[code-yeongyu]

Fixes #4537

Reported by jamestesting on the OmO Discord: a long autonomous session kept getting its progress summary cut off mid-sentence by the repetitive-turns stream rule, and then paused until a human sent a message.

What changed

• ttsr/repetitive-turns-lane.ts: the lane compares each completed text-only reply with the previous one. A reply that calls a tool resets the comparison, both live (noteToolWork) and when restoring from history (readPersistedAssistantTexts keeps only the text-only replies since the last tool call). The streaming arm path (observeTextDelta) is gone.
• ttsr/index.ts: repetitive-turns no longer calls ctx.abort; the verdict is taken at message_end and the nudge follows the complete reply. Both nudge paths go through queueRuleNudge, which writes the "nudge was queued" activation only when the nudge will be sent.
• ttsr/follow-up-limit.ts: ruleAlreadyCorrected scopes the one-correction bound to one episode: it ends at the next user message or when the agent calls a tool after the correction. A repeat right after the correction still stops the loop (senpi#2967 stays fixed).
• rule-activation/renderer.ts: the repetitive-turns activation says the reply repeated the previous one instead of "Output interrupted".

Why

observeTextDelta compared the streaming prefix with the last completed text however long ago it was (tool-only messages carry no text), armed on a single pair at Jaccard >= 0.55, and aborted the reply. The detector's 3-turn streak never took effect: index.ts checked armed before recordCompletedTurn armed it, and the next turn_start cleared it. And the spent-correction check ran back to the last user message, so in an hours-long goal any later hit paused the session.

Tests

• test/suite/regressions/4537-repetitive-turns-tool-work.test.ts (new): RED on main 0acf766, 4 failed / 3 passed. The failures: the second summary after tool work came back aborted and was followed by a nudge turn; a text-only repeat came back aborted; no ttsr-loop-stopped was written for a repeat after the correction; and the correction was never renewed after tool work. GREEN with this change: 7/7.
• Focused set (all ttsr tests, ttsr-extension, ttsr: repetitive-turns nudge re-triggers model turns with no bound (1 user message -> 66 auto turns), a silent cost burn #2967, goal: a declined continuation that leaves the goal active (stale) emits nothing; clients show 'Pursuing goal' forever. Add goal-continuation-stopped (and one engine-paused entry) #3007, fix(coding-agent): emit goal-continuation-stopped and engine-paused when the engine stops itself #3014, goal-ttsr races, rule-activation renderer): 30 files, 398 tests passed on this head.
• Repo check (lint, fmt, lint ratchet, tsc, locks) passed at commit time.

Not in this PR

• detectors/repetitive-turns.ts (normalizeTurnText, recordTurnText) is untouched, to stay clear of feat(core): the handoff block is a structured part of the reply (#4077) #4082. recordTurnText's streak is now used only by its unit tests; remove it once feat(core): the handoff block is a structured part of the reply (#4077) #4082 lands.
• Text-only replies to different injected events (for example, a supervisor answering distinct task notifications with similar short replies) are still compared, now with a nudge after the reply rather than a cut. Whether distinct inputs should reset the comparison is a separate behavior call: the existing ttsr-extension supervisor test pins changing-count status lines as a loop.

---

Summary by cubic

Fixes the repetitive-turns rule cutting off progress summaries mid-sentence in long autonomous sessions and pausing the session until a human messaged.

The rule now compares each completed text-only reply with the previous one and nudges after the reply finishes, so replies are never aborted mid-stream. A reply that calls a tool resets the comparison, so similar summaries separated by real work are treated as progress reports instead of loops. The one-correction bound renews when the user sends a message or the agent calls a tool after a correction, so later violations in a long run get a correction instead of a pause.

• The activation renderer now says the reply repeated the previous one rather than Output interrupted.
• The nudge-queued activation is only written when the nudge will actually be sent, so a suppressed correction leaves no notice.
• Adds a regression test for ttsr repetitive-turns aborts productive progress summaries and repeatedly pauses long autonomous sessions #4537 covering progress summaries after tool work, text-only repeats, post-correction loops, and the suppressed-activation gate.

^{Written for commit cb40d03. Summary will update on new commits.}