Files
agent-estate-wiki/incidents/2026-11-02-job-radar-stage1-iteration-cap-recurrence.md
T

6.3 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
Job Radar Stage 1 Hit the Iteration Cap Again (2026-11-02) — Incomplete Run Recorded `completed`, Output Suppressed incident active 2026-11-03 2026-11-03 2026-11-03 2026-11-10 high
cron
job-radar
iteration-cap
silent-failure
notion
camofox
live-sqlite-cron-executions
live-sqlite-cron-incidents
live-state-db
live-jobs-json
session-history

Job Radar Stage 1 Hit the Iteration Cap Again (2026-11-02) — Incomplete Run Recorded completed, Output Suppressed

Headline

Third occurrence of the Stage-1 budget class, and the first that the cron ledger reports as healthy. On 2026-11-02 the 07:00 dca8482e4f76 (Daily Marketing Job Radar — Stage 1 (Discovery)) ran for 9 h 01 m 19 s (started 07:00:33.378239 → finished 16:01:52.496335 +08:00), consumed the harness's entire tool-calling iteration budget, and closed with:

Field Value
status completed
error NULL
delivery_outcome suppressed

No handoff artefact was written. Nothing reached Notion, no cover letters were generated, and Stage 2 was never triggered. Because the run's own status is completed with no error, a health check that inspects executions.status alone reports this day as green — the failure is invisible in the ledger.

What the run did and did not do

The session (Daily Marketing Job Radar - Discovery · Nov 02 16:01, local-368c8fe9…) shows the discovery half genuinely worked:

  • Camofox was restarted at run start after its HTTP server wedged (see systems/browser-backend).
  • SEEK 51 unique roles, LinkedIn 60 unique job IDs, Indeed live, WA Government 35 results — 197 raw candidates, 164 new after three-level dedup against the 479-row Notion board.
  • A URL-canonicalisation bug was found and fixed mid-run: au.seek.com vs au.seek.com.au had been dropping SEEK and Indeed entries from the consolidated list; the fix was a job-ID-based canonicalisation.
  • Adverts were individually verified fresh and active; salaries, closing dates and JD bodies were pulled.

Then the ceiling hit. The runtime injected "You've reached the maximum number of tool-calling iterations allowed. Please provide a final response…" while the agent was still reading JD text for the borderline finalists, and the run terminated in the scoring phase.

The agent's own closing report was explicit — and it was not delivered:

"Status: INCOMPLETE — handoff file NOT written, Stage 2 NOT triggered. … Because the mandatory deliverable (the handoff JSON) was not produced, today's roles have not been handed to Stage 2."

That text is delivery_outcome: suppressed, so Anthony saw nothing on 11-02 — not the roles and not the warning that the roles were missing.

Downstream: both Stage 2 slots behaved correctly

Slot Session Ran Result
11:30 local-92665337… (Write and Letters) 11:30 → 12:25 Exited [SILENT] after verifying Stage 1 was alive but had no handoff
14:30 (safety net) local-05a3f00b… (Write and Letters) 14:30 → 14:30 Exited [SILENT] — 6 messages, 2 tool calls

Both declines were correct: no handoff existed to consume. The safety-net design worked as intended — the gap is that Stage 1 cannot finish inside its budget.

Recurrence pattern

Date Incident page Failure shape
2026-10-27 incidents/2026-10-27-job-radar-budget-exhausted Tool-call budget exhausted before the Notion write stage
2026-10-28 incidents/2026-10-28-job-radar-stage1-iteration-cap Stage 1 hit the agent iteration cap before the handoff
2026-11-02 this page Same cap; run recorded completed and its report suppressed

Three hits in seven days, each ending in the same place: discovery complete, handoff absent, Notion write stage never reached.

The "green row, no artefact" anti-pattern

This is the fourth appearance of the class systems/scheduled-tasks §9 of the maintenance playbook already documents (a job whose durable ledger entry says success while the thing it was supposed to produce does not exist). The new variant here is that the ledger flipped from red to green:

Date cron_incidents row for dca8482e4f76 executions.status
2026-10-27 dca848_e93a538ade6b (failure_type: auth) non-clean
2026-10-28 dca848_0a2f03b5952b (failure_type: timeout) non-clean
2026-11-02 none raised completed, error: NULL

The two earlier caps each raised an incident whose error text is literally the run's INCOMPLETE report. The 2026-11-02 cap raised nothing at all (cron_incidents holds no row for dca8482e4f76 with a first_seen_at on 2026-11-02; fleet-wide unclosed rows were 0 at the 2026-11-03 audit). The run's delivery_outcome is suppressed, so no message reached Anthony either.

Detection rule: for any job whose deliverable is a file (a handoff JSON, a report, a generated page), executions.status is not sufficient evidence. Verify the artefact exists at its expected path, or that cron_incidents stayed quiet for a reason, before treating the slot as served. Silence from cron_incidents is not the same as success — here it is the symptom.

Open questions

  1. Is the iteration budget the right control, or is the task too large for one agent turn? Three caps in a row suggest Stage 1's work (four job boards → dedup → individual advert verification → JD reading → scoring) does not fit a single turn at any budget.
  2. Should the harness record a cap-hit as something other than completed? A distinct terminal status (or a non-null error) would make this visible to every existing health check.
  3. Should the run's final response be delivered even when the job "succeeded"? Suppressing an INCOMPLETE report converts a loud failure into silence.
  4. Did the 11-02 roles get picked up on the next day's run, or is 2026-11-02's candidate set lost? The 164 deduped candidates were never persisted.