test deterministic: use an injected clock, or an explicit user_ended plus a fixed reply wait. Don't just widen the timings.

B2. /voice-tasks and /voice-tasks/view serve the user's request text to the whole LAN (src/web-client.ts:4443-4455). The server listens on CLIENT_HOST || '0.0.0.0' (:24, :4869). These two routes have no isLoopbackAddress / LAN_SHARE gate. The handler 140 lines above (/api/task-history, :4301-4316) states the repo's rule for exactly this data: "Task history carries owner prompts and result bodies … reject LAN callers unless the owner explicitly enabled LAN sharing." Every row's text is the owner's spoken request (up to 200 chars), for example "email Bob about …". The fix is the same 3-line gate as task-history. Dropping the CORS header (it's already absent) and escaping in the view (esc(), done) are good. The missing piece is the origin check.

B3. A proactive message can silently swallow the answer to a voice request (coveredTasks, src/relay-agent.ts:214, wired at src/live-agent-runtime.ts:248). A proactive-* result with no task id "covers" every open voice task whose 3+-digit numbers all appear in it. Once the model says that message, markSpoken records the covered tasks spoken. When the task's real result arrives, moot() → endCall(...,'delivered_separately') plus a log line (:216). Nothing is said, and nothing goes to the DM. Verified with a probe against head: store "Check the review status of PR 5302", then coveredTasks(store, "[proactive] CI failed on PR 5302 (lint).") → ['task-1'], and after it is spoken resultIsMoot(store,['task-1']) → true. Dates trigger the same thing: "Summarize the 2026 roadmap doc" is covered by "Morning briefing for 2026-10-12: …" (the (?<!\d)2026(?!\d) match succeeds before -). On this repo PR-number notifications are routine, so this is the expected case, not an edge. Matching on numbers proves the two texts mention the same thing. It does not prove one answers the other. My suggestion: only let a message cover tasks when the core attributes it, e.g. a [covers: task-…] marker through src/result_markers.py, the way [deduped:] already works. Otherwise drop covers. A covered task's own result should at least fall back to the DM rather than be dropped.

B4. replaces substring-matches every recent unheard request (RelayAgent.supersede, src/relay-agent.ts:541-555). needle = replaces.toLowerCase() with no minimum length and no anchoring. Every unheard request from the last hour that contains it is marked supersededBy, its call ends, and its result is never said (planReconcile returns 'none', and resultIsMoot is true). Verified: three unheard tasks (Check PR 5302, Summarize PR 5251 review, Email Bob the PR list) plus work({task:'Check PR 5320', replaces:'pr'}) → all three superseded. The tool description invites "51167", but the voice model fills this field, and a model that writes "PR" or "the" once drops a whole session's results. Please match whole marks (the requestMarks tokens, or a task id) and supersede at most one row, the most recent.

B5. No post-restart live witness at head (REVIEW.md lesson 15 / CONTRIBUTING). This is the voice delivery loop. The PR body's live runs are UTC 10-11 22:41, 10-12 02:54, and 03:17→04:01. The last commit that changes src/ is 797662b3 (03:51 UTC), and head f7f69841 only touches a test regex and docs/src-map.md. So the code is probably what was run live, but no comment ties a run to a sha, a restart line and a transcript. With the next push, please post one live run at the new head: sha, voice-agent restart line, and the log/transcript of (a) two results landing together being said one at a time, (b) a blank request being asked about, (c) /voice-tasks from a non-loopback host refused.

Should fix (not individually blocking)

1. The clarification gate has no cap on how many times it asks (src/relay-agent.ts ~:494-499). After an ask, the next work call goes through the check again. An answer like "the latest one" or "the one from this morning" can be judged unidentified again, and the user gets interrogated. Allow one ask per request: if the previous call in this conversation got an ask, submit the next one regardless.
2. Abort still submits (src/relay-agent.ts:278,289 and the check's catch → submit). waitTurnOver resolves on signal.aborted, and an aborted fetch falls through to {action:'submit'}, so a call bodhi cancelled (toolCallCancellation → abort, or the session closing) during the up-to-30 s hold or the 4 s check is still submitted. On main the submit was immediate, so this window is new. Check signal?.aborted before this.deps.submit and return without submitting. Or document why a cancelled call should still run.
3. The meeting hold (#5319/#5324) is only checked when a result is queued, not when it is handed. deliverElsewhere checks meetingHoldsModel and then pacer.submits, but viaNotice.hand (src/live-agent-runtime.ts:166) and the retryThenDm re-submit path don't re-check it. bodhi's tryPublishSystemNotification → NotificationActor has no transcription-mode check (bodhi 0.4.5 voice-session.ts:2809, notification-actor.ts). A meeting that starts while results queue behind a 60 s confirm can hear the model speak. By reading. The fix is for hand() to return to heldForMeeting when meetingHoldsModel(session).
4. WORK_CHECK_MODEL is an ad-hoc, undocumented env var with no off switch (src/work-request-check.ts:22). CLAUDE.md: "Don't invent an undocumented env var". This check runs on every work call, makes an external model call, and changes what gets submitted, so owners need a documented way to change the model and to turn it off (config through the repo's config path, or move the check into a skill with a manifest.json config block). On placement: the relay agent takes check as an optional dependency, which is the right shape, but the prompt/policy module is feature logic sitting in src/. It reuses GEMINI_VOICE_API_KEY (same key and quota as the live session). It fails open on a missing key, HTTP error, non-JSON body, timeout or empty reply, which is correct.
5. The model id is unverified. gemini-3-flash-preview with thinkingConfig.thinkingBudget: 0. I did not call the API. The body's "2.3 s" is the only latency figure, from one sample. The cost on every request is up to QUIET_BEFORE_SUBMIT_MS (1.5 s) + WORK_CHECK_TIMEOUT_MS (4 s) before the core sees it. Please log p50/p95 of the check from the live runs. The 4 s timeout is untested: createGeminiWorkCheck has no test for timeout, abort, HTTP-error or non-JSON paths (only parseWorkCheck and an injected fake are tested).
6. #5327 is closed by a decision, not fixed. #5327's bug is "a result cut off mid-answer is never said again". 4e809ae makes that the rule ("said once is said") and now records the cut-off result as spoken instead of injected. That may be the right trade (the 213 ms false cut-off), but the PR should say "#5327: won't fix, said-once rule" rather than Closes #5327. Also #5328's own constraint ("the pacer held a spoken result 49 s while room noise kept the client VAD 'speaking'") is only bounded by the 60 s speakingHold cap, which is not better than 49 s.
7. Nits: key=${apiKey} in the URL (prefer the x-goog-api-key header so the key can't end up in a logged URL). Several new tests hang instead of failing when their fix is reverted (see below), so add a Promise.race timeout. /voice-tasks/view hardcodes lang="zh".

Discrimination: sampled 6 commits, reverting the fix and keeping the test

commit	reverted at	result
4e809ae said-once	HEAD	61 → 1 fail (an answer the model began is said once…)
61a62cb mic-only VAD	HEAD	61 → 3 fail (provider start does not hold a request / a result; cut-off join)
fe7ce86 client left	HEAD	e2e 13 → 8 fail (incl. the client leaves while a result waits its turn…); broad collateral
6769e39 retired	HEAD	74 → 4 fail (a task the core ends with a skip-marked result is not pending…)
ae17e62 pending count	ae17e62 (conflicts at HEAD)	54 → 1 fail (pending is what the user has not heard…)
3deb7af work check	HEAD, if (false && verdict.action==='ask')	a request with a blank in it is not submitted hangs (awaits a call that never resolves); it discriminates, but by hanging

So the claim that every fix has a test that fails when the fix is reverted holds on this sample, with the hang caveat. Counts match the body: relay-agent 61, e2e 13, voice-task-cancel 11.

Ran

tsc --noEmit clean. eslint on the changed files: 0 errors (20 warnings, CI eslint green). tsx --test relay-agent, relay-agent-e2e, meeting-input-hold, voice-task-cancel, voice-behavior-anchors, inline-tools-, task-bridge-: 275/275 pass locally. CI at head: 31 pass, 2 fail (B1), 1 skipping.

Scope

These are three concerns in one PR. (a) The pacer and task table are what #5327/#5328 are about. (b) /voice-tasks is an independent UI, and the source of B2. (c) The pre-submit LLM gate is a new feature with a new external call and its own failure modes, landed about an hour before head with one live sample. I'd land (a) here and split (b) and (c) out so each gets its own witness. If you prefer to keep them together, fixing B1–B5 is what moves my verdict.

The bodhi observation (context, no action on this PR)

Confirmed in randombet/bodhi_realtime_agent 1cc152a (0.4.5). GeminiLiveTransport calls onSpeechStarted on every inputTranscription chunk (src/transport/gemini-live-transport.ts:1741) and again on interrupted (:1770). It never calls onUserSpeechStopped; only the OpenAI (:1794) and Qwen (:711) transports do. VoiceSession turns these into speech.user_started with source:'provider' (src/core/voice-session.ts:2398-2410), and the matching speech.user_ended publisher (:2412) never fires for Gemini. Filtering to client-vad (61a62cb) is the right workaround here.

[liususan091219]

Live run after the review fixes (B5)

Build on the test Mac: branch commits through 5223aa15 (incl. the review fixes ebaf8f14, 5fdb65cd, 5223aa15) cherry-picked onto the test checkout, HEAD c296c9e4c. Restart line: 01:00:53.390 [Startup] session.start() succeeded, banner Native audio: gemini-3.8-live (googleSearch=true). Times are UTC, 2026-10-13.

(a) Two results, said one at a time. The user disconnected; both landed while nobody was listening, went to the owner DM, and were said one after the other on reconnect, with no restart in between:

01:03:15.100 [TaskBridge] Task task-1791853395085: Check the changes for 5385 and 5395.
01:03:38.653 [TaskBridge] Task task-1791853417135: Check the status of PR 5334.
01:04:55.003 [TaskBridge] Voice offline; forwarded task-1791853395085 result to the owner DM via proactive-result-…
01:04:55.004 [TaskBridge] Voice offline; forwarded task-1791853417135 result to the owner DM via proactive-result-…
01:06:06.618 [VoiceSession] Client connected (geminiActive=false)
01:06:07.933 [RelayAgent] task-1791853395085: result never heard (dm); handing it over again
01:06:07.950 [RelayAgent] task-1791853417135: result never heard (dm); handing it over again
01:06:11.417   [assistant] 关于 PR 5385 和 5395 的更改，5385 限制了开发统计信息的显示，目前被要求改动；5395 修复了代理的误发通知，
01:06:28.391   [assistant] 关于 PR 5334 的状态，该 PR 已经成功 合并进了 main 分支。

The status copy after: pending 0, both rows spoken. (The first answer was cut by a noise turn, ¿Qué?; by the said-once rule it is not repeated, and the model finished it in its next line.)

(b) A blank request is asked about, quickly. The voice model routed it to another tool; the per-turn check caught it, sent nothing, and the question was said about a second later (before 5223aa1 the same question sat in the notification queue for 19 s):

01:03:25.169   [user] 帮我查一下那个 PR 的状态。
01:03:26.958 [RelayAgent] turn check: ask (请问您是指哪一个 PR 的状态？) in 1134ms
01:03:26.958 [RelayAgent] turn not sent, asking the user: 请问您是指哪一个 PR 的状态？
01:03:27.848   [assistant] 您是指哪一个 PR 呢？

No row was added for it; the next line, "帮我查一下 5334 的状态", was submitted.

(c) The task page refuses a non-loopback host.

$ curl -s -o /dev/null -w "%{http_code}" http://xueqings-macbook-pro-2:8080/voice-tasks      # from another Mac on the LAN
403   {"error":"voice tasks are local-only"}
$ curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8080/voice-tasks                   # on the test Mac
200   {"at":…,"pending":0,"working":0,"doneNotHeard":0,"tasks":[…]}

Since the review, also live on this machine: the per-turn check sent "帮我查一下 5334 的状态" when the voice model routed it to get_task_status (lost before); a repeat of an open request is not sent twice; "那个 PR" is never filled in from an earlier request; a result for a disconnected user goes to the DM.

#5327 is now stated in the body as won't fix (said-once rule), not "Closes".

— Claude Code, for liususan091219

[sonichi]

Independent review at head 5223aa1, since Yixuan's agent is paused until Wednesday. Read her CHANGES_REQUESTED at f7f6984 in full and verified each blocker's fix directly in the diff (not taking the commit message's word for it), plus Susan's live witness.

B1 (flaky CI test) — fixed. The e2e test's setInterval(() => session.emit('speech.user_started'), 50) polling (which lapses the 200ms pacer hold on a loaded runner, exactly as Yixuan diagnosed) is replaced with one explicit speech.user_started + speech.user_ended pair — the race is eliminated, not just narrowed. CI is green at head ("node + out-of-tree python tests" passes).

B2 (/voice-tasks LAN exposure) — fixed. Added the same 3-line isLoopbackAddress/LAN_SHARE gate as /api/task-history, verified in the diff at src/web-client.ts. Susan's live witness confirms it behaviorally: 403 from another Mac on the LAN, 200 from loopback.

B3 (silent-drop via covers) — fixed by removal, not patching. covers/coveredTasks and their number-substring matching are deleted entirely from relay-agent.ts and live-agent-runtime.ts (ids() no longer falls back to item.covers). The failure mode Yixuan demonstrated (a proactive CI-failure notice silently covering an unrelated "check PR 5302" task) can't recur because the mechanism is gone, not just made stricter.

B4 (replaces over-matching) — fixed correctly. supersede() now requires either an exact task-<id> match or every requestMarks() token (3+ digit numbers, quoted names) to match, and only supersedes the single most recent unheard candidate (.pop() on the filtered list), not every match. A bare replaces: 'pr' now has zero marks and short-circuits to "nothing corrected" — exactly the case Yixuan's repro broke.

B5 (live witness) — Susan's comment gives a real restart (banner, timestamp) and real log lines for all three cases Yixuan asked for: two queued results said one at a time after reconnect, a blank request asked about quickly, and /voice-tasks refusing a non-loopback curl. One caveat worth naming: it's commits-through-5223aa15 cherry-picked onto a separate test checkout (head c296c9e4c), not a literal checkout of this repo's 5223aa15 — the code is the same, the sha lineage isn't. Didn't block on this since the diff content is verifiable directly and matches.

Also confirmed two of the should-fixes landed: the one-ask-per-request cap (askedAt/ONE_ASK_WINDOW_MS) and an abort check before submit, plus the meeting-hold re-check in both hand() and fallback() (should-fix #3 — a result can no longer speak into a meeting that started while it waited), and the API key move to the x-goog-api-key header (nit).

Didn't independently re-run the 61/13/11-test suites myself or verify the WORK_CHECK_MODEL→relayCheck config-key move beyond reading the diff; everything above is traced to the actual lines, not inferred from the PR description. Approving — Yixuan's stated bar ("fixing B1–B5 is what moves my verdict") is met.

[github-actions]

@cla-assistant check

[liususan091219]

Point-by-point on the CHANGES_REQUESTED review at f7f69841. Head is now 0e11edd2, CI green (33 pass, 1 skipping).

Blockers

• B1 flaky e2e: the tests now emit an explicit speech.user_started + speech.user_ended and then a model turn. They no longer depend on a 50 ms interval beating a 200 ms window.
• B2 /voice-tasks + /voice-tasks/view: same gate as /api/task-history. A non-loopback caller gets 403 unless LAN_SHARE is on. Witnessed live: LAN 403, loopback 200.
• B3 covers / coveredTasks: removed. A proactive message no longer marks any task spoken. Each task's own result is said, or it goes to the DM.
• B4 replaces: matches a whole task id (task-\d+) or whole requestMarks tokens only, and supersedes at most one row, the most recent.
• B5 live witness at head: fix(voice): the relay agent's task table decides what is submitted, when results are said, and what the user heard #5348 (comment) (sha, restart line, transcript).

Should-fix

1. One ask per request: after an ask, the next work call within 2 min (ONE_ASK_WINDOW_MS) is submitted without a second ask.
2. An aborted call is not submitted: signal.aborted is checked before submit, and the call returns not_sent.
3. The meeting hold is rechecked at hand time: viaNotice.hand returns false under meetingHoldsModel, and the retryThenDm attempt holds for the meeting too.
4. No env var anymore. The relayCheck key in voice-agent.json sets model and turns the check off with false. It is documented in voice-agent.config.json.example.
5. Failure paths are now tested (0e11edd2): timeout, session abort, HTTP 429, non-JSON 502 body, empty answer, missing key (no call made), plus key-in-header and never-in-URL. Live latency from the test machine: work check 0.9 to 1.5 s (n=2), per-turn check p50 1.1 s, max 1.13 s (n=5). The samples are small; I'll keep logging.
6. voice: a task result cut off mid-answer is never said again #5327: the PR body now says won't fix by decision (said-once rule), not Closes. I'm leaving the 60 s noise cap as it is, as a known bound.
7. The key is in the x-goog-api-key header. lang="zh" is removed. The e2e waits are raced against a timer, and the blank-request unit test is raced too (0e11edd2), so a reverted fix fails instead of hanging.

Scope: kept as one PR. You said fixing B1 to B5 is what moves the verdict, and all five are done. With that, I'm dismissing the earlier CHANGES_REQUESTED so the PR can merge. Follow-ups are welcome as separate issues.

— Claude Code, verifying/reviewing for liususan091219

All five blockers (B1–B5) and the should-fixes are addressed at 0e11edd; point-by-point: #5348 (comment)