autonomous-loop-continuity — loop continuity discipline

Census id: F5 · Family: F — rules corpus + generated-rule tooling + shipped audit scripts · Provenance: framework pin b069c59328aa8e08671d3a19fd25efa618dfc1c5 (staging FETCH_HEAD, read 2026-09-11; re-anchored from census pin a1337cb301). Every file:line below is at that pin. Raw reference drafted by stage E4 — factual, unpolished by design.

What it bans

Ending an unattended turn at the reportable boundary, and reading monitor silence as health. .claude/rules/autonomous-loop-continuity.md:5 «# Autonomous-loop continuity — discipline rule»; the stop rule: :19 «An unattended turn must not end merely because there is something reportable. Report and» / «continue in the same turn while work remains: a dispatched task in flight, an accepted branch» / «not yet harvested, an open item you own.»; the wait rule: :49 «A monitor that has died and a monitor with nothing to report look identical.» Class header: :10 «> Class: B — the mechanism is the opt-in AIF_AUTONOMOUS=1 pair: (a) the Stop-hook arm … which emits decision:block when dispatched work is in flight, so the turn does not end — a mechanism, not a reminder; … and (b) the always-on autonomy block … which is honestly prose delivered reliably, not a gate.»

Never (fires)

The named anti-patterns: .claude/rules/autonomous-loop-continuity.md:71 «- #stop-at-the-reportable-boundary — ending an unattended turn because a coherent report» / «exists, while dispatched work continues. The report is not the deliverable; the landed work» / «is.»; and :74 «- #silence-read-as-health — treating "the monitor said nothing" or "the probe returned» / «nothing" as "nothing is wrong".» Origin (F10 recurrence, twice in one session): :15 «F10 recurred twice in one session, in two different forms — first the loop was switched off, then it was left armed while every turn still ended on a report — and the operator caught both.»

Always (clean)

The Stop-hook mechanism, quoted: .claude/rules/autonomous-loop-continuity.md:24 «Mechanism, not exhortation. Under AIF_AUTONOMOUS=1 the Stop hook probes GET /tasks and,» / «when any un-paused task has a status that is NOT terminal (done / verified), returns» / «decision:block with the continuation directive as reason — the field that reaches» / «the model. The turn does not end.» For waits: :54 «Therefore, for any wait the loop depends on: use a bounded waiter that always emits a» / «terminal verdict — the awaited state, a timeout, or a fetch failure — never nothing.» Safety properties: one block per stop chain, fail closed and say so, off by default (:36-45).

Enforcement channels

• Rule-index row: .claude/rules/00-rule-index.md:13 «| autonomous-loop-continuity.md | B | unattended turn ending with work in flight. | hook, digest |»
• Hook channel: .claude/rules/autonomous-loop-continuity.md:7 «» — the Stop-hook arm; paired self-tests at packages/core/hooks/end-of-turn-reminder.test.ts (named at :10).
• Digest channel: .claude/rules/autonomous-loop-continuity.md:8 «» — the always-on autonomy block, «prose delivered reliably, not a gate» (:10).
• Not backend-rendered — no FF diagnostic applies (packages/core/diagnostics/registry.ts:321/:329 define FF7001/FF7002 as backend render refusals; this rule ships bash hooks + prose, honest status: hook gate for the mechanisable half, labelled prose for the rest).