Files
agent-estate-wiki/incidents/2026-10-18-degenerate-stream-canned-substitute.md

5.9 KiB

title, type, status, created, updated, verified_on, review_after, confidence, tags, sources
title type status created updated verified_on review_after confidence tags sources
Degenerate Model Stream Delivered a Canned Non-Answer — FreshRSS Midday Briefing incident active 2026-10-18 2026-11-10 2026-11-10 2026-11-24 high
cron
agent-loop
delivery
freshrss
provider
live-sqlite-cron-executions
live-sqlite-cron-incidents
live-sessions
live-logs

Degenerate Model Stream Delivered a Canned Non-Answer — FreshRSS Midday Briefing (2026-10-18)

Re-verified 2026-11-10 (nightly audit): no recurrence in the 22 days since. ecc1fc13655a has logged 6 runs since 2026-10-18, 0 failed (executions.db; newest 2026-11-09 12:00:41 completed, no error). review_after extended to 2026-11-24. Status kept active: the degenerate-stream class has no recorded fix, only absence of recurrence.

What happened

The 2026-10-18 12:00 slot of FreshRSS Midday News Briefing (ecc1fc13655a, 0 12 * * *, delivery telegram:-1004321904721:346, pin custom:omni-101 / smart-route) delivered a placeholder sentence to Telegram instead of the midday briefing, then failed.

The delivered message, stored verbatim as the run's assistant output:

Your request was not processed. Send it again if you still want me to carry it out.

Timeline (AWST; executions.db row c53b69ce19e24d888635b38b7552d200)

Time Event
12:00:56 run claimed
12:01:11 run started
12:08:11 progress_at — canned reply written and delivered (delivery_outcome: delivered)
12:08:22 incident ecc1fc_21a72aa11cc4 opened — state alerted, failure_type: agent
12:08:25 run ended failed — RuntimeError: Model output entered a repetition loop and the stream was cut mid-loop; discarding the degenerate response.

Impact

  • The user-visible Sunday midday news briefing for 2026-10-18 was replaced by a non-answer. The run's own incident alert fired, but the Telegram message that arrived did not read as a failure — it read as a terse prompt. A delivered message is not evidence of delivered content.
  • last_status: error, failure_streak: 1; the row clears on the job's next clean slot (2026-10-19 12:00).
  • The 2026-10-18 18:01 79f9bb41ecf5 weekly review run correctly listed it (“Oct 18 (today): FreshRSS Midday News Briefing — repetition loop, stream cut”), so the failure was visible in the weekly health summary even though the delivered message was not.

Cause and open questions

  • Confirmed: the model stream hit a repetition loop; Hermes's degenerate-stream guard cut the stream, discarded the response, and substituted a canned “not processed” line — which was then delivered on the job's normal route before the run aborted.

  • Only recorded instance in the retained logs: the string repetition loop appears in agent.log* only on 2026-10-18 (2 lines: the guard message and the cron error). Log retention is short, so this is “first recorded”, not “first occurrence”.

  • Open: should the guard suppress delivery (or re-queue the run) instead of delivering a non-answer under the job's delivery route? To a reader scanning a digest topic, a canned sentence is indistinguishable from content.

  • Open: whether the repetition loop is a model/provider-quality symptom. The job rides the same custom:omni-101 alias family that produced the 2026-10-14 → 10-15 connection-refused storm and the 2026-10-16 → 10-17 resource-pressure 503 burst. Single-node topology unchanged.

  • Second recorded surface: an interactive session, 2026-10-20 18:37 AWST — and a different guard fired. Session 20261020_183748_6d40ea (source: desktop, model: minimax/minimax-m3.1-flash:free, 2 messages, 0 tool calls, end_reason: agent_close) holds one user question — “i want to use the nous oauth dashboard aut for this gateway how do i do that. can i also set up a gateway token of …” — and one assistant reply: the identical canned substitute. The turn-end log names the trigger: Rebuilt-message restart limit (3) exceeded; ending turn instead of refunding the iteration budget indefinitely. → reason=rebuilt_restart_limit_exceeded model=qwen3.5:397b api_calls=1/150 tool_turns=0 response_len=0 (single occurrence in the retained logs). No incident row was created — the ledger relays cron failures only, so a user's unanswered question never enters it. The question itself was never answered → current-state.

  • Open: whether the canned sentence should be suppressed for interactive sessions (where the user is present and would rather see the real error) while kept for cron (where silence looks like success).

Recurrence 2026-10-20 (interactive desktop session, rebuild-limit path)

Same visible outcome — a canned non-answer persisted as the assistant's reply — reached by two routes on the same day: the 2026-10-18 cron path (repetition-loop guard) and, on 2026-10-20 18:37, the rebuilt-message restart limit: the conversation loop rebuilt the message list three times, gave up, and ended the turn with response_len=0, leaving the canned line as the only thing the user saw. Correlation with the same-day provider change is not established — the session's own pin was minimax/minimax-m3.1-flash:free, while the turn-end line names model=qwen3.5:397b; the two disagree and the audit cannot resolve which model the failing API call actually went to. The durable conclusion holds regardless: the canned substitute is emitted by the harness's failure path generally, not only by the repetition-loop guard. It is a delivery-integrity problem, not a stream-parser one.

No secrets recorded; job IDs, incident IDs, timestamps and status codes only.