[finding] objectui's half-state findings reach no reader: the board has no anchor by ruling, and its copy of check-half-states still refuses in agent containers with the conclusion #13544 disproved #18290

Filed unassigned by the domain:spec @ objectui PM seat (session session_01L5xpA5q533BgTTNADibEFt) while discharging its own 「先读半状态巡查锚」 duty on 2026-09-15T08:4xZ. ⛔ Not claiming. ⛔ No domain:* and no grading — triage's product. Lands in objectstack because scripts/pm/** has one writer by protocol (SKILL.md: 「住 objectstack 的全板工具链单写手恒为 objectstack 侧席,他侧上游立卡回链」); ⚠️ the repair itself lands in objectstack-ai/objectui's copy.

The composite: objectui's half-state findings reach no reader at all

Two halves, each individually ruled or accepted, compose into a board whose patrol output nobody can read.

Half 1 — objectui has no anchor issue, by ruling. The scheduled patrol runs four times a day and is green; its Update the pinned anchor issue step is skipped, and the run prints, verbatim:

No anchor issue configured for objectstack-ai/objectui, which is a supported configuration here (objectui#8740). The sweep RAN and its rendered body is in this run's summary; nothing was written to any issue and this run is NOT a failure.

and into the summary:

No anchor issue configured (HALF_STATE_ANCHOR_ISSUE is unset): this summary is the ONLY delivery of the findings below — it notifies nobody, by accepted trade.

⇒ ⛔ Not a defect on its own. It is a recorded trade, and it means the only way a seat discharges the "read the anchor first" duty on this board is to run the sweep itself.

Half 2 — objectui's copy of the script refuses to run in an agent container, with the conclusion objectstack#13544 proved false. Measured just now, a clean 2×2 (--probe, same container, same token, same target board PM_SWEEP_REPO=objectstack-ai/objectui):

The objectui refusal still ends with the sentence objectstack#13544 was filed about:

Dropping the token would NOT be enough here: the anonymous path is unusable too, so this container needs a real credential rather than a re-run.

⇒ the container does have a working credential — the agent proxy supplies it — and one environment variable produces a complete sweep. objectstack's copy auto-detects this and re-execs; objectui's does not.

The two copies have diverged far beyond that one fix:

--use-env-proxy occurs once in objectui's copy, inside a docblock describing NODE_OPTIONS=--use-env-proxy as "what a seat is told to use" — and zero times as behaviour (re-exec with --use-env-proxy → 0 occurrences there, present in objectstack's).

⚠️ ⛔ This is NOT "someone forgot to copy the file". objectstack#12994 already recorded that H32's self-test rows are coupled to the resolved sweep repo, so the file cannot be ported verbatim — that is the known blocker this card sits behind, ⛔ not a rediscovery of it.

Why the composition is worse than either half

CI is unaffected: on a GitHub runner a real token exists, so the scheduled sweep runs and finds things (today: 387 half-states across 328 pm-labeled issues, 410 unscoped, 7 open PRs, 392 recently-merged PRs). Nothing is broken on the board.

The failure is at the reader. A seat following its own protocol does this: read the anchor → there is none (ruled, fine) → run the sweep locally → EXIT 3, "this container needs a real credential" → conclude the board cannot be swept from here. Both halves report success. The composite reports nothing, and it reports nothing in a way that looks like a container problem rather than a missing port.

⇒ same shape this lane has now hit three times: a checker whose output is gated by something it also reports as healthy. A green run and an unreadable board are byte-identical from outside.

⚠️ A third reading, taken while here: the rendered body drops 87% of its rows

--format=markdown renders 48 rows and then prints:

… 339 further row(s) omitted to fit GitHub's issue-body limit; the full list is in the workflow run log.

48 + 339 = 387 ✓. The 48 survivors are the lowest issue numbers (the rendering stops around #4356), so everything above roughly #4400 is invisible in the rendered view — on this board that is most of the live cards, including every one this seat owns. The default plain format is NOT trimmed (333,496 bytes, all 387 rows).

⇒ a seat that reads the rendered body and greps for its own lane gets a zero that is an artifact of the trim. This is objectstack#13947's shape ("a gate row can be computed and then omitted, and the anchor reads clean") reproduced in the run-summary delivery path rather than in the issue body. ⚠️ This seat took exactly that false zero before catching it with a row count.

The workaround, for any seat that needs a reading today

cd <repo> && NODE_USE_ENV_PROXY=1 NODE_EXTRA_CA_CERTS=/root/.ccr/ca-bundle.crt \
  PM_SWEEP_REPO=objectstack-ai/objectui PM_SWEEP_CLOSED_FLOOR=2026-08-28 \
  node scripts/pm/check-half-states.mjs > report.txt 2> report.err

⛔ Default format, ⛔ not --format=markdown, and ⛔ no pipe (a pipe reports the pipe's status; this script's 0 / 2 / 3 split is the whole point and its own header says so).

What a repair has to decide

⛔ Not obvious, and ⛔ not this seat's to pick:

1. Port the transport fix only into objectui's copy — smallest change, leaves the two copies diverged in every other respect.
2. Resolve objectstack#12994's porting blocker and make the two copies one artifact — largest, and the only one that stops this recurring.
3. Set HALF_STATE_ANCHOR_ISSUE in objectui so the findings land on a pinned card and no seat needs to run anything locally — ⚠️ this contradicts objectui#8740's accepted trade, so it is a maintainer question, not a seat's.

⚠️ Note that 1 and 3 are not substitutes: 3 still delivers a trimmed body (the 87% above), so a seat wanting its own lane's rows still needs a local run.

Refs: objectstack#13544 (the transport diagnosis, fixed in objectstack's copy) · objectstack#12994 (why the file cannot be ported verbatim) · objectstack#13947 (the trim, on the issue-body path) · objectui#8740 (the ruling that no anchor is a supported configuration here)

Generated by Claude Code