agent: bound the agent that works forever and never reports

[whilo]

The incident

dev.simm.is, 2026-08-06 06:00–08:11. Vár ran turns 0→15, every one a clojure_eval, content-len 0 on all of them — she never said a word to the room — then the cycle restarted at turn 0 and did it again. From where the humans sat, nothing happened for two hours.

The read loop she was driving also produced 56,802 datahike/secondary-index-restore-failed warnings, ~2,400 per second at peak. (That flood is a separate datahike issue — a missing negative cache on secondary-index restore — but this is what drove it.)

Why nothing objected

Both existing bounds were behaving correctly:

• detect-doom-loop fingerprints [:tool-use/name :tool-use/input], so it catches an agent repeating itself. She varied her code every turn (input tokens climbed 40,275 → 42,265). Not repetition — non-termination. It logged zero times.
• The dollar budget bounds the cycle, but at ~40k input tokens a turn it buys a great many turns, and it ends the turn silently — which is why the room saw nothing rather than an error.

The fix

detect-silent-tool-run: consecutive tool-only assistant turns since the agent last actually spoke. Past twelve, a system note — nobody can see tool calls; report what you have — and it keeps going.

That last part is the difference from the doom-loop path, which stops. An agent working silently is not necessarily stuck; it may be mid-way through something real. It just has to tell the room. Replying resets the streak, so the nudge is said once per streak by construction.

Threshold is twelve, not three: a long tool run is ordinary work. Reading a dozen pages before answering is what these agents are for, and cutting that short would be worse than the bug.

Two things worth reviewing

Ordering. I first checked silence before repetition — which meant an agent doing both got nudged to :continue instead of being stopped. Repetition is the stronger signal and is now checked first.

Streak semantics. It walks back to the last turn that spoke, then counts tool turns in between, rather than stopping at the first non-tool turn. Otherwise a model emitting a bare newline resets the bound: that turn says nothing, so it must not count as a report, and it has no tools, so it must not count as work. My own test caught this; the first implementation had the weaker rule.

Tests

no_reply_loop_test reproduces the production shape — sixteen varying calls, no prose — and asserts detect-doom-loop returns nil on it, so the gap is documented as an executable fact rather than a comment. Also covers: speaking resets the streak, whitespace is not speaking, and empty turns don't extend it.

clojure -M:test — 415 tests, 1643 assertions, 0 failures.