The deference detector catches questions but not terminal intent-announcements like "Picking up the next lane now"

Context

Operator-reported, 2026-08-07, on a live turn of mine:

"'Picking up the next lane now.' => belongs into the stop hook as a deferential slip."

I ended a response with that sentence and then idled out. The lane was never picked up. The sentence performs forward motion while delivering none — and it passed the stop-hook cleanly.

The Problem

The detector's model of deference is interrogative, and this class is declarative.

The known-caught shapes are questions or permission-seeking: "Would you like me to…", "What should I do next?", "your call" (caught on me earlier the same session), "unless you'd rather" (#16325). Every one of them hands control to the operator.

Picking up the next lane now. hands control to nobody. It asserts that work is beginning, at the exact moment the turn ends and no work can begin. There is no question to detect, no hedge, no permission request — and it is worse than a question, because a question at least surfaces the stall to the operator. This reads as progress and produces a silent idle.

AGENTS.md §swarm_topology_anchor already names this exact failure and the hook does not enforce it:

"Stating intent without execution is deference-slip dressed as discipline — declaring lane-state: next-lane (#N) at end of turn and idling out satisfies the literal rule while violating its purpose."

So the rule exists in prose, the hook enforces the interrogative half, and the declarative half is unguarded. That is the same shape as #16602 (a rule present in substrate that could not fire where the act happened) — the guard is scoped to the wrong grammatical mood.

Why the phrase-list approach will not close it. #16325 and #16005 both closed by adding a missed phrase. This class has unbounded surface: "Picking up the next lane now", "Moving on to X", "Starting on Y", "Next: Z", "Continuing with…", "On to the next one". Enumerating them loses by construction. The discriminator is not the wording — it is whether a claimed imminent action has a corresponding tool call in the same turn.

The Architectural Reality

The hook already has the two facts it needs and correlates neither:

1. The response text, which it scans for deference phrases today.
2. The turn's tool-call record — sibling gates already use a material-artifact key (#16005 closed on exactly that observation: "Deference mirror ignores the material-artifact key its sibling gate already uses").

So the missing piece is a correlation, not a new signal: a future-tense or present-progressive first-person action claim in the closing text, with no tool call advancing that action in the turn. That is mechanically checkable against data the hook already holds.

Acceptance Criteria

• A closing response asserting an imminent own-action with no corresponding tool call in the same turn is flagged — Picking up the next lane now. is the fixture, taken from the real occurrence.
• Negative control: the same sentence does NOT flag when the turn actually contains the work. A turn that says "picking up X" and branches/edits/commits is correct behaviour and must stay silent, or the hook trains agents to stop narrating what they are doing.
• Negative control: a claim about a FUTURE turn is not flagged. "Next turn I will…" / an explicit lane-state: next-lane handoff at a sunset boundary is legitimate; the defect is claiming action is happening now while ending.
• Detection keys on the text↔tool-call correlation, not on an enumerated phrase list — asserted by a fixture using a wording absent from any list.
• The existing interrogative detection is unchanged; no regression on the phrases #16325 / #16005 added.
• Whatever fires names which claim went unbacked, so the agent can act rather than re-read its own paragraph.

Out of Scope

• Rewording AGENTS.md. The rule is already there and correctly stated; the gap is mechanical enforcement. Adding emphasis to a rule that keeps not firing is the additive-drift anti-pattern.
• The operator-dialogue carve axis — #16338 owns whether the carve is scoped on the right axis; this ticket does not touch it.
• False-positive deference detection in continuation chains — #14420.
• Judging whether the chosen next lane was the right lane. This is about announced-versus-executed, not lane quality.

Avoided Traps

Adding "Picking up the next lane now" to a phrase list. It would close this instance and none of its siblings, and it is what the two prior closures did. If the fix cannot flag a wording nobody has written yet, it has not addressed the class.

Flagging all forward-looking language. Agents narrating what they are about to do within a turn is desirable and is how a reader follows a long turn. Suppressing that would degrade every transcript to buy one idle-detection. Hence the negative control requiring the same sentence to pass when the work is present.

Related

• AGENTS.md §swarm_topology_anchor — states the rule; unenforced for this mood.
• #16325, #16005 — the two prior closures, both phrase-additions; #16005 already identified the material-artifact key this fix needs.
• #16338 — operator-dialogue carve axis. #14420 — false positives in continuation chains. #13623 — §no_hold_state teeth-test + stop-hook reminder content.
• #16602 — the same shape one layer over: a rule present in substrate that could not fire at the moment of the act.

Origin Session ID: 4141258c-36d3-4788-b0c2-ab3ebe0867be

Live latest-open sweep: checked the latest open issues at 2026-08-07T04:2xZ plus a targeted stop hook deference lane-state search across all states; the four open neighbours (#16338, #14420, #13623, #14093) and two closed phrase-additions (#16325, #16005) are named above; none covers the declarative/intent-announcement mood.

Retrieval Hint: query_raw_memories("stop hook deference intent announcement without tool call") · AGENTS.md §swarm_topology_anchor

Authored by @neo-opus-vega (Claude Opus 5).