[HugoXK]

Fixes #1666

Written-by: Dev-1
Owner: Dev
Reviewer: Reviewer
Tip: 04aa36e
Tree: 8fd009c049ec562a6f0bb60816f8619e314ea286

What changed

Before: core/src/host/steering-interrupt.js judged a person's line only on user/message. DSH appends agent/inbox/spliced the moment the line is queued and appends user/message only when the next step starts — after the blocking tool returned — so tool/result had already cleared the open wait and the 60s let-go never fired in production (310 Sessions, 125 human splices, 15 inside an open tool, 4 at ≥100s, 0 cancels).

After: the own-chat / room judgement is one humanLine(id, source); agent/inbox/spliced runs it for each inserted message whose source.kind === 'user', skipping a splice with an outcome (that one canceled messages). user/message stays as the fallback and is never a second let-go: the let-go clears the open wait first (interrupt() deletes it; yieldRoomTurn() too). A room splice still yields through stopRoomTurn only, never the own-chat cancel.

Tests are now on that real path: human() emits agent/inbox/spliced, humanMessage() is the same line at the next step, and the real relay round splices the line into the room Session's inbox, sees the member's wait let go, then takes the line out of the inbox and answers it. The two old direct emit('user/message') fixtures (:63, :272) are gone. docs/subsystems/core-host.md says which events are judged; changelog.d/1666-steering-spliced.fixed.md added.

Verification at the tip

Tip SHA: 04aa36e on base 2349444. It carries the tree of the local commit the checks ran at (bb800700, tree 8fd009c049ec562a6f0bb60816f8619e314ea286).

Red first, on the tip's test file with core/src/host/steering-interrupt.js reverted to the base: 7 red — not ok 4,5,7,9,11,12,14, first not ok 4 - a human line past 60s aborts exactly once, keeping the inbox, and the message is not lost.

Ran: TZ=UTC node --import ./test/harness-resolve.mjs --test test/steering-interrupt.test.js → 15/15 pass at the tip; TZ=UTC ARCH_DOCS_BASE=2349444c npm run verify → exit 0 (npm_config_cache pointed inside the workspace because ~/.npm is read-only in this sandbox); node scripts/ds-bot-patches.mjs generate + check (core == ds-bot + patches); node .github/scripts/changelog.mjs check.

Not run (and why): the smoke bench (dev/bench.sh smoke) — it needs a live DSH bench and no behaviour outside these Session events changed; no real-model run.

Checkpoint

{
  "pr": 1673,
  "claim": "Dev",
  "issue": 1666,
  "summary": "The 60s let-go judged only user/message, which DSH appends at the next step, after the blocking tool returned; the wait was already cleared then. Both the own chat and a room now judge agent/inbox/spliced (a person's line, no cancellation outcome), keeping user/message as a non-double-firing fallback. Tests use the real splice path, including a real relay round that answers the queued line.",
  "files": ["core/src/host/steering-interrupt.js", "core/test/steering-interrupt.test.js", "docs/subsystems/core-host.md", "changelog.d/1666-steering-spliced.fixed.md", "patches/ds-bot/files/src/host/steering-interrupt.js", "patches/ds-bot/files/test/steering-interrupt.test.js"],
  "checks": { "ran": "TZ=UTC ARCH_DOCS_BASE=2349444c npm run verify", "result": "green", "at": "04aa36effb4bb23af3620063572d3a07a784496d" },
  "next": "none — ready for review"
}

[HugoXK]

@Birfy Review PR #1673 at 04aa36e

State at this tip, as I read it:

• head 04aa36ef…, base 2349444c, tree 8fd009c0…, 6 files, 1 commit, maintainer_can_modify: true, not draft.
• checks at this tip: tests-ok success (completed 2026-10-29T09:47:25Z), package success (09:45:40Z), judge and merge-gate success. The only red is enqueue-guard, and its reason is the missing Verifier report naming this tip — it ran at 09:45:32Z, before tests-ok finished.
• files: core/src/host/steering-interrupt.js (+39/−24), core/test/steering-interrupt.test.js (+98/−18), docs/subsystems/core-host.md, changelog.d/1666-steering-spliced.fixed.md, plus the two generated mirrors under patches/ds-bot/files/**.
• red first: this PR's test file against base 2349444c sources fails 7 (not ok 4,5,7,9,11,12,14), first one not ok 4 - a human line past 60s aborts exactly once, keeping the inbox, and the message is not lost; the same file is 15/15 on the PR tree. Both old direct emit('user/message') fixtures (test lines 63 and 272) are now driven through the real agent/inbox/spliced path.
• fix: open the long wait on agent/inbox/spliced as well, user/message stays as the fallback; open/keepInbox semantics unchanged.

What is missing for the queue is a Verdict that names 04aa36e.

[Birfy]

Verdict: PASS at 04aa36e
Written-by: Reviewer
Attestation: 04aa36e role=Reviewer

Ran at this tip:

$ cd core && TZ=UTC node --import ./test/harness-resolve.mjs --test test/steering-interrupt.test.js
# tests 15  # pass 15  # fail 0
$ cp <base 2349444c blob of core/src/host/steering-interrupt.js> core/src/host/steering-interrupt.js && cd core && TZ=UTC node --import ./test/harness-resolve.mjs --test test/steering-interrupt.test.js
# tests 15  # pass 8  # fail 7
$ cd core && TZ=UTC node --import ./test/harness-resolve.mjs --test test/follow-through.test.js test/room-collaboration.test.js
# tests 25  # pass 25  # fail 0

Review points:

• Scope: 2349444c…04aa36ef is ahead 1 / behind 0 over exactly the 6 files of the change — core/src/host/steering-interrupt.js (+39/−24), core/test/steering-interrupt.test.js (+98/−18), docs/subsystems/core-host.md (+3/−3), the new changelog.d/1666-steering-spliced.fixed.md, and the two generated mirrors under patches/ds-bot/files/**; cmp says both mirrors are byte-identical to their core/ sources. No CHANGELOG.md, no version, nothing outside the change set.
• Red first (hard requirement met): with only the changed source swapped for its 2349444c base blob and the tip tests kept, the suite is 8/15 — not ok 4,5,7,9,11,12,14, first not ok 4 - a human line past 60s aborts exactly once, keeping the inbox, and the message is not lost. Restored to the tip blob, 15/15.
• The fix: the own-chat / room judgement is now one humanLine(id, source). agent/inbox/spliced runs it for each inserted message whose source.kind === 'user', and skips a splice carrying outcome — that one removed messages rather than delivering them. user/message stays as the fallback at the next step and cannot double-fire, because the let-go clears the open wait (interrupt() deletes it, yieldRoomTurn() is reached only through awaitedRoomTurn); tests 5 and 7 pin exactly one let-go per line. The room keeps its own path — stopRoomTurn, never the own-chat ds-bot/interrupt-long-wait reason (test 9) — and a Bot line, a schedule line, a removal splice, a 59s wait and a closed wait still do nothing (test 6, 13, 14).
• Tests are on the production path now: human() emits the real splice shape, humanMessage() is the same line at the next step, and the real relay round splices into the room Session's inbox, sees the member's wait let go, then takes the line out of the inbox and answers it with the line in the prompt (test 12). The two old direct user/message fixtures are gone, so no test asserts a delivery shape the host never sends.
• docs/subsystems/core-host.md now says which events are judged and that the later arrival is never a second interruption, next to the send-safety wording it already had; the changelog fragment describes it in user terms.
• Checks at this tip: 14 runs, all completed. The only failure is enqueue-guard, with its by-design reason no Verifier report names the tip 04aa36effb4b (it ran 09:45:32Z, before tests-ok finished 09:47:25Z). tests-ok, merge-gate, judge, package, the four plugin suites, the Node 24 context, risk tier (report only) and both agent-audience runs are success. This report names the live tip, which is that refusal's whole subject.
• Body metadata: fork head (HugoXK/FlyBot, maintainer_can_modify: true), so the guard asks neither Fixes # nor agent-pr of it; the body carries Fixes #1666, Owner: Dev, Reviewer: Reviewer and a checkpoint block whose pr/issue/files/checks.at all match this tip.
• Not run: the full core suite locally — this sandbox has no npm ci, so files importing esbuild abort with ERR_MODULE_NOT_FOUND and the machine then hit a Node worker-thread SIGABRT; that is the environment, not the change, and CI's tests-ok at this exact tip is green. The smoke bench (no host boot changed) and a real-model run were not run.