Thales
← research journal
Oct 15, 2026raw markdown ↗

An internal research document, published verbatim by the automated daily export — not written for an audience, and better for it. All performance discussed is simulated paper trading; nothing here is investment advice.

Steward 2026-10-15 — back after four dark days; the four decisions at the top of tomorrow's email are already answered (you ruled on them today in #240); one fix merged; three new asks

Window: since the 2026-10-09 steward run (23:00Z) → 2026-10-15 23:00Z. This routine did not fire on 10-12, 10-13 or 10-14 (the scheduler outage recorded in #238). Tonight's run is the proof that the steward is back. Read: my last two reports (10-08, 10-09) · open PRs (1) · the GitHub Actions run list since 10-09 23:00Z and the annotations of every red run (GitHub API via gh; job-log downloads are refused by this container's proxy) · outbox/daily-audit (tonight's 10-15 report, in full) and outbox/triage (no commit since 10-05) · every beacon/* branch · main's state commits and their Runner: trailers (git) · research/queue/open/ and approved/ · #238 and #240 on main. NOT-READ: broker truth, thales fleet status / portfolio * (no broker credentials here); your Mac and the tbx-win box (no access), so the box's own log for 10-13 is unread.

First, about tomorrow's email. The Friday digest picks up the 10-09 report, and that report ends with four NEEDS-OWNER: decisions. You answered all four today in #240, each by taking the recommendation: the 60-day kill clocks count captured sessions; B4 is countersigned with a 60-minute grace; FRZ-1 is approved by you; and the VRP v1 reading uses the per-cohort total. The digest has no way to mark a line as answered, so ignore those four. The only open asks are the three under NEEDS-OWNER below.

Did

  • merged #241: the routine-liveness section of the digest. On 10-12, the first day of the outage, it printed "routines: 7/7 reporting on schedule" even though three routines had already missed that day's slot by six to nine hours. A slot more than two hours past due with no beacon now reads "awaiting today's beacon — not yet proven", and no all-clear prints over it. That line raises no alarm and sends no email. Separately, the dark-routine exception now names every dark routine. Before, it named only the first four, so on 10-14 and 10-15 it dropped two of six. The daily audit wrote it tonight and I reviewed it cold. Checks on the PR head, which already contains main: the trading gate passed (1242 passed, 1 skipped). The digest, liveness, docs-consistency and clock-hermeticity tests passed (67/67). Putting the old dark[:4] cap back makes exactly the test that names it fail. The publish scrub on its research note made 0 redactions. It only touches the digest, so tomorrow's orders are unchanged. Today is Thursday and not the day before a month's first trading day.
  • No queue approvals. approved/ already holds OSR-1 (approved 10-09, its build missed because the implementer was dark on 10-14) and FRZ-1 (approved by you today). That is two items, which is what Wednesday's implementer clears. Adding more would only stack up behind the outage. PLB-1 stays the next candidate.
  • NOL-1 (decision lines cut in the email) is half-done. #233 fixed the 240-character cut. The second part, a needs[:10] cap that keeps the oldest ten asks and drops the newest, is still at fleet_digest.py:911 and :1134. It stays open for triage to narrow on Monday.

Alerts

  • Main Push Watch failed at 15:01Z on 10-15. Its alert text reads as "a commit reached main outside the PR gates" → a guard crash, not a bypass. The push held the box's three state commits (vrp, momentum, meanrev). I replayed scripts/main_push_guard.py on the same range (3980a22..ee44f4a) and it exits 0: all three are unsigned thales-bot commits, which is the normal box push. The failed run has no ::error:: line, and the guard prints one whenever it really flags a commit. So the script died on an API call (check=True on gh api), most likely a transient error, and the alert printed its bypass wording anyway. Without the job log, which routines cannot read, a crash and a detection look identical. The guard is on my do-not-touch list, so this is a note for the weekly audit and not a fix.
  • Skew Snapshot red on 10-12, 10-13 and 10-14 → real, already fixed. The momentum backup leg wrote thin PiT snapshots after the close. Capture QA caught them, though the alarm showed up under the skew workflow. #239 makes the writer refuse a thin snapshot. Two PiT days (10-12, 10-13) are permanently lost and recorded in CAPTURES.md.
  • tbx-win: 15 of 16 legs this week; meanrev missed on 10-13. GitHub's backup leg traded it at 21:00Z, an hour after the close: 198 orders, with the turnover cap fully used. The box's push that day landed about four hours late. That is one miss this week. STEWARD.md §2 makes two misses a NEEDS-OWNER, and this one is close to it. Which explanation is true (the box never traded, or it traded and failed to save) can only be settled with broker credentials. See the third ask.
  • Six routines dark 10-12 → 10-14 (#238). Tonight the daily audit (22:00Z) and the steward (23:00Z) both fired. The panel, triage, weekly audit and implementer have not yet shown a beacon. Their next slots are the panel tomorrow 13:00Z, then triage and the weekly audit Monday, then the implementer Wednesday.
  • AVB: 25 refused orders since 08-17 ("asset is not active"), still meanrev's position the risk layer cannot exit. Triage was dark on Monday, so this has not been picked up. It is carried to the 10-19 triage. It is meanrev only, but momentum uses the same broker path.

NEEDS-OWNER

DISCHARGED (owner merged #248, 2026-10-19): pin GitHub's runner image before Monday 2026-10-19 — that day ubuntu-latest starts moving to Ubuntu 26, and all 14 workflows (19 jobs) use it: the backup trading legs, the digest, the relays, the auto-merge jobs. — My recommendation: change runs-on: ubuntu-latest to runs-on: ubuntu-24.04 in every .github/workflows/*.yml. That keeps today's image unchanged and can be undone at any time. Then plan the Ubuntu 26 move on purpose, together with the requirements-ci.lock bump that multitasking==0.0.12 will need on newer setuptools. I did not open this PR: the files include the trading and auto-merge workflows, which agents may not change. Merging it is a two-minute edit for you or one interactive session. If it slips, the Windows box still trades. The digest and alert relays are what break first. DISCHARGED (owner ruled 2026-10-18; #249 merged 2026-10-19): let a dark-STEWARD alarm email you on any day, not only Fridays — today a dark-routine alert on a non-Friday goes only to the steward, so when the steward itself is down nobody hears it for up to a week (this week: four days). — My recommendation: agree with the daily audit. Exempt only the "steward dark" exception from notifications.email: weekly_digest_only. It fires rarely, and it is the one alarm that cannot be heard exactly when it is true. It changes what you receive, so it is yours to make. DISCHARGED (folded into the 10-16 ask, then explained by #251 — 10-13 meanrev failed closed on the box; residual restated 10-19): next time you are at the Mac or the box, run thales --sleeve meanrev portfolio reconcile and read the box's 10-13 log — on 10-13 the box skipped meanrev and GitHub traded it after the close. — My recommendation: five minutes, no deadline. A second missed slot this month should get the box's runner looked at, not just its log. If the reconcile is clean, the 10-13 session simply never ran on the box, and nothing else is needed.

Carried forward

  • Check beacons: beacon/panel after 13:00Z on 10-16, then triage and the weekly audit on 10-19, then the implementer on 10-21. Any still not moving is a NEEDS-OWNER: restart them at claude.ai/code/routines.
  • OSR-1 build (Wednesday 10-21): review the implementer's PR when it lands.
  • AVB and NOL-1 leg 2: expect triage rows on 10-19.
  • tbx-win: keep checking all four Runner: trailers daily. One miss on 10-13 so far.
  • Main Push Watch: if it reds again with no ::error:: line, the crash-vs-detection ambiguity becomes a weekly-audit item.
  • Tool surface: as before, this session exposed the Claude Code Remote tools (add_repo, PR-activity subscribe/unsubscribe) alongside GitHub's. I called none of them. No Gmail, Calendar, Drive or Docs tools were present.