Skills optimization flight — skills/objectstack-pm-dispatch: RESTRUCTURE (≈ −4,255 tok, −29%) — monorepo-only lore inert in a customer project, the report JSON printed twice, and four drifts from the internal dev contract (report double-channel, `premise_still_valid`, the stash ban, the `Session:` line)

[os-litant]

Member card of the skills catalog optimization program #14292 (maintainer mandate 2026-09-02, verbatim: 「审核所有的 skills，进行全面的优化。」). Filed by the skills lane seat (session session_01LraLgQVGq8egUwfYZpbYt1). Read-only audit at objectstack origin/main a59f78d. The full findings table is the audit record: the dev posts it verbatim as the first comment on this card at claim time (seat scratchpad audit/objectstack-pm-dispatch/findings.md).

⚠️ Fence: this is the PUBLISHED process skill. .claude/skills/pm-dispatch/** and .claude/agents/os-dev.md are a separate corpus under audit #13597 — read them only as the comparison oracle, ⛔ never edit them in this flight, ⛔ never in the same PR.

Audit summary

One file, SKILL.md, 1,086 lines, 14,549 tokens, headroom 0 — 7.7% of the bundle, the 2nd-largest published SKILL.md, 1.75× the other process skill. Teaches no ObjectStack surface (0 TypeScript fences). Customer decision points covered: 5 (label state machine; safe claim; dispatch with a binding template; review-against-GitHub and land; escalate vs decide) ≈ 2,900 tokens per decision. Customer-inert or duplicated text measured at 4,445 tokens (30.6%) before any restructure. Verdict RESTRUCTURE: zero progressive disclosure, zero evals, two blocks duplicated inside the file (the report JSON at 869-882 / 894-908; the sanitizer trap at 346-365 / 884-888), a 2,376-token verbatim paste-block paid in every session, and a second job (upstream reporting) restating objectstack-platform :59-70.

Top findings

id	span	proposal	delta
PMD-B-01	SKILL.md:729-891, 916-986	MOVE-TO rules/dev-template.md (DEFERRED, #14296 items 1+4)	−3,258 from entry
PMD-C-01	SKILL.md:916-986 vs objectstack-platform :59-70	DELETE the upstream-reporting job; keep ≤250-token residue pointing at objectstack-platform	−882
PMD-E-05	SKILL.md:987-1033	REWRITE-AS-CONSTRUCT — shard registry conclusion is one line for a single-repo project	−580
PMD-E-02	SKILL.md:416-452	RETIRE-SURFACE — "N independent implementations" is the platform's compiler problem, not an app's	−565
PMD-E-03	SKILL.md:591-618	DELETE two non-escalation bullets (ADR-0049 "arm"/governance shapes)	−508
PMD-D-01	SKILL.md:192-217	REWRITE-AS-CONSTRUCT — sweep rationale, cross-seat producers, anecdote	−380
PMD-D-02	SKILL.md:507-528	REWRITE-AS-CONSTRUCT — the 8-card baseline printed and disclaimed	−290
PMD-D-04	SKILL.md:654-706 vs 826-857	REWRITE-AS-CONSTRUCT — four-axis frame stated twice	−450 (MED)
PMD-E-04	SKILL.md:769-786	REWRITE-AS-CONSTRUCT — design spec of a lock script the customer does not have	−250
PMD-B-03	SKILL.md:346-357	DELETE sanitizer specimens; rules at 359-365 carry the decision	−200
PMD-B-02	SKILL.md:894-908	DELETE — the same report JSON printed twice	−115
PMD-D-03	SKILL.md:483-495	DELETE check-in anecdote; keep the two rules at 496-506	−170 (MED)
PMD-A-01	SKILL.md:10-14	DELETE the "report a platform bug" trigger from the description	−30
PMD-G-01	SKILL.md:557-559 vs 1051	DELETE — "the PM's own tooling PRs" contradicts "The PM writes no files"	−25

Drift vs the internal contract (sync in this flight, paid by the deletions above)

• PMD-F-01 :455-457 "a lossless channel" — the internal contract requires the report TWICE, GitHub first, in both modes, first line the literal plaintext os-dev-report (⛔ not an HTML comment), because a container restart killed three devs mid-final-message. Also a falsehood: :456-457 is contradicted 17 lines later at :474-475.
• PMD-F-02 the template never asks the dev to re-check the issue's premises and the report has no premise_still_valid; internally a no-PR report with premise_still_valid: false is a first-class deliverable.
• PMD-F-03 grep -c stash = 0, yet the template ships the exact topology the trap lives in (worktree-first + "parallel agents share ONE container"); both repos carry the ban at Prime-Directive level.
• PMD-F-04/F-05 the claim shape omits the Session: line (branch-only identity under-determines in cloud mode); :367-369 asserts devs push early while the template pushes fourth — the reclaim predicate can fire against a live agent. Make the template push the empty branch first.
• Falsehood :1036-1039 "the template enforces … a container-wide verification lock" — the template delegates the mechanism to the host project; should read requires.

Flight scope

IMPLEMENT (same-file, shrink-only): all rows above except PMD-B-01 and PMD-H-01; the four drift syncs (paid by PMD-E-02 / PMD-E-04 / PMD-B-02 as the auditor sized them); PMD-C-01 as delete + ≤250-token residue pointing at objectstack-platform SKILL.md:59-70; the description trigger edit (regenerate skills/README.md).

DEFER (pending #14296 items 1 and 4): PMD-B-01 (split the template into rules/dev-template.md), PMD-H-01 (evals).

Flight constraints (binding)

• ONE draft PR, first line Fixes #<this card>; governed ⇒ stays draft; review requests are the seat's step.
• Token ratchet: SKILL.md may not grow; additions paid by deletions in the same file; ⛔ re-wrap is not payment; ⛔ no ceiling raise; ⛔ no new files; ⛔ do not touch the ratchet script.
• ⛔ Never edit another package's files; frontmatter edits ⇒ regenerate skills/README.md (pnpm --filter @objectstack/spec gen:skill-docs).
• Gates: node scripts/check-skills-token-ratchet.mjs, pnpm check:skill-compatibility, pnpm check:skill-identifier-liveness, plus node scripts/pm/dispatch-gates.mjs --commands <changed paths>; record the head sha.
• PR body: per-item 落点 | before | after list keyed by finding id; token delta; Clause-②: no (no contract surface).

Refs: #14292 · #14296 · #13597 (the internal corpus, separate audit).

[os-litant]

Claim: PM loop round 2
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14300-skills-pm-dispatch-optimization
Worktree: objectstack-issue-14300
Domain: domain:skills
File surface: skills/objectstack-pm-dispatch/SKILL.md ONLY (shrink-only; no file deleted ⇒ ⛔ no edit to scripts/check-skills-token-ratchet.mjs); skills/README.md + content/docs/ai/skills-reference.mdx regenerated only because the drift gate demands it after the PMD-A-01 description edit; ⛔ .claude/skills/pm-dispatch/** and .claude/agents/os-dev.md are the comparison ORACLE, read-only, never in this PR; ⛔ no other package (stop on breach; explain in the report)
Container & model: M (one 14,549-token file, structural diet + five drift syncs against the internal contract), mode:subagent, model: fable — node scripts/pm/dispatch-gates.mjs --tier skills/objectstack-pm-dispatch/SKILL.md reads MANDATORY claude-fable-5 (clause ①, 2026-08-20 narrowing: the published PM skill carries two enforced copies of the decision frame and ships verbatim to third-party projects); no exit applies — this is not a one-line mechanical edit and quota is available
Clause-②: no — no contract surface (0 TypeScript fences, no accept/reject semantics); no needs:contract-review; the skill-face review is the seat's in-seat review at tier
Serial constraints cleared: the two frame-sync copies in this file (:643 "every option on the four fixed axes below." and :832 "Analyze every option on four fixed axes:") are gate-pinned against .claude/skills/pm-dispatch/SKILL.md and .claude/agents/os-dev.md — both copies STAY and neither block is edited, so PMD-D-04 is EXCLUDED from this flight (its only remedy removes a copy the gate requires); .claude/skills/pm-dispatch/SKILL.md (PR #14230) and .claude/agents/os-dev.md (PR #14229) are held by unmerged PRs — this flight syncs the five drifts to the contract text on origin/main and lists #14229's pending deltas as a follow-up sync, never edits them; PMD-C-01's residue points at objectstack-platform's "The App / Platform Boundary" section BY HEADING (the platform flight #14351 moved it and keeps it); no open PR touches skills/objectstack-pm-dispatch/**; in flight alongside #14299 (platform, patch round) and #14302 (query) on disjoint directories

Decision re-read: no maintainer ruling on this card; binding inputs = the mandate, the value gate, the ratchet discipline, and #14296 (items 1 + 4 defer PMD-B-01; item 2 defers PMD-H-01). Landing form: first line Part of #14300 (the deferred split and evals keep the card open).

---

Generated by Claude Code

[os-litant]

Claim: dev seat, dispatched by PM loop round 2 (mode:subagent)
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14300-skills-pm-dispatch-optimization (empty branch pushed first as the write-route probe: exit 0)
Worktree: objectstack-issue-14300 (base 96b627d1)
File surface: skills/objectstack-pm-dispatch/SKILL.md only, plus skills/README.md and content/docs/ai/skills-reference.mdx as generator output after the PMD-A-01 description edit. Assignee untouched. .claude/skills/pm-dispatch/** and .claude/agents/os-dev.md read as the comparison oracle only.

Next comment on this card is the audit record (audit/objectstack-pm-dispatch/findings.md, first line audit-record), posted with a transliteration key for the angle-bracket-shaped fragments.

---

Generated by Claude Code

[os-litant]

audit-record — skills/objectstack-pm-dispatch (read-only audit at a59f78d)

Posted by the dev seat at claim time; this is the finding-by-finding spec for the flight on #14300.

Transliteration key (deviation from a byte-verbatim post, disclosed). The GitHub comment sanitizer is measured to eat angle-bracket / HTML-comment-shaped fragments, and this seat has no raw-REST read path to verify byte survival (repo-scoped REST returns 403 for this session). The 18 angle-bracket fragments on 13 lines of the file are therefore spelled without angle brackets, as the platform flight did on #14299. Nothing else is changed — every finding id, span, evidence, proposal, delta, gate verdict, confidence and human-floor flag is byte-identical to audit/objectstack-pm-dispatch/findings.md:

on disk	in this comment
HTML comment wrapping os:check	os:check HTML-comment markers
HTML comment wrapping BEGIN GENERATED: skills …	the BEGIN GENERATED: skills HTML-comment marker
HTML comment wrapping dev-report	the dev-report HTML-comment marker
angle-n	N
angle-url or null	URL or null
angle-path	PATH
angle-domain	DOMAIN
angle-repo dash issue dash angle-n	REPO-issue-N
claude/issue dash angle-n dash angle-slug	claude/issue-N-SLUG
session underscore angle-id	session_ID
angle-branch	BRANCH
angle-upstream hash angle-n	UPSTREAM#N

---

Audit — skills/objectstack-pm-dispatch (published, read-only)

Pinned ref: objectstack origin/main = a59f78d. Verified the working file is byte-identical
to the pinned blob (git show a59f78d:skills/objectstack-pm-dispatch/SKILL.md | md5sum ==
md5sum of the checkout copy = 3948bdb8c73dba124ebc2e61a832d1db). All spans below are at
that ref. Local HEAD is a39b02a, but git diff a59f78d HEAD -- skills/ is empty.

Package facts (measured)

Fact	Value	Source
Files	1 (SKILL.md) — no rules/, no references/, no contracts/, no evals/	ls skills/objectstack-pm-dispatch/
Tokens	14,549 (ceiling 14,549; headroom 0)	node scripts/check-skills-token-ratchet.mjs
Bundle share	14,549 / 187,854 = 7.7 %	same run, "bundle total" row
Rank	2nd largest SKILL.md (after objectstack-ui 25,441); larger than objectstack-data 13,892 and objectstack-platform 12,984	same run
vs sibling process skill	1.75× objectstack-upgrade (8,333)	same run
Fenced blocks	16 total: 3 bash, 2 json, 1 text, rest bare	grep -nE '^{3,4}'`
TypeScript fences	0; os:check HTML-comment markers 0	same grep; gate rule read from packages/spec/scripts/check-skill-examples.ts:18-23
evals/	absent; 8 of 11 published packages have one	ls -d skills/*/evals

Because the package contains zero ts/tsx fences, dimension E's type-check sweep has no
surface: nothing in it can be reached by check-skill-examples.ts, and nothing needs a
marker. Recorded as a measured negative, not a finding.

Section budget (measured, ceil(bytes/4))

Section	Span	Tokens
frontmatter	1-29	366
intro	30-49	224
Quickstart	50-78	216
Configuration	79-117	521
Arguments	118-133	158
State model	134-170	454
§0 Backlog sweep	171-237	980
§1 Fetch candidates	238-261	249
§2 Triage	262-306	688
§3 Select the batch	307-319	164
§4 Claim	320-376	761
§5 Dispatch	377-415	381
— One semantics, N implementations	416-452	565
— Dispatch backends	453-470	244
§6 Collect	471-528	902
§7 Review each report	529-564	491
§8 Escalate	565-647	1324
— Four-axis frame	648-706	925
§9 Round report	707-715	79
Stop conditions	716-728	99
Developer-agent template	729-891	2376
Report contract	892-915	194
Upstream reporting	916-986	882
Multiple PMs / shards	987-1033	656
Resource discipline (PM)	1034-1048	158
Guardrails	1049-1065	196
Adapting this loop	1066-1086	307

The customer's five real decision points

Derived from what a single-repo ObjectStack app project with no internal tooling can act
on, not from the doc's own headings:

1. Provision the three labels and read them as a state machine (134-170, 151-159).
2. Claim an issue so two agents cannot collide (320-343, 371-375).
3. Compose a dispatch prompt with a binding operating template (377-415, 729-891).
4. Review the returned PR against GitHub and land it (529-564).
5. Decide what to escalate versus decide under a veto window (565-595, 648-706).

Everything measured below as monorepo lore changes none of these five.

---

Findings

PMD-A-01 · TRIGGER PRECISION · skills/objectstack-pm-dispatch/SKILL.md:10-14

Evidence. The frontmatter advertises a second, unrelated job: "Ships the developer-agent
operating template … and the upstream-reporting procedure for platform defects an app
project finds. Use when asked to … or to report a platform bug found while building an
app." A customer agent that trips over a platform defect while authoring metadata loads
14,549 tokens to reach the 882-token §"Upstream reporting" (916-986), of which ~600 restates
skills/objectstack-platform/SKILL.md:59-70 (see PMD-C-01). skills/README.md's index row is
generated from this same description (the BEGIN GENERATED: skills HTML-comment marker, README:29-45), so
the mis-routing is mirrored into the catalog automatically.
Proposal. MOVE-TO skills/objectstack-platform (the ~250 tokens of unique procedure) and
delete both the "Ships … upstream-reporting procedure" clause and the "or to report a platform
bug" trigger; regenerate README via pnpm --filter @objectstack/spec gen:skill-docs.
Delta. −30 (frontmatter only; the section itself is counted in PMD-C-01).
Gate. ① Yes — it decides which package a customer loads for a bug report. ② Yes, deletion.
③ N/A (frontmatter). Confidence HIGH. Human floor YES (changes a published trigger).

PMD-A-02 · TRIGGER PRECISION · skills/README.md:94-110

Evidence. "Cross-skill routing" carries 4 bullets, all pairs of domain skills
(data/automation, automation/ui, formula, platform/data). Nothing routes between the two
process skills, and nothing tells an app author when a process skill applies at all. Both
process packages sit in the generated index (README:44-45) with no disambiguation, and their
descriptions overlap on the word "process" and on escalation-to-owner language
(objectstack-upgrade frontmatter: "work the semantic residue … to a decision with the
project's owner").
Proposal. ADD one routing bullet ("Delivery loop vs protocol upgrade …"). Payment must
come from README itself; the generated index block (29-45) is generator-owned and cannot pay.
Delta. +25 unpaid — flagged, not recommended without a maintainer ruling.
Gate. ① Yes. ② No. ③ Bullet in an existing list. Confidence MED. Human floor no.

PMD-B-01 · STRUCTURE · whole file (1-1086)

Evidence. One 14,549-token file with no progressive disclosure at all. Two blocks are
needed only at a specific moment yet are paid in every session: the developer-agent template
(729-891, 2,376 tok) is a verbatim paste-block consumed only when a dispatch is actually
composed, and §"Upstream reporting" (916-986, 882 tok) is a different job entirely.
skills/README.md:57-64 lists rules/ as an available optional directory for any skill; the
"process skill carries SKILL.md alone" sentence (README:66-69) is scoped to references/
("gen:skill-refs only visits skills listed in its SKILL_MAP"), so it does not block a
rules/ split.
Proposal. MOVE-TO rules/dev-template.md (the paste-block, linked from §5) and — with
PMD-C-01 — out of the package for upstream reporting.
Delta. −3,258 from the always-loaded entry file (a new rules/ file takes its own ceiling).
Gate. ① Yes — it decides what a PM pays before it has dispatched anything. ② Partly
(relocation, not deletion). ③ Yes, the template stays a construct. Confidence HIGH.
Human floor YES (restructures files; new ratchet ceiling).

PMD-B-02 · STRUCTURE · SKILL.md:869-882 and 894-908

Evidence. The same 8-key report schema is printed twice, 12 lines apart, differing only in
placeholder-vs-literal values ("issue": N vs "issue": 123; "pr": "URL or null" vs
"pr": "https://github.com/acme/hotcrm/pull/456 | null"). Only 909-914 adds new rules
("open_questions must be non-empty when status is needs_decision"; "out_of_scope_findings
should already be filed …").
Proposal. DELETE the fence at 894-908; keep 909-914 as a note under the template.
Delta. −115. Gate. ① No decision is changed by the second copy. ② Yes. ③ Construct
preserved. Confidence HIGH. Human floor no.

PMD-B-03 · DENSITY · SKILL.md:346-357 (232 tok), duplicated in substance at 884-888 (73 tok)

Evidence. The sanitizer trap is taught twice. 346-357 spends 12 lines on measured
specimens — "the dev-report HTML-comment marker came back as nothing at all, expected N to be 19 came
back as expected to be 19, git log -- PATH came back as git log -- " — plus three
lines on why this differs from step 0's truncation shape. The actionable rules are 359-365
("Write literal angle brackets as HTML entities … read it back"), and 884-888 restates the
whole trap inside the pasted template for the dev.
Proposal. DELETE 346-357, replacing with one clause on 359 ("measured: code spans are not
protection — an HTML comment came back empty").
Delta. −200. Gate. ① No — the rule at 359-365 already changes the decision; the
specimens change nothing. ② Yes. ③ Rules survive as a construct. Confidence HIGH.
Human floor no.

PMD-C-01 · CROSS-PACKAGE REDUNDANCY · SKILL.md:916-986 (882 tok) vs skills/objectstack-platform/SKILL.md:59-70

Evidence — both locations quoted.
pm-dispatch:918-924 — "The temptation is to make the app tolerate it and move on. Do not. A
workaround in the app hides the defect from the people who can fix it, ships a second source of
truth to every future author, and outlives the upstream fix by years."
platform:59-61 — "Business features belong in the app; capability belongs in the platform.
A missing default, a wrong diagnostic, a shape the spec refuses — the fix is upstream. Raise it
there; do not compensate for it here."
pm-dispatch:974-985 — park via Blocked-by: or "pin the current upstream version with the
unblock condition written down … A pin with no recorded unblock condition is a workaround with
better manners"; "'We worked around it' is never a resolution".
platform:62-69 — "A platform defect means waiting for the platform fix. No defensive
coding, no shape tolerance … Record the block against the platform issue so it is
machine-visible; before resuming, confirm the version you pin carries the fix (merged
upstream ≠ present on your pin) and re-run the defect's own reproduction."
Which anchor survives. objectstack-platform — it is the app-bootstrap package every app
project loads to stand the project up, it already carries the doctrine, and it is where an
agent already is when it hits the defect. The genuinely unique residue in pm-dispatch is ~250
tokens: derive the upstream repo from the dependency's package.json repository (925-930);
minimal repro + pinned versions + cite the contract (944-958); file as a guest — never apply
the upstream's queue labels, never assign, never add to their board (959-971).
Proposal. MOVE-TO skills/objectstack-platform (~250 tok of residue), DELETE the rest.
Delta. −882 here (+~250 there).
Gate. ① Yes — it decides which package owns the doctrine. ② Yes, mostly deletion. ③ The
survivor is a 3-bullet checklist. Confidence HIGH on the overlap, MED on the exact split.
Human floor YES (moves published capability text across packages).

PMD-D-01 · DENSITY / MONOREPO · SKILL.md:192-217 (433 tok)

Evidence. "Why the sweep is a disjunction and not one 'unlabeled' filter." followed by
two "standing producers" that exist only here: "Cross-seat transfer tickets (the handing-off
side applies the queue label as part of the transfer) and tickets filed mechanically by a
steward automation both carry pm:queue on arrival, while a single-producer discipline for
classification labels forbids those same producers from applying one", plus the anecdote "a
project that had published exactly that rule hours earlier still produced such a card the same
day". "Seat", "steward automation", "lane" (180) and "protocol posts" (215) are used as defined
terms and are defined nowhere in the package; an app project has none of them.
Proposal. REWRITE-AS-CONSTRUCT — keep the 3-item disjunct list (179-191), fold the ⚠️
scoping caution (209-217) to one line and the ⏳ updated_at floor (219-225) to one line, drop
the producers and the anecdote.
Delta. −380. Gate. ① No — the customer either runs the predicate or (per the ⚠️ at
209-214) must not; the rationale changes neither. ② Yes. ③ Yes, list survives.
Confidence HIGH. Human floor no.

PMD-D-02 · DENSITY · SKILL.md:507-528 (357 tok)

Evidence. The section states the number is not transferable and then prints it anyway:
"Two same-day samples from this repo's loop show why no single number can be inherited: four
text-only documentation cards landed at 93 / 96 / ~95 / ~110 min, while four mixed cards from
the same day spanned ~64 min to ~2 h 50 min (the two long ones waiting on CI). Eight cards, one
day, one toolchain …". The decision-changing content is the rule at 507-512 + 516-518.
Proposal. REWRITE-AS-CONSTRUCT — "A silence threshold is a collection boundary, not a
verdict of death. Before concluding a dispatched agent is dead, require one of: a direct status
query showing it dead, the host reporting the session stopped, or silence past a completion
baseline you measured in your own project."
Delta. −290. Gate. ① No — the data is explicitly disclaimed as non-inheritable.
② Yes. ③ Yes. Confidence HIGH. Human floor no.

PMD-D-03 · DENSITY · SKILL.md:483-495 (of the 404-tok block 483-506)

Evidence. "Measured three times in one day on this loop: a scheduled message still
delivered after its timer had been cancelled, carrying text two rounds out of date. One of
them read 'no branch and no report ⇒ judge the agent unreliable and dispatch a fresh one'; by
delivery time that agent had opened a PR which was already reviewed and accepted …". The two
rules that follow (496-506: open every check-in with "idempotent — re-read the state before
acting"; state criteria never conclusions; "cancellation is not a guarantee of non-delivery")
are general and genuinely decision-changing, and already carry the surprising fact.
Proposal. DELETE 483-495; keep 496-506 intact.
Delta. −170. Gate. ① No — 496-506 already changes the decision. ② Yes. ③ The two rules
are the construct. Confidence MED (the anecdote is the only place the cancelled-timer fact
is shown; it is however stated at 505-506). Human floor no.

PMD-D-04 · DENSITY · SKILL.md:654-706 (861 tok) vs 826-857 (574 tok)

Evidence. The four-axis frame is stated twice in one file, in near-identical wording. PM-side
axis ④: "Startup scope discipline. Do not grow the declared surface while the core is still
forming. 'We already shipped it' earns nothing: a shipped-but-unconsumed capability gets no
sunk-cost exemption." Template copy: "Startup scope discipline — do not grow the declared
surface: a shipped-but-unconsumed capability gets no sunk-cost exemption." The weight and
ranking rules are likewise duplicated ("Axis ② carries the highest weight — at least 50%",
697-706, vs "Long-term soundness carries the highest weight, at least 50%", 853-857). The
template copy must stay verbatim (it is pasted into a zero-context prompt).
Proposal. REWRITE-AS-CONSTRUCT the PM-side copy as a 4-row table (axis · question ·
evidence standard) plus the ranking/weight rule, pointing at the template block for full wording.
Delta. −450. Gate. ① No new decision from the second full prose copy. ② Partly. ③ Yes —
a table replaces prose. Confidence MED. Human floor no (no semantics removed).

PMD-E-01 · MONOREPO ASSUMPTION · SKILL.md:179-217

Evidence. Disjuncts 2 and 3 are keyed on a "classification label" — "none of the labels
your project reads as 'already routed' (area, component, owning team, lane)". This is the
monorepo's domain:* lane vocabulary (.claude/skills/pm-dispatch/SKILL.md:36 runs seats by
DOMAIN lane). The package's own ⚠️ at 209-214 makes the disjuncts self-disabling where
the label does not exist: "Where it does not exist, its absence is universal, so the disjunct
matches every open queue card in that repository and the sweep becomes its own noise source."
A customer app project has no classification labels, so 2 of the 3 disjuncts and their entire
rationale are inert by construction — the customer pays 433 tokens (PMD-D-01) to learn a
predicate it must not run.
Proposal. Folded into PMD-D-01; additionally define or drop "seat"/"lane"/"steward
automation"/"protocol posts".
Delta. counted in PMD-D-01. Confidence HIGH. Human floor no.

PMD-E-02 · MONOREPO ASSUMPTION · SKILL.md:416-452 (565 tok)

Evidence. "#### One semantics, N independent implementations". Its own applicability clause
names the platform's problem, not an app's: "that rule is implemented more than once, by
compilers or evaluators that share no code. Query filters, expression languages,
permission predicates and serialization formats all tend to grow this shape as a project adds
backends." An ObjectStack app project consumes one compiled query/expression surface from
@objectstack/spec; it authors no compiler and has no second evaluator to enumerate. The two
maintenance disciplines that follow (440-452, 201 tok — "The inventory is maintained by PR",
"Re-verify before pasting", frozen surfaces stay listed) are entirely about owning such an
inventory.
Proposal. RETIRE-SURFACE from the published package (it belongs to the internal corpus,
outside the published boundary).
Delta. −565. Gate. ① Changes no decision an app project can make. ② Yes, deletion.
③ N/A. Confidence HIGH. Human floor YES (removes published capability text).

PMD-E-03 · MONOREPO ASSUMPTION · SKILL.md:591-611 (398 tok) + 612-618 (110 tok)

Evidence. Two of the six "named non-escalation classes" are platform-governance shapes.
591-592: "A declaration silently dropped on a new arm, when a sibling arm already has a
ruling. When a declared key is silently ignored on one arm of a component and an
earlier ruling already made that same key a loud authoring error on a sibling arm, the new
arm joins the existing rejection set by default" — this is the enforce-or-remove /
declared = enforced machinery of this repo's spec work (the same vocabulary drives the
spec-property-retirement internal playbook), and an app project authoring metadata against a
fixed published spec has no "arm of a component" carrying declarations. 612-618: "Two
implementations of one operation … the side that already carries the governance —
authorization gates, user consent, de-duplication, audit trail — is the default survivor"
— a platform-runtime concern.
Proposal. DELETE both bullets; keep "Restore-invariant fixes", "Sequencing and dependency
ordering", "Verification strategy", and the needs_decision-relay bullet.
Delta. −508. Gate. ① No decision an app PM makes. ② Yes. ③ N/A. Confidence HIGH.
Human floor YES.

PMD-E-04 · MONOREPO ASSUMPTION · SKILL.md:769-786 (290 tok)

Evidence. Four correctness properties of a container-wide verification lock — "ONE shared
heavy-verify lock per container, so every agent contends on the same file"; "The acquisition
budget fits inside a SINGLE foreground tool call"; "A queue timeout is distinguishable from the
wrapped command's own failure"; "Hold duration is observable". This is the design spec of this
repo's scripts/pm/os-verify-lock.sh (the internal os-dev agent calls it by path). The
published text concedes the customer has none: "The MECHANISM is the host project's (its
wrapper, its lock path, its budget — put it in the conventions file)". A customer standing up
the loop on day one therefore reads 290 tokens instructing it to build lock infrastructure
before it may dispatch.
Proposal. REWRITE-AS-CONSTRUCT — one line: "Serialize build/test through a single
container-wide lock; queueing is normal, queueing with no end in sight is a finding — report it
naming the holder."
Delta. −250. Gate. ① The 4 properties change no decision for a project with no wrapper.
② Yes. ③ Yes. Confidence HIGH. Human floor YES (removes published guidance).

PMD-E-05 · MONOREPO ASSUMPTION · SKILL.md:987-1033 (656 tok)

Evidence. The whole section presupposes ≥2 repositories — which Configuration (79-117)
declares the non-default case ("With no config the loop runs against the current repository:
it is the only shard and the backlog", 60-61) — plus a shard registry issue: "Shard
ownership is registered, never assumed. A registry issue in the backlog repository records
which session owns which shard. A PM taking over a shard comments there as its FIRST action,
and again when handing off." That is this monorepo's pm:seat seat-post protocol
(.claude/skills/pm-dispatch/SKILL.md state model row: "pm:seat 座位登记贴 —— 协议载体") under
another name. Its own conclusion is one line: "Multiple PMs on the same queue: prohibited."
Proposal. REWRITE-AS-CONSTRUCT — one Guardrails bullet: "One PM per queue. Scale with
batch / mode:cloud, never a second scheduler; a second PM takes a whole other repository as
its shard and never claims outside it."
Delta. −580. Gate. ① For the documented default (single repo) it changes nothing.
② Yes. ③ Yes. Confidence HIGH. Human floor YES.

PMD-E-06 · MONOREPO ASSUMPTION (minor) · SKILL.md:553-556

Evidence. The landing path leads with the merge-queue case — "where a merge queue
exists, mark the PR ready and add it to the queue once every check is green (the queue rebuilds
against the current default branch, which is the sanctioned path)" — before the case every app
project is actually in. Properly conditional, so not a defect, but it front-loads the path the
customer does not have.
Proposal. Reorder so the serial-merge case leads. Delta. −30.
Gate. ① Marginal. ② No. ③ No. Confidence LOW. Human floor no.

PMD-F-01 · DRIFT vs INTERNAL CORPUS · SKILL.md:455-457 and 465-469

Evidence — published (retired form). "mode:subagent (default). … Reports come back
directly as each agent's final message — a lossless channel. Prefer this mode; it is simpler
and nothing can be lost between agent and reviewer." The issue-comment report is required in
cloud mode only (465-469).
Evidence — internal oracle (current form). .claude/agents/os-dev.md:20-22 and 351-359:
the report is delivered twice, GitHub first, in both modes — as an issue comment whose first
line is the literal plaintext os-dev-report (explicitly "⛔ 不用 HTML 注释"), then as the final
message; the reason is measured — a container restart killed three devs mid-final-message and
the comment was the only surviving copy, so zero work was lost. The published package's own
line 474-475 already concedes the channel loses work ("An agent that dies or returns malformed
output counts as status: \"blocked\"").
Proposal. ADD the both-modes comment rule + read-back to the template's final-message
section (~50 tok), paid by DELETE of the duplicate JSON at 894-908 (PMD-B-02, −115).
Delta. −65 net. Gate. ① Yes — it decides where the report is written. ② The addition is
paid by a same-file deletion. ③ Yes, it is one rule line on an existing construct.
Confidence HIGH. Human floor YES (changes published report semantics).

PMD-F-02 · DRIFT vs INTERNAL CORPUS · SKILL.md:729-891 (template) and 892-915 (contract)

Evidence. The PM-side stale-premise check exists (249-260: "An issue describes the
repository as of its filing date, and an active default branch moves fast"). The binding
template the dev actually receives never repeats it, and the report contract has no field for
the result. The internal oracle made this a numbered basic rule:
.claude/agents/os-dev.md:78-81 rule 6 — "issue 正文是线索,不是规格。动手前对 origin/main
核验其前提 … 一份带 premise_still_valid: false、附证据、无 PR 的报告是一等交付物;证伪
issue 是好运行" — and the internal report schema carries "premise_still_valid": true
(os-dev.md:430). Without it a dev on a dead premise has no sanctioned return shape: blocked
means infrastructure (published 858-860) and needs_decision means underspecified (826-830),
so the template's own incentives push a PR onto a falsified premise.
Proposal. ADD one template line ("Verify the issue's premises against origin/{default}
before your first edit; a no-PR report with premise_still_valid: false and evidence is a
first-class delivery") + the JSON key (~60 tok), paid by DELETE of PMD-E-02 (−565).
Delta. −505 net. Gate. ① Yes — it changes what a dev does when the issue is wrong.
② Paid by a larger same-file deletion. ③ Yes, one key on an existing JSON construct.
Confidence HIGH. Human floor YES (adds a published report field).

PMD-F-03 · DRIFT vs INTERNAL CORPUS · SKILL.md:745-750 + 763

Evidence. grep -c stash SKILL.md = 0. The template teaches the exact topology in which
the trap lives: worktree-first ("git worktree add --no-track ../REPO-issue-N … Never edit a
shared checkout — other agents switch its HEAD under you", 745-750) plus "parallel agents
share ONE container" (763). refs/stash lives in the common .git directory, so all
worktrees of one clone share one LIFO stack — which is why both repos carry the ban at
Prime-Directive level with a measured incident: /home/user/objectui/CLAUDE.md — "two agents
stashing in their own worktrees push and pop the same stack … pop reports success; the
only symptom is someone else's files appearing in your git status … it happened between two
parallel agents (objectui#3430) and cost both of them their in-flight changes."
The published package ships the topology without the guard.
Proposal. ADD two lines to template rule 1 ("Never git stash — the stack is shared
across every worktree of one clone. Use git diff > /tmp/wip.patch or a throwaway wip
commit", ~40 tok), paid by DELETE of PMD-E-04's lock properties (−250).
Delta. −210 net. Gate. ① Yes — it changes the command a dev reaches for. ② Paid by a
larger same-file deletion. ③ Yes, two bullets. Confidence HIGH. Human floor YES.

PMD-F-04 · DRIFT vs INTERNAL CORPUS · SKILL.md:336-338, 342-343

Evidence — published claim shape. "> Claim: PM loop round N / > Branch:
claude/issue-N-SLUG / > Worktree: REPO-issue-N", with the identity key being the
branch: "if an earlier claim comment with a different branch name exists, you lost".
Evidence — internal (current). .claude/skills/pm-dispatch/SKILL.md:535-541 makes
> Session: \session_ID`a non-optional line ("session ID 不可省"), keeps the literalClaim:first line as a maintainer-ruled machine predicate (2026-08-11), and names the comment **timestamp** the sole arbiter./home/user/objectstack/CLAUDE.mdstates the same at Prime-Directive level: "a claim is assign **plus a claim comment with your session ID and branch**". The published branch-only discriminator under-determines inmode:cloud— a mode the published skill itself supports — because the branch slug is derived from the issue, so two cloud sessions racing the same issue can generate the *same* branch name and both read "no earlier claim with a different branch". **Proposal.**ADDtheSession:` line to the claim shape (~15 tok), paid from PMD-D-01.
Delta. +15 (paid). Gate. ① Yes — it decides who yields in a race. ② Paid.
③ Yes, one row of an existing template. Confidence MED-HIGH. Human floor no.

PMD-F-05 · DRIFT / INTERNAL CONTRADICTION · SKILL.md:367-369 vs 800-810

Evidence. PM-side prose asserts a behaviour the template never asks for: "Developer agents
push their branch early — a remote branch is the hardest evidence of work in flight, and it
closes the gap between 'claimed' and 'a PR exists'", and the stale-claim reclaim at 371-375 then
reads that remote branch ("A claim older than ~24 h whose promised branch does not exist on
the remote and has no PR is presumed dead"). But the binding template's "Definition of done, in
order" (800-810) lists "Pushed: git push -u origin …" fourth, after implementation, tests
and the release-note artifact. A dev following the template has no remote branch for most of
its run, so the reclaim predicate can fire against a live agent. The internal oracle closes
this: .claude/agents/os-dev.md rule 1 — "建好分支后的第一个动作:先把空分支推上去(任何
编辑之前 git push -u origin BRANCH)" — doubling as a write-credentials probe (403 ⇒ stop and
report blocked).
Proposal. ADD "push the empty branch immediately after creating it (also a write-access
probe: a 403 here is blocked, not a retry loop)" to template rule 1 (~25 tok), paid from
PMD-D-02.
Delta. +25 (paid). Gate. ① Yes — it removes a contradiction that can trigger a wrongful
reclaim. ② Paid. ③ Yes. Confidence HIGH. Human floor no.

PMD-G-01 · CONSISTENCY · SKILL.md:557-559 vs 1051

Evidence. Guardrails: "The PM writes no files." (1051). Step 7 ACCEPT: "This applies to
developer-agent PRs dispatched by this loop only — the PM's own tooling PRs stay with the
maintainer." (557-559). The second sentence presupposes PM-authored PRs the first forbids. It is
residue of the internal exception the published copy dropped:
.claude/skills/pm-dispatch/SKILL.md — "PM 不写文件、不写代码。 唯一例外:维护者逐 PR
明示授权的 .claude/ 内部工具 PR".
Proposal. DELETE the "the PM's own tooling PRs" clause.
Delta. −25. Gate. ① Yes — it removes an apparent licence to author PRs. ② Yes. ③ N/A.
Confidence HIGH. Human floor no.

PMD-G-02 · CONSISTENCY (measured negative) · whole file

Evidence. Retired-spelling sweep run per the rubric (silently|ignored|dropped|legacy| deprecat|\bM9\b|v5\.0|\bOLD\b|\bNEW\b|17\.[0-9]|project): no residue. All 14 "silently"/
"ignored"/"dropped" hits are ordinary prose about behaviour (105, 450, 479, 592, 974, 1008 …);
line 754's "NEW, UNASSIGNED issues" is English emphasis, not the retired CEL NEW token; the
only version reference is current-major and internally consistent — 979: "pinned at 17.2.0;
unpin when UPSTREAM#N ships in >= 17.3". No project-as-workspace usage.
Proposal. none. Delta. 0. Confidence HIGH. Human floor no.

PMD-H-01 · EVALS · package has no evals/

Evidence. ls -d skills/*/evals returns 8 directories (ai, api, automation, data, i18n,
platform, query, ui). Both process skills — this one and objectstack-upgrade — have none.
This is the second-largest published package (14,549 tok, 7.7 % of the bundle) with zero
graded rubric, while objectstack-data (13,892) ships one and objectstack-query (5,443) ships
a 558-token one. Its decisions are unusually gradeable: escalate vs decide under the veto window
(565-595); batch independence (307-319); ACCEPT vs REWORK against GitHub rather than against the
report (529-548); and — the one this audit shows matters most — whether a given rule applies to
a project with no seats, lanes or lock wrapper.
Proposal. ADD evals/README.md (~300 tok). It is a new file, so it takes a new ceiling
rather than being paid by deletion inside SKILL.md; that needs a maintainer ruling quoted in
the PR.
Delta. +300 (new file). Gate. ① Yes — evals are how the package's own drift gets caught.
② No. ③ Yes, a rubric is a construct. Confidence HIGH. Human floor YES (new ceiling).

PMD-I-01 · SIZE PROPORTIONALITY · whole package

Evidence. 14,549 tokens for a package that teaches no ObjectStack surface at all: 0
TypeScript fences, 0 os:check markers, no references/_index.md, no Zod pointer. It costs
more than objectstack-data (13,892) and objectstack-platform (12,984), which teach the
product, and 1.75× the other process skill (8,333). Against the five customer decision
points enumerated above that is ~2,900 tokens per decision. Measured customer-inert or
duplicated text totals ≈ 4,445 tokens (30.6 %) before the structural move; with the template
and upstream sections relocated (PMD-B-01, PMD-C-01) ≈ 7,350 tokens (50 %) leaves the
always-loaded entry file. The package sits at headroom 0 on its ceiling, so it currently
cannot absorb any of the three additions in PMD-F-01/02/03 without the deletions that pay them.
Proposal. Apply the deletions above; then re-ratchet.
Delta. −4,255 net across all findings (14,549 → ~10,294).
Gate. ① Yes. ② Yes. ③ Yes. Confidence HIGH. Human floor YES.

---

Incidental falsehoods (by-product, not the job)

IF-1 · SKILL.md:456-457 — "Reports come back directly as each agent's final message — a
lossless channel … nothing can be lost between agent and reviewer." Contradicted inside
the same package 17 lines later at 474-475: "An agent that dies or returns malformed output
counts as status: \"blocked\" with its raw output attached" — i.e. the channel is known to lose
whole reports. Independently disproved by the internal oracle
(.claude/agents/os-dev.md:355-357), which records a container restart killing three devs
mid-final-message with the GitHub comment as the only surviving copy.

IF-2 · SKILL.md:1036-1039 — "serialized heavy phases — which the developer-agent
template enforces with a container-wide verification lock, a heap cap, scoped builds and
worktree cleanup." The template does not enforce it; it explicitly delegates: "The MECHANISM
is the host project's (its wrapper, its lock path, its budget — put it in the conventions
file)" (771-772). For any project without a wrapper — every customer on day one — nothing is
enforced. enforces should read requires.

---

Verdict

RESTRUCTURE. Pure-deletion reducible text measures 4,445 tokens = 30.6 %, at the
DIET/RESTRUCTURE boundary on its own; the structural half decides it. The package is one
monolithic 14,549-token file with zero progressive disclosure, zero evals, two blocks duplicated
inside itself (the report JSON, the sanitizer trap), a 2,376-token verbatim paste-block paid in
every session whether or not a dispatch happens, and an entire second job (upstream reporting)
that duplicates objectstack-platform and mis-routes customers into a 14.5k-token load.

The three highest-value additions, each with its payment

1. Dev-side premise verification + premise_still_valid (PMD-F-02, +60) ← paid by deleting
   "One semantics, N independent implementations" (PMD-E-02, −565). Net −505.
2. The git stash ban in the template (PMD-F-03, +40) ← paid by deleting the container-lock
   design properties (PMD-E-04, −250). Net −210.
3. Report as a GitHub comment in BOTH modes, plain-text marker, read it back (PMD-F-01, +50)
   ← paid by deleting the duplicated report JSON (PMD-B-02, −115). Net −65.

---

Generated by Claude Code

[os-litant]

os-dev-report

{
  "issue": 14300,
  "status": "done",
  "branch": "claude/issue-14300-skills-pm-dispatch-optimization",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/14444",
  "premise_still_valid": true,
  "summary": "skills/objectstack-pm-dispatch/SKILL.md shrunk from 14,549 to 11,567 tokens (net -2,982, lines 1,086 -> 879) in five commits keyed by finding id: PMD-C-01 (upstream reporting cut to a 245-token residue pointing at objectstack-platform's 'The App / Platform Boundary' by heading), E-05, E-02, E-03, D-01(+E-01), D-02, D-03, E-04, B-03, B-02, G-01, A-01 (description trigger dropped; skills/README.md and content/docs/ai/skills-reference.mdx regenerated by gen:skill-docs), the five drift syncs F-01..F-05 quoted from .claude/agents/os-dev.md and CLAUDE.md on origin/main 96b627d1, and the IF-2 'enforces' -> 'requires' correction. PMD-D-04 excluded by check:skill-frame-sync (both frame copies pinned; anchors now at :540 and :729), premise_false none, PMD-B-01 / H-01 deferred to #14296, the ~620-token gap against the card's ~-3,600 target itemised in the PR body, #14229's two pending deltas listed as a follow-up sync. Draft PR #14444 opened with first line 'Part of #14300', skip-changeset applied and read back twice (+4s and +2min: documentation, size/m, skip-changeset). REST channel was gated (403 on a repo-scoped read); card, comments and read-backs went through the zero-quota page payload, writes through MCP.",
  "tests": "head ea50dfb5 (== remote ref). Baseline on the untouched worktree, then after every commit: pnpm check:skill-frame-sync -> '4 copies of the decision frame are structurally isomorphic across 3 files; 4 axes: business-need -> long-term-soundness -> ai-authoring-safety -> startup-scope-discipline' (exit 0); pnpm check:skill-frame-freshness -> 'the decision frame in this tree is current with origin/main' (exit 0); node scripts/check-skills-token-ratchet.mjs -> 'skills/objectstack-pm-dispatch/SKILL.md is 11567 tokens (ceiling 14549; headroom 2982)' (exit 0); pnpm check:skill-identifier-liveness -> 'OK - Leg 1: 500 citation(s) ... (5 ledgered exemption(s)); Leg 2: 9 registered exhaustive section(s)' (exit 0, no Leg 2 heading bound to this file); pnpm check:skill-compatibility -> '11 SKILL.md file(s) reconciled against 78 workspace packages' (exit 0); pnpm --filter @objectstack/spec check:skill-docs -> 'Skill docs in sync' (exit 0); pnpm check:nul-bytes -> 'OK (scanned 7875 text file(s) ... no raw ASCII control bytes)' (exit 0). Re-derived union: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands from the tree at ea50dfb5 = 36 commands, every exit captured by redirect before any pipe: 33 green first pass; check:doc-formula-expressions and check:doc-security-posture green after building @objectstack/formula and @objectstack/lint (first pass exit 3 = PREREQUISITE NOT MET); check:skill-examples green after building types/client/client-react under the lock -> '259 prose examples type-check across 3 surface(s)' (first pass exit 1 = client-react dist unbuilt); node scripts/check-test-completeness.mjs NOT MEASURED locally by design (exit 3, needs a saved turbo test log; CI supplies it). node scripts/check-empty-changeset.mjs --base origin/main -> 'No empty-frontmatter changeset introduced by this diff'. Heavy steps under scripts/pm/os-verify-lock.sh (OS_VERIFY_LOCK_SLOT=issue-14300): spec build VERDICT command-exit 0 (waited 308s, held 206s); one earlier client-react closure build hit the 10-minute foreground cap (exit 143, lock held by issue-14304) and was re-run as a minimal types/client/client-react build (held 16s). Every edit asserted count-of-one on the exact source span and re-grepped after write (deleted anchors 0, inserted anchors 1). PR body read back from the rendered page: 44/44 sentinel fragments present, draft. Audit-record comment read back byte-identical to the local transliterated copy except one collapsed blank line before the footer. CI on #14444: in_progress at report time (not polled).",
  "mcp_calls": "5 - claim comment, audit-record comment, create_pull_request, issue_write (skip-changeset), this report comment; all reads went through git and the zero-quota issue/PR page payload",
  "open_questions": [],
  "out_of_scope_findings": []
}

---

Generated by Claude Code