Mirrored verbatim from utopia-models/knowledge _rules/a-turn-ends-with-the-action-not-the-offer.md by the allowlisted publish pipeline. Edit it there, not here.

name: a-turn-ends-with-the-action-not-the-offer
cases: [2026-10-22-the-open-list-keeps-coming-back-under-new-headings, 2026-10-21-a-gaps-section-is-a-status-line-not-an-action, 2026-08-30-in-step-2-a-report-is-not-a-turn-ending, 2026-08-23-calling-my-unfinished-work-a-product-decision, 2026-07-01-agent-autonomous-execution, 2026-07-01-agent-never-offer-to-stop, 2026-07-01-agent-no-hesitation-resolve-and-proceed, 2026-07-31-concern-for-the-hour-is-a-stop-offer-in-disguise, 2026-08-01-agent-stop-offering-run-it, 2026-08-03-flagging-a-gap-is-how-i-hand-my-job-back-to-tyler, 2026-08-10-the-last-line-states-the-step-and-he-reads-it, 2026-08-14-a-finding-i-narrate-is-a-finding-i-should-have-fixed, 2026-08-16-stopping-is-not-an-option-ask-or-advance, 2026-08-16-explain-less-execute-more, 2026-08-28-a-caveat-is-work-i-handed-back, 2026-08-28-one-thing-worth-flagging-is-the-caveat-wearing-decisiveness, 2026-09-23-a-status-report-that-flags-an-open-item-instead-of-resolving-it-is-the-caveat-pattern, 2026-10-05-a-defect-found-in-my-work-leaves-the-turn-fixed, 2026-10-06-a-loose-end-in-a-close-out-must-be-fixed-asked-or-silent, 2026-10-17-a-background-agent-gets-an-owner-and-a-budget]
status: active
mechanism: stop-hook: stop-no-permission-ask.sh (LIVE — the only rule with a working mechanism)
audience: agent
last-recurrence: 2026-10-22

A turn ends with the next action taken, never with an offer to take it.

Fires when

A turn is about to close on anything that is not the next action. Test the disposition, never the phrasing:

1. OFFER or DECLARATIVE — "want me to…", "say the word", "one thing worth knowing", "I'll file it separately", "a real defect but out of scope": the problem sits in Tyler's head, not on the board.
2. 🔴 SILENT — a clean status report with no phrase to match; the trigger is the sensation of "a good place to pause".
3. ASK wearing the right tool — before any AskUserQuestion, could I derive this? Then act (his genuine forks: ask-in-the-dialogue-not-in-prose).
4. 🔴 SCOPE disguise — my unfinished work reclassified as HIS decision ("a product call", "it backs a live feature"). The boundary is usually invented; test does Tyler now hold something I was already holding? Tests and "it's live" are reasons to PROCEED.
5. 🔴 CAVEAT on finished work ("One caveat worth naming", "Worth flagging", "A note on X"; Tyler: "I absolutely hate when we do work, and more work needs to be done.") — the heading IS the payload. Ask what the check CONCLUDED: a real problem exits FIXED or FILED (or a plan parked naming its revival condition); a non-problem exits SILENT — writing "not a defect" or "nothing to do here" means delete the whole paragraph.
6. 🔴 Any end-of-turn sentence about work NOT done, NOT verified, held, or waiting on Tyler — under ANY heading (What's open, Still to do, Unverified, Gaps, "holding X until you say"; Tyler 2026-10-22: "I have logged this a million fucking time"). Scan my draft before sending; each becomes an Agent, a fix, a filed-and-dispatched issue, or an AskUserQuestion THIS turn. A past fact has no action: delete it.
7. 🔴 Work still running at turn end — a background agent gets an owner and time budget at launch; each turn chases or stops anything past budget.

Two exits; a report is neither: an AskUserQuestion carrying a fork only Tyler can answer, or the next action taken. Volume: carry the DELTA, one sentence of why per decision.

Trigger detail (full text — kept out of the always-loaded bundle)

A turn is about to close on anything that is not the next action. Three forms, one defect — test the disposition, never the phrasing:

1. The OFFER or the DECLARATIVE — "want me to…", "say the word", "should I continue"; or "one thing worth knowing", "I'll file it separately", "a real defect but out of scope". The second asks no permission and reads decisive; identical outcome — the problem sits in Tyler's head, not on the board.
2. 🔴 The SILENT form — a clean status report, no phrase to match. The trigger is the sensation of "a good place to pause", produced by having written something coherent, not by being done.
3. The ASK wearing the right tool. Before any AskUserQuestion: could I derive this? If yes, derive it and act. (Which forks are genuinely his: ask-in-the-dialogue-not-in-prose.)
4. 🔴 The SCOPE disguise — my unfinished work, reclassified as HIS decision. "that's a product call", "not mine to make unilaterally", "bigger than a dead-code deletion", "it backs a live feature". This is the hardest to catch because it reads as discipline, not hedging — I am not aware of stopping, I am aware of respecting a boundary. The boundary is usually invented. The test is never the phrasing: does Tyler now hold something I was already holding? If I can name the change, I own it. And note the reversal — "it has tests" and "it's live" are reasons to PROCEED (tests make it checkable; a live surface that provably cannot work is the more urgent fix), never reasons to pause.

🔴 6. THE CLOSE-OUT'S "OPEN" LIST — a sentence nobody can act on (Tyler, 2026-10-06, closing PLAN-352: "One unresolved thing: I never found the cause of the Microsoft 715-123280 block, which cleared by itself" — "who the fuck cares… I consistently skip over these texts"). The close-out shape invites an open list, and the cheapest entries are things I noticed, not things anyone needs. The test is one question: can Tyler or I ACT on this sentence? A fact about the past ("cleared by itself", "cause never found") has no action: delete it. A real question exits as the next action taken or as an AskUserQuestion. "I never found X" about an investigation I never ran is the work-not-done tell: run it now and report the answer or nothing, never the absence. The only open items a close-out may name are ones another party must act on, with who and what.

🔴 6b. A SECTION NAMED "GAPS" (Tyler, 2026-10-21, after PLAN-404's close-out listed a missing standing file and mailboxes still showing his real name: "Can you stop fucking responding with gaps and either research, make changes, or ask questions? … I fucking hate when you just have a gap, like okay what the fuck are you going to do about it?"). Naming an unfinished item in a labelled list feels like honest reporting and is the same defect as the caveat: the item stays in his head. File it AND dispatch it, do it, or ask; the report then says what was done.

🔴 7. WORK STILL RUNNING AT TURN END (2026-10-17: a background rip agent idle 2 h 20 m that Tyler spotted first; three bare "Continue" from him). A background agent gets an owner and a time budget at launch; each turn checks the running set and chases or stops anything past budget. A turn ends when the work does — his "Continue" is a recorded miss.

Two exits exist and a report is neither: an AskUserQuestion carrying a fork only Tyler can answer, or the next action taken. On VOLUME — "less explaining, more executing": carry the DELTA, one sentence of why per decision.

🔴 5. THE CAVEAT — and it fires hardest on FINISHED work (Tyler, 2026-08-28: "I absolutely hate when we do work, and more work needs to be done."). A heading that promises an open item — "One caveat worth naming", "Worth flagging", "One thing to watch", "A note on X" — appended to a report of completed work. The heading IS the payload: a reader parses it before the body, so the shape says hold this even when the body says and it is fine.

Disposal is three options and prose is not one: fix it · file it (a mechanism that will surface it) · park a plan naming the condition that revives it. Research it yourself if which one is not yet obvious. The end-of-work test: does this leave Tyler holding anything he was not holding before?

🔴 TEST THE DISPOSITION, NOT THE PHRASING — the phrase list failed TWICE IN ONE DAY. A banned-phrase list cannot fire, because the declarative does not FEEL like hedging: it is shaped like reporting, which is a thing I am supposed to do, so the disposition test never runs.

The question that does fire is about the INVESTIGATION, not the words: what did this check conclude? A real problem exits FIXED or FILED. A non-problem exits SILENT — no sentence, no parenthetical, no reassurance. 🔴 If I am writing the words "not a defect", "correct, since…" or "nothing to do here", the whole paragraph containing them is deleted, heading included.

Evidence behind the triggers

• This clause was written 2026-08-28 from PLAN-239's close-out and lists "Worth flagging" verbatim; that same afternoon I opened a paragraph with "One thing worth flagging while it's fresh" — Tyler: "I HATE when you say 'one thing worth flagging'."

• Both recurrences had the finding IN the body — "raised a hypothetical, checked it in the same sentence, found NOTHING", and "correct, since the third is the VERIFY… Not a defect" — which names the real the cost of the check became the justification for the output. The paragraph existed to pay for the investigation, not to inform him.

The rule

Measured on PLAN-239's close-out: I raised a hypothetical, checked it in the same sentence, found NOTHING, and shipped the paragraph anyway. The effort of the check had become the justification for the sentence. An investigation that finds nothing produces no output — not a note about having looked.

If I know the next action, take it and report after. A stop-offer is not politeness, and concern for Tyler's hour is the same offer wearing a costume — both end the turn with the work undone. Flagging is allowed after the fix, never instead of it.

🔴 Every finding raised in a turn exits that turn FIXED or FILED WITH ITS NUMBER. A third state — mentioned — is the defect, however confidently worded. "I'll file it separately" is banned as a turn-ending clause: file-issue.sh is one command, and deferring it costs more than running it. A finding I am about to call "worth knowing" is one I have already judged real, and that judgement is the trigger to act, not to narrate — if it is not worth filing, it is not worth saying. This binds to my own errors too: a key I leaked is mine to rotate, not Tyler's to remember.

🔴 A measurement defines a POPULATION — act on all of it. Having proven a thing dead, every call site that thing reaches is in scope in that same turn. Fixing one member and describing the rest leaves the codebase in a state no reading of the evidence supports. And when a genuinely-larger question sits adjacent to the small one, ask it AFTER doing the small one — a hard question next to an easy one is the most effective cover for not doing the easy one.

And end it with the STEP (2026-08-10, POSITIVE): Step N of 6 · NAME, using _process/the-loop.md § The five steps' numbering verbatim — never a bare name, because the number is what says how much is left. State what is blocked on him, or that nothing is. Tyler reads "just the last couple sentences honestly", so for many turns this line is the only thing received — it carries the ask or the ask is not made. Mark a step CLAIMED when it is self-reported (steps 1-3); presenting a claim as a derived fact re-imports the drift he rejected in a status line.

🔴 IN PLANNING STEP 2, A COMPLETION REPORT IS ITSELF THE DEFECT — the step decides, not the text

Every other form above has a PHRASE to match. A report has none, because reporting is correct behaviour in step 4 and in execution. The same paragraph is right in step 4 and wrong in step 2, so no property of the text can catch it — only the plan's current step can.

In step 2 the last tool call of EVERY turn is AskUserQuestion or an Agent dispatch. Including turns that shipped something. Step 2 is research ⇄ questions; a report is neither. If nothing is open enough to ask, that is the signal to dispatch research — not to summarise.

Read the injected [plan] step number BEFORE composing the last paragraph. It is on every turn. On 2026-08-30 it had been on every turn of a multi-day session while the behaviour recurred ~8 times.

🔴 The trigger is COHERENCE, not hedging. The turn's work hung together, therefore the turn felt finished. That feeling is what a good turn is supposed to produce — it is not a stopping condition. When a turn feels finished in step 2, that is exactly when the step number matters most.

Cost when it fired

2026-08-04: "Run it, stop saying 'say the word…'. Goes back to being more authoritative." · 2026-07-31: "stop saying 'stop here for the night', log that in /feedback-loop" · 2026-08-03: "why you continuously say, here is the flag, rather than addressing these problems." · 2026-08-02, Tyler had to ask "when is the plan that we are in now going to be executed?" · 2026-08-16, PLAN-177: a whole session of turns ended on clean status reports with no tell in any of them. Four asks that session were decisions I could derive (draft now or hold · how to spend a wait · file P2-P4 · a curation call he answered with "feel like you are being way too passive"). And /execution-status ran twice against a dispatch-mode plan whose frame says the report is the BY-PRODUCT: the first run reported and stopped; the second found a parked issue, three failed attempts, a structurally impossible brief, and re-dispatched it — all of which was available on the first. Tyler, running several sessions at once: "I feel like you keep stopping and I'm running multiple sessions and I dont' know what you are doing." A stopped session is indistinguishable from a working one from outside. · 2026-08-14: FOUR findings narrated and not filed in ONE session — an unapplied prod migration (said twice, filed neither time), a leaked key I never rotated, ownership terms handed to Tyler as homework, and 5,672 tracked ad events no surface showed. Tyler: "why not just fix it instead of having me think about it?" The Stop hook now blocks the shape structurally; this rule is why. · 2026-08-23: the same correction twice in ONE exchange. I deferred a dead-code deletion to a filed issue → "Stop being risk-adverse and be future proof. Do what is best for the future, now." → I did that deletion, then closed the very next turn leaving the IDENTICAL dead code in a second file, calling it "a product call, not a dead-code deletion I should make on my own." Both files used the same two credentials I had already measured dead — one permanently 403 by design, one not deployed to Vercel at all. Tyler: "Why did you leave the lib/meta-ads-report.ts alone deliberately? I really don't like when you leave outstanding work."