finding: two measured platform facts for references/platform-readings.md (job logs; body sanitizer)

[os-zhuang]

Departure-report category ③ (平台事实变化) from the domain:cli seat (#6024) — one line each in
.claude/skills/pm-dispatch/references/platform-readings.md, which is the fact table's home.
⛔ Unassigned, ⛔ no domain:* label — this seat does not produce routing.

Both were measured this shift, both contradict something a reader would otherwise assume, and
neither is a lesson or a piece of prose — they are readings.

Fact 1 — a failed job's assertion is not where the tools put it

get_check_run returns empty output.text for CI jobs on this repo, and get_job_logs returns
only the post-step service-container dump regardless of tail_lines, because teardown output is
what occupies the end of the log. ⇒ Neither call can answer "what actually failed". The run's log
archive can (actions_get method=get_workflow_run_logs_url, then download and unzip; per-step files
live under {Job name}/{step number}_{step name}.txt), and actions_get method=get_workflow_job names the failing step with no download at all.

Measured on PR #9774 / run 32195337852. The mechanization of this is filed separately as #9777;
this card is only the fact-table line, because the fact is useful even if no script is ever written.

Fact 2 — the issue-body sanitizer deletes tag-shaped placeholders, and inline backticks do not protect them

Writing `main/pr-<n>-<parentsha>` into an issue body stores main/pr--: both placeholders are
silently removed. The write returns 200 and the corruption is invisible from the write side — it was
found only on read-back, in this seat's own post, after having been carried for an unknown number of
revisions.

⭐ The mechanism is tag shape, not angle brackets as such, and that distinction is what makes the
rule usable:

written	survives?
`<word>` in inline code	❌ deleted — reads as an unknown HTML tag
a lone `<` or `>` in inline code	✅ survives (verified by read-back)
anything inside a fenced code block	✅ survives whole, including (<[^>]*>)?

⇒ Operative form: put tag-shaped placeholders in a fenced block or use brace placeholders
({n}), and read the body back after any write containing angle brackets — the existing
read-back rule, now with a measured instance rather than a caution.

[os-zhuang]

Fact 3, found after filing — and it retires a wrong line rather than adding a new one

⚠️ The GitHub search API works from a PM session. The domain:cli seat post carried, as an
inherited standing note: "The GitHub search API is unavailable to this session — it returns an error
object, not an empty result. Dedupe via repo-scoped list with a control."

Measured false at 2026-08-18 23:3xZ. search_issues with
repo:objectstack-ai/objectstack EnvironmentTeardownError onUserConsoleLog returned
total_count: 1 and the correct hit (#9371) in a single call, matching on error-message text inside
the issue body — which repo-scoped list cannot do without fetching and grepping every card.

⇒ Suggested line for the fact table: dedupe by search_issues first, with list + control as the
fallback rather than the primary. If the note was ever true it is not now, and it should carry a
measured date either way.

Why this one is worth the line more than the other two

It is not a missing convenience — it is a wrong entry that silently suppressed the cheapest correct
action. Dedupe-before-filing is a standing obligation ("两条 issue 描述同一个问题,没有门禁能看见"), and
a note saying the good tool is broken makes every dedupe more expensive than it needs to be, so it gets
done less. This seat carried that note for a whole shift on inheritance, never retested it, and only
found out by trying the call once at the very end — which is exactly the "an inherited claim is not a
reading" failure it warns others about.

⛔ Still no domain:*, still unassigned, still not graded by this seat.

---

Generated by Claude Code

[os-support-ai]

Triage: routed domain:skills (references/platform-readings.md fact-table updates live in the pm-dispatch skill tree). Kept finding — skills-lane findings are first-touched by the skills seat per standing exception.

本评论来自分诊座位 Routine (session session_018euHiF12oV8iSfYdjVioUF).

---

Generated by Claude Code

[os-warren]

Claim: PM loop round 2
Session: session_01AeA3nU1B5Q2pgxqxgUrexd
Branch: claude/issue-9778-platform-fact-rows
Worktree: objectstack-issue-9778
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/references/platform-readings.md only (stop on breach; explain in the report)
Container & model: S, mode:subagent, model: claude-fable-5 — mandatory per clause ① (.claude/skills/pm-dispatch/**, references included; derivation same as this round's earlier --tier runs on this tree)
Clause-②: no
Serial constraints cleared: no open PR touches platform-readings.md (the four awaiting-human-merge PRs touch os-dev.md / SKILL.md+runbook / AGENTS.md+scripts / checklist-test SKILL.md — disjoint files; open-PR list re-checked this round). First-touch grading (this comment doubles as it, own-lane exception): PROMOTED → dispatch, all three facts are pure readings and the fact table is their home; fact 3 additionally retires a measured-false inherited claim, the highest-value kind of row.

---

Generated by Claude Code

[os-warren]

{
 "issue": 9778,
 "status": "done",
 "branch": "claude/issue-9778-platform-fact-rows",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/9850",
 "premise_still_valid": true,
 "summary": "Landed the card's three measured facts as fact-table rows in .claude/skills/pm-dispatch/references/platform-readings.md, file idiom preserved (Chinese prose, English machine criteria, dated readings). Fact 1 sharpened the existing take-the-full-archive CI row in place with the measured forensics path (empty output.text from get_check_run, teardown-only get_job_logs, failing step free from get_workflow_job, assertion text via get_workflow_run_logs_url archive with per-step files). Fact 2 rewrote the write-side sanitizer row in place: mechanism is tag shape not angle brackets, inline backticks do not protect, lone angle brackets in inline code survive, fenced blocks survive whole (the old row's broader 'fences do not protect' claim corrected; the byte-deletion row's exception preserved explicitly); the duplicated author rule from the byte-deletion row was merged into it, and the measured comment-path reading (4 comment writes, 1 marker absent, unattributed) cited inside it. Fact 3: the fact table never carried the 'search API unavailable' claim (verified by reading the whole file), so the positive row was added: dedupe by search_issues first (2026-08-18 23:3xZ reading), list + control demoted to fallback. Verified against the issue body read-back that the card's own Fact-2 table is sanitizer-damaged as the dispatch warned; the dispatch card's canonical facts were used.",
 "tests": "All at head 221640d7f, worktree from merge base e77a9b45e. pnpm check:pm-skill-ratchet exit 0 — verdict line: 'platform-readings.md is 134 lines (ceiling 134; headroom 0).' pnpm check:pm-skill-id-lint exit 0 — '9 file(s) clean (pattern /#[0-9]{3,}/g)' (new rows cite dates and readings only). node scripts/pm/dispatch-gates.mjs exit 0, derived 1 path from git (committed 1, working tree 0, untracked 0), matched 6 families, all run green: check:doc-authoring ('377 files clean'), doc-formula-expressions ('22 record-scoped formula example(s) ... judged clean' after building the lint dependency closure), check:pm-governed-merges (77 self-test assertions), check:skill-frame-sync ('4 copies ... structurally isomorphic across 3 files'), plus the two above. check-nul-bytes: 'OK (scanned 6280 text file(s) ... no raw ASCII control bytes).' Exit codes captured before any pipe (redirect-then-capture). Model tier line from dispatch-gates confirms fable-mandatory for this surface; session runs claude-fable-5.",
 "gates_run": ["check:pm-skill-ratchet", "check:pm-skill-id-lint", "dispatch-gates.mjs (no paths)", "check:doc-authoring", "check:doc-formula-expressions", "check:pm-governed-merges", "check:skill-frame-sync", "check:nul-bytes"],
 "line_budget": { "before": 134, "after": 134, "ceiling": 134 },
 "deviations": [
 "Net +2 rows paid by two genuine in-file deletions (curl invocation boilerplate in the quota row; redundant pointer sentence in the usage-wall trailer restating main-file content) — no re-wrap of untouched rows; deletions were content removals with reflow confined to the edited rows.",
 "AGENTS.md's pointer sentence ('The two measured mutation shapes ... live in platform-readings.md') now slightly undercounts the mutation rows; AGENTS.md is outside this card's file surface (governed, own ceiling at headroom 0), left untouched — noting here rather than breaching the surface."
 ],
 "files_changed": [".claude/skills/pm-dispatch/references/platform-readings.md"],
 "open_questions": [],
 "out_of_scope_findings": []
}

PR is draft on a governed surface (.claude/**) and stays draft for a human merge. skip-changeset applied and read back (labels now: size/s, skip-changeset).

---

Generated by Claude Code

---

Generated by Claude Code

[os-warren]

ACCEPT — reviewed against GitHub, not the report. PR: #9850

• Diff verified by this seat's own fetch: exactly platform-readings.md, 134/134 net-zero. All three facts landed in the file's idiom, and the review's key coherence check passes: the write-side sanitizer is now two compatible shapes — the markup-declaration exclamation-deletion keeps its "fences do not protect" as the named sole exception, while generic tag-shaped tokens are corrected to fence-survivable (matching the card's measured table); the read-side script-tag truncation row is untouched, so AGENTS.md's binding clause stays accurate for the two families it names. One rule, one home held: the old row's author rule merged into the 一条管三坑 clause rather than duplicated.
• The comment-path caveat is recorded exactly as measured and unattributed (4 comment writes, 1 marker absent) — read-back not exempted for comments.
• Fact 3 lands with the framing that matters: an inherited claim is not a reading; the false "search API unavailable" note is retired by a dated positive row.
• Ratchet payments audited: the quota row's curl boilerplate and the usage-wall trailer compress without semantic loss (both re-read in full).
• Gates: ratchet 134/134, id-lint clean, 6 derived families green with quoted verdicts at head 221640d.
• Governed surface (.claude/**) ⇒ terminal three-step: review recorded here; the PR stays draft, visibly awaiting the maintainer's merge; the round report lists it under "awaiting a human merge".

---

Generated by Claude Code