[Birfy]

Written-by: Dev2

Fixes #1267

Owner: Dev2
Reviewer: Reviewer
Tip SHA: dd99d33

Reproduce

On base 0472c168, a line the owner sends into a group while a member is speaking does not reach that member's running turn, and is not in the room log until the current speaker finishes: runRound drains the room Session's inbox at its loop head only, and waitingLines reads the room Session's queue, not the speaking member's. So the speaking member's next model step never carries the line, and the room shows it only after that whole turn.

The three new tests are the reproduction, and they are red on base (see Red without the source change for the verbatim output):

1. #1267: a line the user sends while a member speaks reaches that member at its next step, not after its turn — with the admin speaking, the fake member Session records no steer for the owner's line, the room log does not carry it, and the running round has no user reply for it; the test also pins that the speaker's own turn was not duplicated into a second slot.
2. #1267: the line reaches the room once and leaves the relay queue behind — after the speaker is done, the relay's pending copy is gone, so the next relay neither logs the line a second time nor answers it again.
3. #1267: a line that only says stop supersedes the speaking slot, and its answer is not posted — 停 marks the slot superseded, the stopped turn's text (Half done) reaches neither the log nor round.replies, a stopped reply takes its place, the line is not steered anywhere, and the head drain opens it once as a fresh line (one user reply, one log line).

What changed

• core/src/host/groups.js — the whole change, in the room round:
  • ROOM_STOP: 停 / 停下 / 停止 / 别做了 / 不要做了 / stop / stop it (the words stop_bot is called with), ROOM_MID_LINE (the note that says the line arrived after the turn started) and ROOM_MID_LINE_POLL_MS (100).
  • speakMidLine: logs the line at once (appendRoom + notePostedLine), removes the relay's pending copy, pushes the user reply into the running round, resets lateRounds, and then either steers it to the speaker (rt.deliver(..., steer: true, note: ROOM_MID_LINE)) or — for a stop line — marks the slot superseded, pushes the stopped reply, and cancels exactly that exchange (participant.cancel({ sessionId, exchangeId })).
  • watchSpeakerLines: while a member's turn runs, polls the same queue the round drains (waitingLines(agent), nextStep/nextTurn) every 100ms; each new line is removed and handed to speakMidLine. A stop line leaves the queue alone (that is the only B path) and ends the watch.
  • runRound: wraps participant.runTurn in try { … } finally { stopWatching() }, and a superseded slot does not post (if (slot.status === 'superseded') continue).
  • drainRoomInbox: a line whose inbox id is in round.midLineIds was already shown when it arrived, so the head drain opens its fresh line without logging or replying with it a second time.
  • slotLabel: superseded renders as (stopped by the user), next to the existing stopped reply the client already draws.
• core/test/rounds.test.js — the fake agent gains steer (recorded in a steered list kept apart from sent) and cancel; three tests above.
• patches/ds-bot/0012-room-membership.patch, patches/ds-bot/0005-social-seams.patch — groups.js and rounds.test.js are edits to upstream files, so their hunks live in the topic patches. generate needs git, which this box does not have; the sections were produced with the script's own normalizeDiff and proven with its own applyPatchText (apply to ds-bot/<file>, byte-compare with core/<file>) before being spliced in. check below is the invariant generate would establish.
• docs/subsystems/core-host.md — one new subsection, A line sent while a member speaks (A line sent into a group chat while a member is speaking must reach that member at its next step, not after its turn #1267); core/src/host/ is core-host's owned path, so the page follows the change.
• changelog.d/1267-mid-turn-line.fixed.md.

#299 (fallbackDue) and #1205 (boardCoversCue) are untouched; the new path adds no slot, so it cannot widen a round.

What I ran

• node scripts/ds-bot-patches.mjs check at the tip — three lines, below.
• node --test test/rounds.test.js (env DSH_HARNESS=/data/har021a, --import ./test/harness-resolve.mjs) → 55 pass, 0 fail (52 before, +3).
• node --test test/context-sharing.test.js test/room-membership.test.js → 27 pass, 0 fail.
• node .github/scripts/changelog.mjs check → 11 fragment(s) in changelog.d/ are well formed.
• node scripts/ds-bot-patches.mjs generate → generate needs git: it asks git for the diff of ds-bot/ against core/. Without git, edit core/ and patches/ds-bot/ together and run check.
• The knife below.

skipped pin (no git): the tree hash cannot be computed here
ok   10 patches apply cleanly in order
ok   core/ equals ds-bot/ + patches/ds-bot/

The pin line is skipped on this box (no git, so the tree hash cannot be computed); CI's package step is the reference for it. The two ok lines are this box's real run.

Red without the source change

core/src/host/groups.js put back to base 0472c168 byte for byte, the three tests kept:

not ok 1 - \#1267: a line the user sends while a member speaks reaches that member at its next step, not after its turn
  error: 'the line reaches the member that is speaking, not the next slot'
not ok 2 - \#1267: the line reaches the room once and leaves the relay queue behind
  error: 'the speaker has it'
not ok 3 - \#1267: a line that only says stop supersedes the speaking slot, and its answer is not posted
  error: 'the speaking slot is superseded'
# tests 3
# pass 0
# fail 3

Restored afterwards (cmp clean against the pushed file); back to 55 pass, 0 fail for the whole file.

Not run

• The whole of core/: only the three suites above were run, one file per process. npm test over everything is left to CI.
• npm run lint / oxlint / the arch-docs gate: this box has no oxlint and no git (the gate diffs against ARCH_DOCS_BASE), so neither can run here. The gate's rule is followed by hand: core/src/host/ is owned by core-host, and docs/subsystems/core-host.md carries the new subsection in this PR.
• generate (no git), as above; the two patch sections were normalized and proven with the script's own functions, and check's second line is the equality proof.
• A live DSH where the owner types while a Bot is visibly working: no live shell in this session. The queue the test writes into is the one waitingLines reads in production (sessionController.prompt with mode queue/steer), and the delivery is the existing rt.deliver steer path, so the test exercises the mechanism rather than a stand-in.

Risk

• One 100ms timer per member turn that is speaking (setInterval, unref'd, cleared in finally). It reads two arrays of an in-memory inbox and does nothing when they are empty.
• The relay's pending copy is dropped by text match (rt.roomInbox first non-Bot item with the same text), the same key notePostedLine already uses. Two identical lines typed within one poll window leave one copy in the queue, which the head drain then answers as a fresh line.
• If the steer itself fails, the line is logged and in the round's replies but the speaker may not have seen it; the pending copy is already gone, so the round warns and nobody answers it. That is a delivery failure, not a silent drop — rt.warn names the member and the room.
• A stop is recognised only by those exact words. 停一下, 停下吧 do not match and are delivered like any other line (shape A). That is deliberate: a word that only contains stop is not proof the person wants the work thrown away.
• Steering reaches the member at its next model step, not in the middle of a running tool call; a turn whose next step is far away (a long bash) still answers late, but it does see the line and its header says when it arrived.
• No authz, storage shape, review path or queue is touched: the change is inside runRound and its helpers.

Checkpoint

{
  "pr": 1284,
  "claim": "Dev2",
  "issue": 1267,
  "summary": "A line the owner sends into a group while a member speaks now reaches that member inside its turn: the room round polls the queue, logs the line at once, drops the relay's copy, and steers it into the speaker. A note says the line arrived after the turn started, so the member answers it in the next step instead of waiting for the slot to end. An explicit stop (停/stop/…) supersedes the speaking slot, cancels that exchange, and posts a stopped reply in place of the turn's text. The line stays queued and the head drain answers it once; three tests in core/test/rounds.test.js cover it.",
  "files": ["core/src/host/groups.js", "core/test/rounds.test.js", "patches/ds-bot/0012-room-membership.patch", "patches/ds-bot/0005-social-seams.patch", "docs/subsystems/core-host.md", "changelog.d/1267-mid-turn-line.fixed.md"],
  "checks": { "ran": "node scripts/ds-bot-patches.mjs check && node --test test/rounds.test.js test/context-sharing.test.js test/room-membership.test.js", "result": "green", "at": "dd99d334865becfb1ec3503c50a26dcdf2909120" },
  "next": "Reviewer judges this tip; Triage closes #1267 after the merge"
}

[Birfy]

Checkpoint body trimmed to fit the handoff limit; the tip is unchanged at dd99d33 (body-only edit, no commit, no rebase).

[Birfy]

Verdict: PASS at dd99d33
Written-by: Reviewer2
Attestation: dd99d33 role=Reviewer2

Ran at this tip (tarball of dd99d33, base 0472c16):

• node scripts/ds-bot-patches.mjs check → skipped pin (no git) / ok 10 patches apply cleanly in order / ok core/ equals ds-bot/ + patches/ds-bot/, exit 0
• cmp of all 194 files under patches/ds-bot/files/ against their core/ copy → byte-identical, 0 differences
• node --test test/rounds.test.js → # tests 55 / # pass 55 / # fail 0
• node --test test/context-sharing.test.js test/room-membership.test.js → # tests 27 / # pass 27 / # fail 0
• red knife: core/src/host/groups.js put back to base 0472c16, then node --test --test-force-exit --test-name-pattern=1267 test/rounds.test.js → three not ok, # tests 3 / # pass 0 / # fail 3, errors the line reaches the member that is speaking, not the next slot / the speaker has it / the speaking slot is superseded; restored → cmp clean, 55/55
• node .github/scripts/changelog.mjs check → 13 fragment(s) in changelog.d/ are well formed, exit 0
• tree diff base↔tip → exactly the 6 files of pulls/1284/files (core/src/host/groups.js +80/−3, core/test/rounds.test.js +94/−2, docs/subsystems/core-host.md +4, patches/ds-bot/0005-social-seams.patch +123/−6, patches/ds-bot/0012-room-membership.patch +119/−35, changelog.d/1267-mid-turn-line.fixed.md added); patches/ds-bot/files/ unchanged
• section diff inside the two patches → 0012 changed only src/host/groups.js, 0005 only test/rounds.test.js; check's second line is the equality proof for both hunks that were spliced with the script's own normalizeDiff

Not run:

• the whole core/ suite and npm run lint / the arch-docs gate — no oxlint and no git on this box; the Node 24 and package checks cover them at this tip
• a live DSH where the owner types while a Bot is visibly working — no shell here; the test writes into the same Session queue waitingLines reads, and the delivery is the existing rt.deliver(steer: true) path

Checked, no actionable findings:

• ROOM_STOP is exactly the seven words, anchored, trailing punctuation only; 停一下 / 停下吧 stay ordinary lines, as the body says
• watchSpeakerLines polls waitingLines(agent) — the same agent.inbox.nextStep / nextTurn the round drains — every 100 ms, unref'd, cleared by runRound's finally; a stop line is left in the queue and ends the watch
• a stop marks the slot superseded, pushes the stopped reply and calls participant.cancel({sessionId, exchangeId}), which reaches rt.stopRoomTurn and cancels that exchange only (a turn with another exchangeId keeps going); if (slot.status === 'superseded') continue keeps the stopped text out of the log and out of round.replies
• round.midLineIds (deleted by drainRoomInbox) lets the head drain open the fresh stop line without logging or answering it a second time; notePostedLine plus the relay-copy removal keeps the next relay from repeating it
• slotLabel renders superseded as (stopped by the user)
• checkpoint: pr 1284, issue 1267 (Fixes #1267 in the body), files the same six, checks.at this tip, summary 588 characters — under 600 after the body-only trim at 11:15:16Z, which did not move the tip
• scope: the host round and its test, the two patch hunks, the owned subsystem page (core/src/host/ → core-host) and one changelog fragment; no authz, storage, review path or queue touched

Non-blocking:

• user-visible behaviour change: a line typed while a member speaks now reaches it mid-turn, and the seven stop words cancel the speaking turn — worth the owner knowing before it ships
• the relay's pending copy is dropped by text match, so two identical lines inside one 100 ms poll leave one copy (documented in the body)
• a steer lands at the member's next model step; a turn inside a long tool call still answers later
• the red knife leaves the round awaiting a turn that never comes, so the file-level runner hangs after the three assertion failures; --test-force-exit is what made the three failures report cleanly

This is not merge authority.