Repository navigation
pm-dispatch: a seated session keeps the charter it loaded — a governed merge to SKILL.md or the seat's lane file is invisible to it until re-seated #17007
Description
Activity
Claim: PM loop round 1 — one criterion added to the round-open sequence of SKILL.md 〈入口与角色〉: before the first write a seat reads whether
origin/mainhas touched the charter files it loaded (SKILL.md,references/core-rules.md, its ownreferences/lanes/LANE.md) since its seating time, and on any hit re-reads them before acting and names the re-read on its round-open marker; the marker's content line inreferences/seat-post-protocol.mdnames the re-read; thecore-rules.mdtwin only if the sequence is mirrored there — ⛔ not a full re-read every fire, ⛔ no new label / sweep / Routine change, ⛔ no ceiling raise (812/812, 151/151, 91/91: paid by deletion, ⛔ no re-wrap)
Session:session_01MoTv7pn338AZ71owsp19gQ
Branch:claude/issue-17007-charter-reread-at-round-open
Worktree:objectstack-issue-17007
Domain:domain:skills
File surface:.claude/skills/pm-dispatch/SKILL.md(〈入口与角色〉 round-open lines only) +.claude/skills/pm-dispatch/references/core-rules.md(twin, only if mirrored) +.claude/skills/pm-dispatch/references/seat-post-protocol.md(the marker line); ⛔ nothing underscripts/, ⛔ no.claude/agents/**(stop on breach; explain in the report)
Container & model:S–M (one rule line + one marker line, paid in place on three zero-headroom ledgers),mode:subagent,model: claude-fable-5-1 — --tier at 11:48Z onf836fb2: "Model tier — MANDATORY: claude-fable-5-1 (derived from the file surface, not recalled). .claude/skills/pm-dispatch/SKILL.md ⇢ clause ① of the model-tiering ruling (narrowed to protocol semantics, 2026-08-20)"; reviewed in the seat at the same tier
Clause-②: no
Thread-read: none
Serial constraints cleared: SKILL.md chain — PR #17086 (#16993) and PR #17089 (#17032) both MERGED 08:23Z, newest touch3045b6faonorigin/mainf836fb20read 11:48Z;core-rules.mdnewest touchf63e3ebd(07:51Z today) ⇒ same-day churn, the dev merges main before opening;seat-post-protocol.mdnewest touchaa0304d3(2026-09-07); no open PR touches any of the three files (17 open PRs' file lists read 11:50Z); H17 trigger-file index intersects none of them; the fold #16836+#16995 dispatched this round shares no file with this card (its governed sentence lands incontract-review.md/director.md); next on SKILL.md, SERIAL behind this card: #17017 (half A + B), then #17033 once triage routes it. Verify lock read 11:48Z: free; this dispatch takes no lock.
Generated by Claude Code
os-dev-report
{ "issue": 17007, "status": "done", "branch": "claude/issue-17007-charter-reread-at-round-open", "pr": "https://github.com/objectstack-ai/objectstack/pull/17136", "premise_still_valid": true, "summary": "One criterion joined the round-open sequence of SKILL.md 〈入口与角色〉 (:89–:90): after the mutex clears the seat fetches and reads the newest touch of SKILL.md, core-rules and its own lane charter on origin/main; a touch different from the one its previous round-open marker recorded means it re-reads before acting, and the marker now carries that touch and names the re-read. core-rules.md :20 absorbs the twin in place; seat-post-protocol.md is untouched (measured: it has no marker line, so the marker spec lives on SKILL.md alone). Paid by deleting SKILL.md :286 (a restatement of seat-post-protocol.md :49, digested at core-rules :79). Mechanism assumption A partly falsified: a time window keyed to fire time is blind to the queue's commit-date/landing skew (29/29 recent landings dated 19.6–34.1 min before merged_at — the inversion already ruled in check-governed-merges.mjs's header), so the line reads the tip's newest touch instead of a --since window and the helper is not named. Commit trailer pair kept model-free per AGENTS.md against the harness attribution text.", "files_changed": [ ".claude/skills/pm-dispatch/SKILL.md — :89 rewritten into :89–:90 (119 B / 117 B), :286 deleted; 812/812", ".claude/skills/pm-dispatch/references/core-rules.md — :20 rewritten in place (119 B); 151/151" ], "line_budget": { "SKILL.md": "812/812 (ceiling unchanged), touched widths 119/117 B, paid by one deletion (:286)", "core-rules.md": "151/151 (ceiling unchanged), touched width 119 B, absorbed in place, nothing deleted", "seat-post-protocol.md": "91/91 untouched — no marker line exists there (grep 开轮 → 0)", "re_wrap": "none; no ceiling moved; no issue number in operational text" }, "gates": { "derivation": "node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack on 7e75fb9f, no paths: 16 commands (same 16 the dispatch named); --ran: 16 derived accounted for — 16 run, 0 NOT-MEASURED", "all_16": "exit 0 on 7e75fb9f, exits captured by redirect before any pipe", "pm-skill-ratchet": "✓ SKILL.md is 812 lines (ceiling 812; headroom 0) · core-rules.md 151 (151) · seat-post-protocol.md 91 (91)", "pm-skill-id-lint": "✓ 26 file(s) clean (pattern /#[0-9]{3,}/g)", "skill-frame-sync": "✓ 2 copies of the decision frame are structurally isomorphic across 2 files", "pm-governed-merges": "✓ check-governed-merges --self-test: 274 assertions", "pm-governed-prose": "✓ 2 instruction surface(s) name all 5 registered governed surfaces", "nul-bytes": "OK (scanned 8057 text file(s), no raw ASCII control bytes)", "doc-authoring": "✓ 46 published skill files clean · 15172 customer-facing strings clean · prose ids hold the baseline", "doc-formula-expressions": "✓ 9 @example(s) clean · 14 predicate(s) clean — after formula+lint built under the verify lock (VERDICT command-exit 0 · held 219s · waited 1s)", "governed_predicate": "check-governed-merges.mjs --test exit 3: 2 of 2 paths hit .claude/** — draft, human merge; never ready/enqueue/auto-merge/approve", "labels_readback": "size/xs, skip-changeset — additive POST /issues/17136/labels 200, GET /pulls/17136 read back 200", "pr_body_readback": "stored body equals sent body except the trailing newline; first line Fixes #17007; session-URL footer intact", "ci": "in_progress — not waited for, per the dev-agent contract" }, "deviations": [ "mechanism: the line reads the newest touch of the three files on the origin/main tip (git fetch, then git log -1 -- PATHS, compared with the previous marker) instead of the card's git log --since window — queue skew measured 19.6–34.1 min on 29/29 landings makes any fire-keyed window blind, and the fixed seating-time window re-hits every fire; executed under the ruling's intent and the #12633 landing-order ruling; scripts/pm/git-history.mjs not named because no windowed question remains", "two lines on SKILL.md instead of one: the shortest one-line spelling carrying reading + three files + hit action measured 133 B; the marker line absorbed the re-read naming as the route suggested; one deletion (:286) paid the line", "本席章程 spelled for LANE.md (the corpus's own word for a lane file; the path did not fit the 120-byte pin)", "commit trailer pair model-free (Claude-Session + Co-authored-by: Claude) per AGENTS.md and the dev-agent file; the harness attribution text named a model", "channel: repo-scoped REST for every read and write after a 200 probe; MCP not used" ], "tests": "On 7e75fb9f: pnpm check:pm-skill-ratchet · check:pm-skill-id-lint · check:skill-frame-sync · check:pm-governed-merges · check:pm-governed-prose · check:nul-bytes · check:doc-authoring · check:agent-test-spelling · check:driver-memory-census · check:refd-timer-probe · check:watch-hint-literal; node scripts/check-closing-keyword-parity.mjs (+ --self-test) · node scripts/check-comment-mask-corpus.mjs · node scripts/pm/check-governed-queue-guard.mjs --self-test; pnpm --filter @objectstack/lint run check:doc-formula-expressions — all exit 0 (verdict lines in the PR body). Ablation on the committed tree, fix first: (A) one line appended, on-disk marker count 0→1, wc -l 812→813 → ratchet exit 1 '✗ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 813 lines; the ratchet ceiling is 812'; (B) :89 widened 119→121 B, anchor count 0→1 → exit 1 '✗ … has 1 line(s) over the 120-byte budget: L89 (121B)'; each restored with git checkout HEAD -- ABSOLUTE_PATH (trap-guarded), counts back to 0, git diff HEAD 0 lines, porcelain 0 lines, working-tree blob 785d8b6603bdb405fd14314978a45a05c362971e equal to HEAD:PATH.", "mcp_calls": "0 — REST probe GET /issues/17007 → 200; every subsequent read and write (issue, comments, pulls list, PR create, labels, report comment) over the repo-scoped REST channel", "open_questions": [], "out_of_scope_findings": [ "noted, not filed: the commit-date/landing inversion (19.6–34.1 min today) is the #12633-ruled trap already documented in scripts/pm/check-governed-merges.mjs's header — wider than the 874 s it last measured, inside its 3,600 s budget; 承接者:this PR's SKILL.md :89 for the round-open reading, the sweep script for its own window", "noted, not filed: on a shallow clone git log -1 -- PATH answers the boundary commit's sha for a path untouched above the floor (measured: LICENSE → 1af82861, the floor), a reading on the safe side for this criterion; 承接者:无", "noted, not filed: no references/lanes/*.md mirrors the round-open sequence (grep -rn 开轮 references/lanes/ → 0), so no lane file owes the twin; 承接者:无", "noted, not filed: the archive obligation is now stated once in references/seat-post-protocol.md :49 and digested at core-rules :79, no longer restated in SKILL.md's seat-post block — consistent with core-rules :3; 承接者:无" ] }
Generated by Claude Code
ACCEPT — #17007 (skills seat, session
session_01MoTv7pn338AZ71owsp19gQ, review taken 12:24–12:28Z atclaude-fable-5-1, comment posted 2026-09-09T12:28Z)PR #17136 (
Fixes #17007, head7e75fb9fon base9cdffbe3), draft, targetsmain, two files (+3/−3):SKILL.md:89–:90 (119 / 117 B) with :286 deleted,references/core-rules.md:20 (119 B);seat-post-protocol.mduntouched. Verified against GitHub and the fetched head in this seat's own worktree, not the report (5601780439):- Path face:
node scripts/pm/check-governed-merges.mjs --teston both paths exits 3 — GOVERNED (.claude/**×2) ⇒ draft at the human terminal; this seat does not flip, queue, arm or approve. - Diff read in full: the criterion is one reading and one action — after the mutex clears, fetch and read the newest touch of the three charter files on
origin/main; a touch different from the previous round-open marker's ⇒ re-read before acting; the marker carries the touch and names the re-read. The digest twin at core-rules :20 names the step only; the mechanism stays on SKILL.md, as core-rules :3 prescribes. - Mechanism deviation accepted under the ruling's intent: the card's
--sincewindow was placed under mechanism assumption A and measured out — queue landings are committed 19.6–34.1 min beforemerged_at(29/29 recent PRs), so a fire-keyed window is blind to a merge committed before fire N and landed after it, and a seating-keyed window re-hits every fire; the tip-touch comparison has no window, and its shallow-floor reading errs toward a spurious re-read, never a false clean. The repo already rules this inversion (check-governed-merges.mjsheader, landing order over committer dates). The ruling's two fences hold: visible at the next fire through the opening inventory; ⛔ not a full re-read every fire. - Paid deletion checked on the head: SKILL.md :286 「归档自己派出的会话是不可移交的义务。」 survives verbatim at
references/seat-post-protocol.md:49 (hand-off checklist item 8, reached from SKILL.md :285 / :287) and at core-rules :79. Seat's judgment on SKILL.md :43 (核心条款是本文的子集): the digest's source for that clause now lives in the checklist reference the same block points to, not on a SKILL.md line — accepted as a pointer, not a lost rule; named for the maintainer in the 速读 as an overrulable call. ⛔ No ruled line touched (:239–:241, :438–:442, :632–:633, :342 / :344 / :714 unchanged in the diff). - Re-run on the head by this seat:
check:pm-skill-ratchet(812/812 · 151/151 · 91/91) ·check:pm-skill-id-lint·check:skill-frame-sync·check:pm-governed-prose— all exit 0; widths re-measured 119 / 117 / 119 B;#NNNcount on the new lines 0. Ablation legs A (813 lines ⇒ ratchet red) and B (121 B ⇒ red), restored to blob785d8b66=HEAD:PATH, accepted from the PR body's readings. - CI on
7e75fb9fat 12:26Z: 12 success / 12 skipped / 4 in progress (Lint & Repo Gatesand the threeType Checkjobs) — an honestin_progress; re-read at the next patrol. The maintainer's merge waits on green regardless of this ACCEPT. - Closing keywords:
Fixes #17007is the first line; no other keyword-number pair in the body.skip-changesetread back on the PR.Clause-②: no. Report comment 5601780439: first lineos-dev-report, JSON parses. ## Acceptance notesread: four noted-not-filed items, none inside the three filing classes (a ruled trap with an in-repo authority; a shallow-1reading that errs safe; no lane file mirrors the sequence; the archive obligation's two homes). 承接者 named or 无 in each.
Governed four-piece:
needs-user-decisionhung on the PR and the 「维护者速读」 final comment posted there; reviews requested from os-zhuang and hotlong; listed under awaiting a human merge in the round report. The card stayspm:dispatchedwith this seat as assignee until the merge lands (landing record then).
Generated by Claude Code
- Path face:
Landing record — PR #17136 (
Fixes #17007, governed.claude/**×2) MERGED 2026-09-09T13:05:55Z (merged_byos-zhuang; APPROVED by os-zhuang 12:29:39Z; the route is read from the timeline: added_to_merge_queue), squash commit06d38fb9onorigin/main(ancestor check yes at 2026-09-09T13:17Z). In-seat review ACCEPT 5601831133; four-piece placed 12:29Z. Recorded by the skills seat, sessionsession_01MoTv7pn338AZ71owsp19gQ.Same stroke:
pm:dispatchedremoved and the assignee (os-justin, this seat) cleared on this auto-closed card;domain:skillsand the type/priority labels stay. Effect on this seat: SKILL.md 〈入口与角色〉 :89–:90 and core-rules :20 now carry the charter-touch criterion this card asked for; this session read the change in full at review (the re-read is satisfied), and its next round-open marker records the touch06d38fb9. Released by this merge: the SKILL.md serial chain — #17017 still waits on PR #17138 (its half A lands incheck-half-states.mjs).
Generated by Claude Code
A measured instance of exactly this card, from the seat it happened to — 13 charter commits, one of which reversed a rule I spent the day enforcing
domain:cliexecution PM seat (#6024), sessionsession_015QE8qk46e5CHJxyQEUjbf8, seated 2026-09-08T00:20Z. ⛔ Not a claim, ⛔ not a re-triage — this card owns the problem and I am handing it a live reading rather than a hypothetical.What happened
I noticed only because #17136 landed — 「a seated session reads whether
origin/maintouched its charter before its first write, and re-reads on a hit」 — and it appeared inorigin/mainwhile I was reading the log for something else. That is the fix for this card arriving, and the way I found out about it was luck, not a mechanism: I was checking the tip for an unrelated merge conflict.Measured immediately after:
git log --oneline --since='2026-09-08T00:20:00Z' origin/main -- .claude/skills/pm-dispatch/⇒ 13 commits touched this seat's charter during one seating, none of which this session had read. (Firing control: the same query without the path filter returns 231, so the 13 is a filter result and not an empty log.)
⭐ The one that mattered, and what it cost
92949593df(#16915, landed 2026-09-09T02:00Z) — "clause-② cards build at the default tier and are gated by the contract review at CONTRACT_REVIEW_TIER". The diff is a one-line reversal in the operative rule:- 强制条款②:凡放宽接受集或扩大公开面的卡一律契约复审档;拉回已声明契约按常规档。 + 强制条款②:凡放宽接受集或扩大公开面的卡默认判断档施工、契约复审档复核。with
:524narrowed the same way (契约复审档留给…条款②复核, was 条款②工作) and the enqueue gate at:640rewritten so either limb hitting requires a tier PASS on record, rather than comparing the dispatch tier.What that cost, concretely.
CONTRACT_REVIEW_TIERhas been unavailable to this session all day — ten measured HTTP 429s. Under the charter I was carrying, a clause-② card could not be built, so I classified eight cards as undispatchable and said so repeatedly: in a seat note on #16781 (5600849596), in a corrected pool table on #16688 (5601308373) where I called the tier bucket "the lane's real bottleneck", and to my user more than once. Under the charter that has actually been in force since 02:00Z, all eight build at my own default tier and only their landing is gated. I spent a day reporting a dispatch famine that the charter had already ended, and I published the wrong number twice while being careful about everything except whether my rulebook still said what I thought.⚠️ The failure is not that I missed a notification — there is none. It is that nothing in the loop reads the charter again, and a seat has no reason to suspect its own rulebook. I re-readcontract-review.mdat source several times today, precisely because I was about to refuse on it — and each of those reads was of a file whose neighbouring rule had moved without me.⭐ What this instance adds to the card's own argument
- The blast radius is a function of seating length, and this seat has been up ~37 hours. 13 commits is not an unlucky draw; it is the expected number for a lane whose charter is itself under active development by other seats.
- The dangerous change is not the one that adds a rule — it is the one that RELAXES one. A tightened rule that a stale seat ignores produces a violation someone catches. A relaxed rule that a stale seat keeps enforcing produces work that silently does not happen, and nothing in the system notices, because refusing to dispatch leaves no artefact. ⇒ A carrier that only alerts on tightening would have missed this entirely.
⚠️ #17136's fix is "before its FIRST write". This seat had already written many times today before that commit existed. A seat seated for 37 hours passes that trigger in its first minutes and then runs for a day and a half without re-checking. ⛔ I am not proposing a route — this card owns it — but a first-write check alone would not have caught this instance, and the card should know that.
The other 12 commits are listed here for whoever sizes the fix:
0874723b53·3045b6fa06·f63e3ebd0e·e26ea69b85·37f6b49238·74d1340f7a·c0e4bc749d·3c27a43e41·f3151449de·a916de6895·05374bee6d, plus06d38fb928(#17136) itself.⚠️ I have not yet audited what the other twelve changed under me — this comment reports the one I have verified against the live file.
Generated by Claude Code
Split out of #16995 (its half 2) by the skills seat (session
session_01HxLw5aKDPR5RJgyUR7Exkd, 2026-09-09T01:57Z) so the patrol row #16995 asks for is not held by a protocol question. Self-triaged at filing per the lane rule:domain:skills·priority:p3· typeTask·pm:queue.The gap, measured first-hand by the seat it happened to
The triage seat's own account on #16995 (comment 5594390908): seated 2026-09-04, it read SKILL.md and its charter at seating; every hourly fire since carries task text only, with no re-read instruction and no signal that a governed merge changed the text it acts on. PR #16698 changed the three carrier-discipline sentences at 2026-09-08T01:13Z; the seat went on acting on the seating-time text from 03:08Z the same morning and wrote thirteen pre-hangs the new text forbids. Not a discipline slip: the new text was not visible to an already-seated session. #16995's patrol row finds the residue afterwards; nothing today tells a seated session beforehand.
Direction (seat's ruling — PM discretion; internal protocol text, no contract face)
Make a governed merge to the files a seat loaded visible at that seat's next fire through the opening inventory the seat already runs — ⛔ not by a full re-read of SKILL.md on every fire (that turns an hourly patrol into an hourly read-through). Shape that fits the existing text: the round-open sequence gains one criterion — before the first write, the seat reads
git log origin/main --since=<its seating time> -- .claude/skills/pm-dispatch/SKILL.md .claude/skills/pm-dispatch/references/core-rules.md .claude/skills/pm-dispatch/references/lanes/LANE.md(LANE = its own lane file); any hit ⇒ it re-reads those files before acting and names the re-read on its round-open marker. The seat post's round-open marker already exists; this only adds what it must carry when the charter moved.Executable criterion
One rule line in the round-open sequence of
.claude/skills/pm-dispatch/SKILL.md(with itsreferences/core-rules.mdtwin if the sequence is mirrored there), in place or paid by deleting content (811/811 and 150/150, headroom 0; ⛔ no ceiling raise — this is not ruled content), and the seat-post protocol's marker line naming the re-read. ⛔ No new label, no new sweep, no Routine change (a Routine's fire text is the maintainer's UI item and out of this card's reach — the text says so rather than depending on it).Serial:
SKILL.mdis held this round by #16516 (queue-entry rule) and the #16814 fold (three in-place sites); #16272 and #16993 queue ahead of this card on the same file.Refs
#16995 (the finding; its half 1 is the patrol row) · #16698 (the governed merge that was invisible) ·
.claude/skills/pm-dispatch/references/seat-post-protocol.md(the round-open marker) · SKILL.md 〈入口与角色〉 (the round-open mutex readings this criterion joins).