wiki(audit): 2026-10-28 window — radar stage-1 iteration cap (no handoff, stage-2 no-oped), stocktake 400 never alerted, Steel removed from the stack

This commit is contained in:
hermes
2026-10-29 00:41:30 +08:00
parent 2cbf0b65c1
commit ec0c30fb30
12 changed files with 374 additions and 27 deletions
+9
View File
@@ -1506,3 +1506,12 @@ not pushed.
- **Unresolved:** whether the two-stage split actually fixes the budget problem — the first real 07:00 → 11:30
cycle (stage 1 handoff → stage 2 writes) had not run at audit time; why a budget exhaustion is classified
`failure_type: auth`; and the 10-27 radar incident row will only close when a clean stage-1 slot runs.
## [2026-10-29] audit | Nightly wiki audit — window 2026-10-28 00:00 → 2026-10-29 00:3x AWST: the radar's two-stage split ran for real and landed nothing, and the stocktake died on a 400 that never alerted
- **New page:** [[incidents/2026-10-28-job-radar-stage1-iteration-cap]] — stage 1 (`dca8482e4f76`) ran 07:00:32 → 10:33:18 (3 h 32 m 46 s), closed `failed`, `failure_streak: 2`, on *"hit the agent iteration cap before the handoff could be written"*. No handoff file existed (`~/.hermes/radar-ref/handoff/` held only `2026-10-27.json`), so **both stage-2 slots (`5fc292ff0f3b`, 11:30:27 → 11:31:31 and 14:30:27 → 14:31:43) `completed` with `delivery_outcome: suppressed`** — the no-op recovery rule proven in production. Second consecutive stage-1 failure at a *different* ceiling (tool-call budget on 10-27 vs iteration cap on 10-28).
- **Updated:** [[incidents/2026-10-27-job-radar-budget-exhausted]] (follow-up section answering its open questions: handoff never written, stage 2 no-oped correctly, its incident row did **not** close, unclosed rows 2 → 4) · [[runbooks/job-radar]] (10-28 slot write-up; "first real cycle" marked as run-and-empty) · [[systems/scheduled-tasks]] (new 2026-10-29 window bullet; prior bullet demoted to *Prior window*; the two in-window job rows refreshed) · [[systems/browser-backend]] (dated correction: Camofox restored, `browser-steel` plugin removed, endpoint dark, `STEEL_BASE_URL` residue in 10 live processes) · [[comparisons/steel-vs-camoufox]] · [[decisions/rollback-steel-to-camofox]] (2026-10-29 update) · [[systems/skills-index]] (stocktake-run note) · `current-state.md` (new 2026-10-28 section) · [[index]].
- **Stocktake folded, not paged:** `1a8e90ea2271` ran 03:00:28 → 03:37:58, closed `failed` (streak 1, delivery `suppressed`) on `RuntimeError: [400] … duplicate field 'reasoning_content' at line 1 column 512203` after regenerating `systems/skills-index.md` (302 skills) and `raw/stocktake/freshness-20261028.txt` (**2 stale / 103 fresh / 4 no-date of 109 pages**); it left both uncommitted and filed incident `1a8e90_f57ab4cafab8` in state **`detected`** with no alert. The audit committed the two artefacts, clearing the dirty tree.
- **Live probes (2026-10-29 00:3x AWST):** `jobs.json` 29 jobs / 26 enabled / 3 disabled, 5 non-clean `last_status`, pins 12 `smart-route` / 6 `free-stack` / 7 unpinned / 1 `deepseek-paid-smart`; `executions.db` — 2026-10-28 settled **168 / 2 / 1 running**, 2026-10-27 now **173 / 1**, 2026-10-29 so far 6 / 1 running; Steel: ping 2.5 ms but CDP `:9223` and REST `:3000` refuse, tunnel `/steel` **502**, `steel.martinwa.org` unresolved; no Hindsight service on this box (`:8888` closed, no `hindsight` units) so [[systems/memory-backend]] and [[comparisons/memory-backends]] were left unchanged.
- **Removed/archived:** none. **Secrets written:** none.
- **Unresolved:** the radar's discovery stage still cannot reach its own handoff write (two ceilings in two days); `1a8e90_f57ab4cafab8` sits `detected` without an alert; both radar incident rows close only on a clean 07:00 weekday slot.