Skills optimization flight — skills/objectstack-automation: RESTRUCTURE (≈ −1,125 tok net / −2,105 gross) — the package teaches a control-flow trio with zero authorings (`parallel_gateway`/`join_gateway`/`boundary_event`) and never mentions the trio the corpus uses (`parallel`/`try_catch`/`loop`); `notify` (#1 node type, 20 uses) has one table cell; the description still promises ETL pipelines removed in spec 17

[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; 25 findings. Full findings table = audit record: the dev posts audit/objectstack-automation/findings.md (seat scratchpad) verbatim as the first comment at claim time.

Audit summary

SKILL.md 1,001 lines / 12,768 tok (headroom 0) · evals/approvals/test-revise-loop.md 1,329 · evals/README.md 414 · generated 835 — package 15,346; all three ratcheted files at zero headroom. Approvals :296-650 = 5,246 tok = 41% of the file; the highest-density block ("Valid-but-silently-wrong" pitfalls 6–10, :836-951) starts at 84% depth. Real usage: 4 *.flow.ts + the showcase's 24 flows + jobs + webhooks; zero *.trigger.ts. Node-type frequency: notify 20, approval 15, screen 9, update_record/get_record/decision 5, connector_action 4, loop 3, try_catch 2, parallel 2, approval_revise 1 — parallel_gateway/join_gateway/boundary_event 0. 14 real patterns; 12 covered; 2 whole defineStack collections (jobs, webhooks) missing. Verdict RESTRUCTURE: wrong control-flow model taught; order inverts value; evals orphaned (grep evals SKILL.md → 0) and inverted (1,329 tok on the least-used node); the generated index misroutes to the ADR-0020-retired state-machine.zod.ts.

Top findings

#	id	span	proposal	Δtok
1	AUTO-H-01	evals/approvals/test-revise-loop.md:23-65 vs SKILL.md:382-442	DELETE the eval's Expected-Output fence + restating Mistakes rows; keep Scenario + Criteria	−700
2	AUTO-B-04	evals/README.md:13-43	planned-structure stub (DEFERRED, #14296 item 2)	−270
3	AUTO-C-02	SKILL.md:596, 600-606	DELETE hook rows + "There is no ctx.record" — anchor objectstack-data/references/data-hooks.md	−208
4	AUTO-B-07	SKILL.md:971-983	DELETE — third statement of pitfalls 6 and 10; keep the os validate block	−190
5	AUTO-B-03	SKILL.md:800-812	DELETE 4 of 5 State-Machine BPs — verbatim restatement of Notes 277-294	−150
6	AUTO-D-01	SKILL.md:152-164	REWRITE-AS-CONSTRUCT: "routes / does-not-route" two-cell row	−150
7	AUTO-D-06	SKILL.md:787-798	DELETE "Flow Design" BPs (generic; #5 dup of :228; #2 wrong)	−137
8	AUTO-D-03	SKILL.md:443-459	DELETE "Re-homing the old process model" (upgrade-skill territory)	−225
9	AUTO-D-05	SKILL.md:926-941	REWRITE-AS-CONSTRUCT: keep the functions fence, drop the rationale	−130
10	AUTO-D-02	SKILL.md:404-409	DELETE ADR-0044 amendment history (stated 3× more)	−115
11	AUTO-B-06	SKILL.md:866-870	MERGE-INTO 736-782	−90
12	AUTO-D-04	SKILL.md:895-902	DELETE spec-17 retirement narration; keep the prescription	−110
13	AUTO-D-07	SKILL.md:826-835	DELETE pitfalls 1/2/5 (engine-implementer advice, BPMN, dup of :691)	−110
14	AUTO-F-01	:76-77, 79, 791-792, 828-829	ADD parallel/try_catch/loop-container rows; mark the *_gateway/boundary_event trio as BPMN lowering, not author-facing (control-flow.zod.ts:42-46)	+120
15	AUTO-F-02	SKILL.md:661-676	REWRITE-AS-CONSTRUCT capability→surface table adding approvals + messaging (notify returns success with skipped: true when messaging is absent — notify-node.ts:313-320)	+40

Incidental falsehoods (fix in this flight)

1. SKILL.md:24-25, :35 — "covers … ETL pipelines" / "designing an ETL pipeline": the L2 ETL layer was removed in spec 17 with no replacement (17.etl-pipeline-layer-retired.ts:7-20); the package declares 17.x and has zero ETL body content. Also the frontmatter/description (AUTO-A-01).
2. SKILL.md:791-792 — "Always handle errors. Add boundary_event nodes": 0 authorings; the showcase's own waiver (coverage.ts:299-300) names try_catch and wait as the author-facing equivalents.

Three funded additions

1. Capability → surface table (AUTO-F-02, +40, paid by AUTO-D-06).
2. ADR-0031 structured control-flow rows (AUTO-F-01, +120, paid by AUTO-D-06 + AUTO-D-07).
3. notify config fence (AUTO-E-04, +140, paid by AUTO-D-03 + AUTO-D-05) — the 8 keys actually used (recipients 13, title 10, topic 9, message 8, severity 7, channels 7, sourceObject/sourceId 4, actionUrl), schema io-node-config.zod.ts:179-218, plus one "Valid-but-silently-wrong" line for the messaging-absent silent success.

Flight scope

IMPLEMENT (shrink-only): rows 1, 3–15 and the findings file at HIGH or MED; the three funded additions; falsehoods 1–2 incl. the description; AUTO-A-02/A-03 (add jobs/webhooks and screen-flow routing to the description — paid by same-file deletions); AUTO-B-01/B-02 reorder (free — bring the "Valid-but-silently-wrong" block up); link the eval from SKILL.md.

ANCHOR RULINGS: hooks → objectstack-data/references/data-hooks.md (AUTO-C-02: delete + pointer). The objectstack-query flight points at this package's :157-164 flow-filter rule — keep that section stable. Screen flows vs views: this package owns screen flows; objectstack-ui points here.

DEFER: AUTO-B-04 (eval stub, #14296 item 2); AUTO-B-05 (generator map build-skill-references.ts:120-129 → out-of-scope card for the spec lane: drop state-machine.zod.ts, add builtin-node-config.zod.ts / io-node-config.zod.ts).

Flight constraints (binding)

• ONE draft PR, first line Fixes #<this card>; governed ⇒ stays draft; review requests are the seat's step.
• Token ratchet: no ratcheted file may 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; follow-ups in the PR body. Generated files untouched. Regenerate the README index only if a drift gate requires it (report which).
• Live surface with zero measured usage ⇒ one row pointing at its schema; retired/tombstoned ⇒ delete. A false claim matching a spec .describe() string ⇒ spec-side twin card.
• Gates: node scripts/check-skills-token-ratchet.mjs, pnpm --filter @objectstack/spec check:skill-examples, 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 keyed by finding id; per-file token delta; needs:contract-review on both carriers (the control-flow rows and the notify silent-success semantics are contract claims).

Refs: #14292 · #14296 · #13658.

[os-litant]

Claim: PM loop round 3
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14308-skills-automation-optimization
Worktree: objectstack-issue-14308
Domain: domain:skills
File surface: skills/objectstack-automation/SKILL.md + skills/objectstack-automation/evals/approvals/test-revise-loop.md (AUTO-H-01), both shrink-only (ceilings 12,768 / 1,329, headroom 0 / 0); evals/README.md untouched (AUTO-B-04 deferred, #14296 item 2); generated references/_index.md untouched; 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 AUTO-A-01/A-02/A-03 description edits; ⛔ no other package (stop on breach; explain in the report)
Container & model: M (1,001-line entry file + one eval, 25 findings, control-flow model rewrite), mode:subagent, model: opus — --tier reads no path-derived mandate; the content limb applies, compensated by the seat's in-seat review at CONTRACT_REVIEW_TIER
Clause-②: yes — the ADR-0031 structured control-flow rows, the notify silent-success semantics (skipped: true when messaging is absent), the capability → surface table, and the two flow-filter rules ported from the query flight are contract claims ⇒ needs:contract-review on both carriers at PR creation
Serial constraints cleared: the query flight (PR #14443, in patch round) points at this package's flow-filter rule (SKILL.md:157-164 at base) — keep that section addressable, and it now carries two funded additions the query flight deleted on this card's ruling (precedence on config.filter with placeholder pass-through and variable shadowing; the static-checkability split: objectstack validate FAILS a {record.…} filter token naming an unknown field or an un-expanded relation hop, the same reference outside a filter stays a warning, an unresolved flow variable is not statically checkable) — paid inside SKILL.md; hooks → objectstack-data/references/data-hooks.md (path kept by the delivered data flight PR #14427); screen flows stay here (objectstack-ui points at this package); skills/objectstack-automation/** byte-identical between a59f78d and origin/main a98b61b3 (seat-measured 07:3xZ); no open PR touches this package; in flight alongside #14305 (ai) and #14307 (upgrade) on disjoint directories

Decision re-read: no maintainer ruling on this card; binding inputs = the mandate, the value gate, the ratchet discipline, and #14296 (item 2 deferred: eval stub). AUTO-B-05 (generator map) goes out as an out-of-scope finding card for the spec lane. Landing form: first line Part of #14308 (the deferred eval stub keeps the card open).

---

Generated by Claude Code

[os-litant]

Dev claim (os-dev subagent, dispatched by the skills lane seat).

Session: session_01LraLgQVGq8egUwfYZpbYt1 (the harness reports the same session id as the dispatching seat — subagent under the same session URL; the distinguishing identity bit is the branch below)
Branch: claude/issue-14308-skills-automation-optimization (empty branch pushed at claim time as the write probe)
Worktree: ../objectstack-issue-14308, base origin/main = a98b61b3ef695431db26097a28ae8e5f1dec8fdf

Read channel: repo-scoped REST is 403 for this seat (GitHub access is not enabled for this session.) ⇒ card body + full comment thread read via the public-repo page payload (zero quota); MCP reserved for writes and one targeted dedupe search.

Body verified complete — it ends with the Refs: #14292 · #14296 · #13658. line. No earlier dev claim on the card; nothing to collide with.

File surface (hard): skills/objectstack-automation/SKILL.md + skills/objectstack-automation/evals/approvals/test-revise-loop.md, both shrink-only; skills/README.md + content/docs/ai/skills-reference.mdx only as generator output. Landing form: Part of #14308 (AUTO-B-04 deferred to #14296 item 2 ⇒ the card stays open).

The audit record follows as the next comment.

---

Generated by Claude Code

[os-litant]

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

Transliteration key. The GitHub body sanitizer is measured to delete angle-bracket /
tag-shaped fragments (even inside backticks), so the two such spans in the source file are
respelled below. The file on the seat's disk is unchanged; this is a lossless rendering.

• OS-CHECK-MARKER (AUTO-E-01 row) — in the source this is an HTML comment whose entire
  body is os:check, i.e. the opt-in marker line check-skill-examples.ts looks for on the
  line directly above a fence.
• IDENT (AUTO-H-01 row) — in the source this is the angle-bracketed placeholder id that
  stood for each lint rule id inside the grep -rn ... packages/lint/src command.

Everything else is verbatim.

---

Audit — skills/objectstack-automation/ (read-only)

Pinned ref: objectstack origin/main = a59f78d. The four package files are byte-identical
between origin/main and the working checkout (git diff --stat origin/main HEAD -- skills/objectstack-automation -> empty), so all spans below are valid at the pinned ref.

Measured baseline

File	tokens (ratchet)	ceiling	headroom
SKILL.md	12,768	12,768	0
evals/approvals/test-revise-loop.md	1,329	1,329	0
evals/README.md	414	414	0
references/_index.md	835	-	generator-owned
authored total	14,511
package total	15,346

Source: node scripts/check-skills-token-ratchet.mjs. Zero headroom on all three authored files.

Internal weight of SKILL.md (bytes/4, measured by sed -n 'A,Bp' | wc -c):

Section	span	tokens	% of file
Frontmatter + intro + When to Use	1-46	481	4%
Flows	47-240	2,245	18%
State machines & Approvals	243-650	5,873	46%
- Approvals alone	296-650	5,246	41%
Triggers	653-782	1,412	11%
Best Practices	785-822	412	3%
Common Pitfalls (1-10)	824-951	1,779	14%
- "Valid-but-silently-wrong" 6-10	836-951	1,630	13%
Blueprint + Verify + References	953-1001	563	4%

Real-usage corpus (dimension E/F basis)

Grepped: examples/** (app-crm, app-showcase, app-todo), packages/apps/**, apps/**,
packages/triggers/*, packages/plugins/plugin-approvals, packages/plugins/plugin-webhooks.
4 *.flow.ts files + examples/app-showcase/src/automation/flows/index.ts (1,759 lines,
24 flows) + jobs/ + webhooks/. Zero *.trigger.ts files exist anywhere.

Node-type frequency across every real flow (grep -o "type: '[a-z_]*'" | sort | uniq -c):

node type	usages	SKILL.md coverage
notify	20	1 table row (SKILL.md:97), no config example
approval	15	355 tokens of tables + 3 examples
screen	9	1 table row (SKILL.md:100), no example
update_record / get_record / decision	5 each	example + notes
connector_action	4	1 table row
assignment / loop	4 / 3	1 table row each
http / create_record	3 each	example (http)
try_catch	2	absent
parallel	2	absent
subflow / script	2 each	table row + pitfall 9
map / wait / delete_record / approval_revise	1 each	table rows; approval_revise gets 875 tok
parallel_gateway / join_gateway / boundary_event	0	3 table rows + BP #2 + pitfall 2

Approval-config knob frequency: behavior 17, approvers 15, group 21, lockRecord 9,
minApprovals 2, maxRevisions 2, onEmptyApprovers 2, decisionOutputs 1,
approvalStatusField 0, escalation/escalateTo/timeoutHours 0, resolveAs 0.

---

Findings

id	dim	file:span	evidence (measured / quoted)	proposal	d tok	value gate (1)(2)(3)	conf	human-floor
AUTO-A-01	A, G	SKILL.md:24-25, SKILL.md:35	Intro: "This skill covers Flows ..., Triggers ..., and ETL pipelines." When-to-Use bullet: "You are designing an ETL pipeline for data synchronisation." The L2 ETL layer was retired at spec 17 - packages/spec/src/migrations/entries/semantic/17.etl-pipeline-layer-retired.ts:7-11 ("the whole L2 layer of automation/etl.zod.ts, its four enums and the ETL factory - 9 defs, 27 exported names"), replacement (removed - no protocol surface replaces it, deliberately); packages/spec/src/shared/retry-policy.zod.ts:138 "ETLPipeline.retry did the same until #6414 retired the L2 ETL layer". The package declares compatibility: Requires @objectstack/spec 17.x (SKILL.md:12) and carries zero ETL body content (grep -n ETL SKILL.md -> only 24 and 35).	DELETE both mentions	-40	(1) yes - stops routing "design an ETL pipeline" into a package with no ETL content (2) yes, deletion IS the fix (3) n/a	HIGH	yes (retired semantics)
AUTO-A-02	A, F	frontmatter SKILL.md:4-5 vs body	Frontmatter promises "...scheduled jobs, and webhooks". The body teaches scheduled flows and inbound webhook flow triggers only. defineJob (packages/spec/src/system/job.zod.ts:170, registered via defineStack({ jobs }) - packages/spec/src/stack.zod.ts:458) and defineWebhook (packages/spec/src/automation/webhook.zod.ts, defineStack({ webhooks })) have zero coverage in any published skill: grep -rn "defineJob|JobSchema|jobs:" skills/*/SKILL.md skills/*/rules/*.md skills/*/references/*.md -> 0 hits. Both are real: examples/app-showcase/src/automation/jobs/index.ts:16 (cron + retryPolicy + handler), examples/app-showcase/src/automation/webhooks/index.ts:17 (triggers: ['create','update','delete']). webhook.zod.ts is already listed in this package's own references/_index.md:18.	ADD a 6-line "Three automation collections: flows / jobs / webhooks" construct - paid by AUTO-D-03 + AUTO-D-04	+180 (paid)	(1) yes - an agent asked "fan task changes out to an external endpoint" currently learns only the http node (2) no - this capability text exists nowhere (3) yes, a table	HIGH	yes (adds published capability text)
AUTO-A-03	A, F	SKILL.md:57, SKILL.md:100	skills/objectstack-ui/SKILL.md:11-13 routes screen flows out to here ("interactive screen flows / wizards (those are *.flow.ts with type: 'screen' - see objectstack-automation)"), and skills/README.md:101-103 repeats it. This package's total screen-flow content is two table rows. Real corpus: 9 type: 'screen' occurrences, and the corpus has had to encode the contract in comments - examples/app-todo/src/flows/task.flow.ts:288-289 "Select options are { value, label } pairs (the screen contract's shape - bare strings fail the executor's config parse, #4277)" and :326-329 "The screen contract has no message/buttons keys ... description + waitForInput: true is the declared way to pause on a message-only confirmation screen". Schema at packages/spec/src/automation/builtin-node-config.zod.ts:423-455, not in references/_index.md.	ADD a screen-node config row-set + the {value,label} / waitForInput pitfall - paid by AUTO-D-01 + AUTO-D-06	+200 (paid)	(1) yes - decides two documented-nowhere authoring choices (2) no (3) yes, a table + one OK/NO pair	HIGH	yes
AUTO-B-01	B	SKILL.md:296-650 vs 836-951	41% of the file is Approvals (5,246 tok) while the highest-decision-density block - "Valid-but-silently-wrong (passes build, fails at runtime)" pitfalls 6-10, 1,630 tok - begins at line 836, i.e. 84% depth, behind ~800 lines. Those five pitfalls each change a specific keystroke (single-brace interpolation, outputVariable holds the record, date-equality never matches, script = registered function, bare CEL).	MOVE-TO - promote the "Valid-but-silently-wrong" block to immediately after the Flows section	0	(1) yes - reorders what an agent reads first (2) no (3) n/a	MED	yes (restructures)
AUTO-B-02	B, D	SKILL.md:131-187	Five stacked blockquotes (893 tok) sit between the "Flow Example" heading and the example itself. Only the first (133-139, 131 tok - edges not next, filter is a where map, fields not values, label required) is about the code that follows. The rest are three separate sub-topics (fault edges 399 tok, readonly/runAs 200 tok, elevate-the-write 151 tok).	MOVE-TO - keep 133-139 adjacent to the fence; relocate 141-186 to a "runAs & failure routing" subsection after it	0	(1) yes - the example is the copy target and sits 57 lines below its heading (2) no (3) n/a	MED	yes (restructures)
AUTO-B-03	B, D	SKILL.md:800-812 vs SKILL.md:277-294	"State Machine Design (ADR-0020)" best practices restate the Notes block 520 lines earlier, item for item: BP1 "Author it as a state_machine validation rule ... not a workflow metadata type (retired)" <- 249-257; BP2 "A state mapped to [] is a final/dead-end state" <- 257-258 + code comment 272; BP4 "Put guards in a sibling script / conditional rule" <- 284-286; BP5 "Put side-effects ... in a record-triggered Flow (ADR-0019)" <- 286-288. Only BP3 (implicit any->any is lenient) is new.	DELETE BP1/2/4/5; MERGE-INTO the Notes block the one surviving sentence	-150	(1) no for the 4 duplicates (2) yes (3) n/a	HIGH	no
AUTO-B-04	B, H	evals/README.md:13-43; SKILL.md (no evals link)	grep -n evals SKILL.md -> 0 hits: the eval tree is an orphan, unreachable from the entry file. evals/README.md:17-27 ships a "Planned structure" ASCII tree naming four files that do not exist (test-quorum-behaviors.md, test-schedule-binding.md, test-decision-edges.md, test-time-relative.md) - verified by find skills/objectstack-automation -type f (4 files total). :29-43 "Format" and "Contributing" are maintainer instructions shipped to every customer install.	DELETE evals/README.md:13-43 (keep the header + current-evals table)	-270	(1) no - a plan for unwritten files changes no authoring decision (2) yes (3) n/a	HIGH	no
AUTO-B-05	B	references/_index.md:16 + absent rows	The generated Core list points at automation/state-machine.zod.ts, which ADR-0020 retired as a record-lifecycle declaration; its own header says so (packages/spec/src/automation/state-machine.zod.ts:15-21: "the top-level workflow metadata type and object.stateMachines are both gone") and its single surviving door is ai/agent.zod.ts:197 lifecycle, marked [EXPERIMENTAL - not enforced] ... Parsed but no runtime consumer yet. Meanwhile automation/builtin-node-config.zod.ts and automation/io-node-config.zod.ts - the per-node-type config shapes (NotifyConfigSchema, screen fields, http timeoutMs) - are in neither list. Root cause: the generator map at packages/spec/scripts/build-skill-references.ts:120-129.	RETIRE-SURFACE (drop state-machine.zod.ts from the map) + ADD the two node-config files	+10	(1) yes - changes which file the agent Reads for a node's config keys (2) partially (3) n/a (generated)	HIGH	yes (generator map change)
AUTO-B-06	B, D	SKILL.md:866-887 vs SKILL.md:736-782	Pitfall 8 (332 tok) and the Time-relative-triggers section (535 tok) open with the same rule. 736-740: "Don't express 'act N days before/after a date' ... as a record_change flow gated on date-equality (end_date == daysFromNow(60)) - that predicate is only evaluated when the record happens to change". 866-870: "Time-relative rules ... are SCHEDULE flows, not record-change date-equality. record.end_date == daysFromNow(60) on a record-* trigger only fires if the record happens to be written on that exact day". Pitfall 8's unique half is the time-component equality failure + the abutting-window $gte/$lt construct (875-887).	MERGE-INTO SKILL.md:736-782 - fold 871-887 in, delete 866-870	-90	(1) no for the restated half (2) yes (3) the surviving half already IS a construct	HIGH	no
AUTO-B-07	B, D	SKILL.md:971-983 vs 842-859, 943-949	"Verify your work" restates two pitfalls. 976-978: "Node values take the single-brace flow-template dialect ... no validator checks its call names, so an unknown function is NOT build-checked - it throws FlowExpressionFunctionError at run time" <- 856-859: "any other name in call position fails the node ... The build does not catch these (conditions are checked, call-position names are not)". 971-974 <- 946-949 ("An UNKNOWN function (PRIOR(), a typo'd name) and a {...}-wrapped field ref both fail the build").	DELETE SKILL.md:971-983, keep the os validate / os build command block (985-991)	-190	(1) no - third statement of the same two facts (2) yes (3) n/a	HIGH	no
AUTO-C-01	C, D	SKILL.md:42 and 943-946 vs objectstack-formula/SKILL.md:160-205	The CEL stdlib is enumerated twice inside this package, and both times immediately defers elsewhere. SKILL.md:42-43: "(now(), today(), daysFromNow(n), daysBetween(a, b), isBlank(v), coalesce(v, fb)) ... See the objectstack-formula skill for the full CEL contract". SKILL.md:943-946 repeats a longer version, then "see objectstack-formula for the full table." Authoritative table: objectstack-formula/SKILL.md:166-201. The frontmatter already says "load objectstack-formula alongside" (SKILL.md:9-10).	DELETE the enumeration in SKILL.md:42-43; keep the pointer	-45	(1) no - an agent that needs the stdlib loads formula (2) yes (3) formula's table is the anchor	HIGH	no
AUTO-C-02	C	SKILL.md:596, 600-606 vs objectstack-data/references/data-hooks.md:364-376	Cheat-sheet rows 2-3 are object-hook rows ("Object hook handler (ctx) | ctx.input (write payload); ctx.result after the write | ctx.previous"), followed by an 8-line paragraph "There is no ctx.record. HookContext declares input / result / previous / session / ql...". objectstack-data/references/data-hooks.md:364-376 carries the same contract as its own table plus the nuance this file omits (:376 "Action bodies additionally receive ctx.recordId and ctx.record"). Hooks are routed to objectstack-data by this package's own frontmatter (SKILL.md:7-9).	DELETE SKILL.md:600-606 and cheat-sheet rows 2-3; keep rows 1 and 4 with a one-line pointer	-208	(1) no - this package's reader authors flows (2) yes (3) the surviving 2-row table stays a construct	HIGH	no
AUTO-C-03	C	SKILL.md:962 vs objectstack-ui/SKILL.md:1878-1897	CRM Blueprint row 4: "Action-to-flow bridge | src/actions/*.actions.ts | Trigger screen flows via Action.type = 'flow' for user-driven automation entry". objectstack-ui already carries the canonical, os:check-marked and therefore type-checked example of exactly this (defineAction({ ... type: 'flow', target: 'lead_conversion', locations: [...] })); actions are objectstack-ui's declared domain (objectstack-ui/SKILL.md:42).	DELETE row 4	-35	(1) no - a lower-information restatement of a verified example one package over (2) yes (3) ui's example survives	HIGH	no
AUTO-C-04	C, D	SKILL.md:731-734 vs skills/README.md:87, objectstack-formula/SKILL.md:383,399	"(Salesforce-flavor OLD / NEW were removed in M9.5 and now fault the predicate.)" is the fourth copy: skills/README.md:87 "Legacy OLD / NEW evaluate to null since M9.5"; objectstack-formula/SKILL.md:383 (translation row) and :399. The link to objectstack-formula is already on the next line (SKILL.md:734).	DELETE the parenthetical	-35	(1) no (2) yes (3) formula's translation table is the anchor	HIGH	no
AUTO-D-01	D	SKILL.md:152-164	212 tokens of justification prose inside the fault-edge blockquote: "It is not a way past a guardrail. ... A failure is a guard when re-running unchanged could never succeed and the fix is to edit metadata ... Never add a fault edge to silence such an error: a dropped filter condition widens the query, so routing it would let a delete_record empty the object while the run reported success."	REWRITE-AS-CONSTRUCT - one row: "Routes: 404 / rate-limit / rejected write / failed subflow. Does not route (fatal either way): missing required config key, dropped filter token, nesting-ceiling recursion, unscoped run."	-150	(1) the surviving row yes; the essay no (2) partially (3) yes, a two-cell row	HIGH	no
AUTO-D-02	D	SKILL.md:404-409 (+ evals/approvals/test-revise-loop.md:61)	ADR narration: "ADR-0044 D3 first prescribed an ordinary wait here and its 2026-07-28 amendment reversed that - a wait is resumeAuthority: 'any', so a raw POST /api/v1/automation/:name/runs/:runId/resume walked the back-edge with no submitter check and no audit row, and could destroy the run outright." The decision it changes is already stated at 401-404, again at 433-436, and again in the eval's mistakes table.	DELETE the amendment history (keep "the window is a service-owned pause (resumeAuthority: 'service')")	-115	(1) no - history of a reversal changes nothing for an author writing today (2) yes (3) n/a	HIGH	no
AUTO-D-03	D	SKILL.md:443-459	"Re-homing the old process model" - a 6-row table mapping pre-ADR-0019 ApprovalProcess.create({...}) concepts onto the flow graph, opening "If you've seen the pre-ADR-0019 ... shape". Cross-major migration is objectstack-upgrade's declared domain (objectstack-upgrade/SKILL.md:4-9). Every row's destination is already taught in "Branching, side-effects & rejection" (SKILL.md:622-636).	DELETE (or MOVE-TO objectstack-upgrade)	-225	(1) no for a new author (2) yes (3) 622-636 is the surviving construct	MED	yes (removes published capability text)
AUTO-D-04	D	SKILL.md:895-902	"The other dispatch forms were retired in spec 17 because none of them ran: config.actionType: 'email' | 'slack' were logger-backed stubs that delivered nothing ... and inline config.script JS was never executed (no server-side sandbox). ... Stored flows convert with os migrate meta --from 16." The same words already ship inside the spec the agent Reads (packages/spec/src/automation/flow.zod.ts:41-43), and os migrate meta is objectstack-upgrade's surface (objectstack-upgrade/SKILL.md:141-145). Maintainer principle 2026-08-27 argues against dual-spelling history in a 17.x-only package.	DELETE the retirement narration; keep the positive prescription	-110	(1) no - an author on 17.x never types actionType (2) yes (3) the surviving sentence is the prescription	MED	yes (retired semantics)
AUTO-D-05	D	SKILL.md:926-941	"The rule is load-bearing: a script step reports no record metrics in the run summary because every write a pure function causes is a downstream node that counts itself. A function that writes anyway makes its run report acted: 0 - indistinguishable from a sweep that silently did nothing." 215 tokens of why, after the rule and its OK/NO code pair are already given (904-924). The effect: 'writes' escape hatch IS decision-changing and real (examples/app-showcase/src/automation/jobs/sweep-project-health.ts:33-38).	REWRITE-AS-CONSTRUCT - keep the 4-line defineStack({ functions }) fence (936-941) with a one-line caption; delete 926-935	-130	(1) the fence yes; the rationale no (2) partially (3) yes, the code fence survives	HIGH	no
AUTO-D-06	D, G	SKILL.md:787-798	"Best Practices -> Flow Design", 5 items / 137 tok. #1 "Keep flows small and composable", #3 "Use variables for all dynamic values", #4 "Prefer get_record over multiple http calls" are generic engineering advice changing no ObjectStack keystroke. #5 "Always set timeoutMs on http nodes. Unset means no timeout at all" duplicates the in-example comment at SKILL.md:228. #2 "Add boundary_event nodes for timeout and error scenarios" is actively wrong guidance - see AUTO-F-01.	DELETE the whole subsection	-137	(1) no (2) yes (3) n/a	HIGH	no
AUTO-D-07	D	SKILL.md:826-835	Common Pitfalls 1-5 (144 tok). #1 "Circular flow references. ... Use a depth counter or visited set to detect cycles" is engine-implementer advice - a metadata author has no such lever, and the platform already enforces a nesting ceiling (referenced at SKILL.md:158). #2 is BPMN-interop (AUTO-F-01). #5 "Scheduled flows without idempotency" is restated at SKILL.md:691.	DELETE #1, #2, #5; MERGE-INTO the Flows section the two surviving clauses	-110	(1) no (2) yes (3) n/a	MED	no
AUTO-D-08	D	SKILL.md:22-25, 29-35	The intro paragraph and the 5-bullet "When to Use This Skill" restate the frontmatter description (SKILL.md:3-10), which is the field the loader routes on (skills/README.md:73). Both are paid on every session; the frontmatter is loaded anyway.	DELETE SKILL.md:22-25; keep the When-to-Use bullets minus the ETL one (AUTO-A-01)	-60	(1) no - routing already happened by the time the body is read (2) yes (3) n/a	MED	no
AUTO-D-09	D	SKILL.md:814-820	"Best Practices -> Trigger Design", 3 items / 88 tok. #3 "Use the start-node condition to narrow when the trigger fires" restates the code comment at SKILL.md:722 and the prose at 655-659. #2 (guard record-after-update self-writes) is the one decision-changing item.	MERGE-INTO SKILL.md:707-729 - keep #2, delete #1/#3	-60	(1) #2 yes, #1/#3 no (2) yes (3) n/a	MED	no
AUTO-E-01	E	SKILL.md - 1 of 13 TS fences marked	packages/spec/scripts/check-skill-examples.ts:18-23: "this gate is opt-in: a self-contained, should-compile block is marked by a OS-CHECK-MARKER HTML comment on the line directly above its fence." Measured: 13 TypeScript fences (10 typescript at 114/188/259/310/420/522/543/667/709/749, 3 indented ts at 875/909/936); exactly one carries the marker (SKILL.md:542). Bundle comparison (grep -rc 'os:check' skills/*/SKILL.md): ui 11, api 6, data 4, ai 4, i18n 4, platform 2, formula 2, upgrade 1, automation 1, query 0, pm-dispatch 0.	ADD markers to the two blocks that can be made self-contained (schedule flow 188-239, record-change trigger 709-729) with their import { defineFlow } line	+30 (paid by AUTO-D-06)	(1) yes - an unverified example is what the gate's own header calls "teaching code that no longer compiles" (2) no (3) yes, the examples ARE the constructs	HIGH	no
AUTO-E-02	E	SKILL.md:310-380, 709-729, 749-775, 667-676	Unmarked fences that would fail a type-check if marked: (a) 310, 709, 749 use the cel`...` tag with no import (grep -n "^import" SKILL.md -> one hit, line 544, inside the marked block); (b) 709-729 omits edges, which FlowSchema requires (packages/spec/src/automation/flow.zod.ts:659 - no .optional()), so FlowSchema.parse rejects it; (c) 667-676 calls defineStack({...}) with no import. None is false - each is a legitimate fragment - but (b) reads as a whole flow (it has name/label/type/nodes), which is the shape an agent copies.	REWRITE-AS-CONSTRUCT - make 709-729 a node-only fragment so it cannot be copied as a complete flow	-20	(1) yes - removes a copyable-but-unparseable flow (2) partially (3) yes	MED	no
AUTO-E-03	E	derived top-5 tasks vs coverage	Top-5 real authoring tasks by corpus frequency: (1) event->notify (20 notify nodes across 9 flows) - no example, no config keys; (2) approval chain (15 approval nodes, 8 flows) - 3 examples OK; (3) scheduled sweep / digest (5 type:'schedule'; flows/index.ts:362,1619; jobs/index.ts:16) - 1 flow example OK, no defineJob; (4) screen wizard (9 screen) - no example; (5) outbound integration (connector_action 4, http 3, defineWebhook 1) - http inside one example, connector_action no example, defineWebhook absent. 2 of 5 top tasks have no canonical example; a 3rd is half-covered.	ADD a notify-config fence + a screen-node fence (AUTO-A-03, AUTO-E-04)	see rows	(1) yes (2) no (3) yes	HIGH	yes
AUTO-E-04	E, F	SKILL.md:97 (whole coverage of notify)	notify is the most-used node type in the corpus (20 usages) and its entire documentation is one table cell: "notify | Send a notification through the messaging service (inbox channel by default)". Real config keys in use (grep -A8 "type: 'notify'" over examples/): recipients 13, title 10, topic 9, message 8, severity 7, channels 7, sourceObject/sourceId 4, actionUrl. Schema: packages/spec/src/automation/io-node-config.zod.ts:179-218. The silent-failure case is undocumented: packages/services/service-automation/src/builtin/notify-node.ts:313-320 - with no messaging service registered the executor logs "[notify] no messaging service registered; notification ... not delivered" and returns success with output: { delivered: 0, failed: 0, skipped: true }. Textbook material for this package's own "Valid-but-silently-wrong" section.	ADD an 8-line notify config fence + one line in the capability table (AUTO-F-02) - paid by AUTO-D-03 + AUTO-D-05	+140 (paid)	(1) yes - the #1 node type has zero authorable detail and a documented silent no-op (2) no (3) yes, a fence + a table row	HIGH	yes
AUTO-F-01	F, G	SKILL.md:76-77, 79, 791-792, 828-829	The package teaches only the BPMN-interop control-flow trio and never mentions the ADR-0031 structured constructs. Measured usage in examples/: parallel_gateway 0, join_gateway 0, boundary_event 0; parallel 2 (flows/index.ts:929, 1207), try_catch 2 (:1009, 1226), loop container 3, map 1. The repo's own coverage waivers say it outright - examples/app-showcase/src/coverage.ts:295-300: "parallel_gateway: BPMN-interop lowering target - the author-facing form is the ADR-0031 structured parallel container (FanOutNotifyFlow)"; "boundary_event: BPMN-interop form; the author-facing equivalents the showcase demos are the ADR-0031 try_catch container (ResilientSyncFlow) and wait timers". The spec agrees: packages/spec/src/automation/control-flow.zod.ts:5-11,42-46 - the constructs are "the native + AI-authored flow model ... the right substrate for LLM authoring", while the gateways "remain author-invisible interchange representations". grep -n "try_catch|ADR-0031|config.body" skills/objectstack-automation/** -> 0 hits. control-flow.zod.ts reaches references/_index.md:23 only as a transitive dependency.	ADD 3 rows to the Control-Flow node table (parallel / try_catch / loop with config.body, plus "the *_gateway / boundary_event trio is BPMN-interop lowering, not author-facing") + DELETE BP #2 and pitfall #2 (AUTO-D-06, AUTO-D-07)	+120, net -127 with its payers	(1) yes - changes which node type the agent emits for fan-out and for error handling (2) no - the payer is elsewhere (3) yes, table rows	HIGH	yes (changes published capability text)
AUTO-F-02	F	SKILL.md:661-676	"Prerequisite - enable the triggers capability" lists requires: ['automation','triggers'] + 'job' + 'queue'. It omits the two capabilities the corpus's two most-used node types need. examples/app-showcase/objectstack.config.ts:92-107 documents them inline: "approvals - ApprovalsServicePlugin, so the approval flow node (ADR-0019) is contributed to the engine"; "messaging - MessagingServicePlugin, so the notify node delivers to the inbox channel (sys_inbox_message rows) instead of degrading to a logged no-op"; the real list is ['ui','automation','approvals','messaging','triggers','job','marketplace','webhooks','realtime']. Confirmed in the executor (notify-node.ts:313-320). An agent following this section authors an approval flow (15 real usages) and a notify node (20) and gets a node type with no registered executor and a silent success that delivered nothing.	REWRITE-AS-CONSTRUCT - a capability->surface table (automation->engine; triggers->record-change/schedule/api binding; job->cron + timeRelative sweep; queue->inbound webhook (503 without); approvals->approval/approval_revise; messaging->notify delivery; webhooks->defineStack({ webhooks }))	+40	(1) yes - the highest-frequency silent failure this package can prevent (2) no (3) yes, a 7-row table replacing prose	HIGH	yes
AUTO-F-03	F	zero-real-usage documented surfaces	Grepped examples/**, packages/apps/**, apps/** for every approval knob the package documents: approvalStatusField 0 (yet promoted as Approval Best Practice #2, SKILL.md:641-643), escalation/escalateTo/timeoutHours 0 (SKILL.md:619), resolveAs 0 (SKILL.md:495,524,533). All three are live implementations, not speculative surface - packages/plugins/plugin-approvals/src/approval-service.ts:1420 (const resolveAs = String(a.resolveAs ?? 'user')) and packages/plugins/plugin-approvals/src/approvals-plugin.ts:293-335 (SLA escalation scan job, ADR-0042). Zero usage in examples is evidence, not proof; I grepped only the three example apps + packages/apps + apps.	no change - recorded so a future trim does not mistake them for dead surface	0	n/a	HIGH	no
AUTO-G-01	G	SKILL.md:37-43 vs 554, 573 + corpus	The package instructs "Use the P`...` and cel`...` tagged templates from @objectstack/spec" for "Every condition / guard / entryCondition / filter value in this skill", then uses cel` in 8 places (329, 366, 372, 374, 524, 527, 533, 723) and plain strings in its only type-checked block (SKILL.md:554 condition: "status == 'submitted'", :573) and in its eval (evals/approvals/test-revise-loop.md:31). The real corpus is unanimously plain strings: grep -rn "cel`" examples/**/flows -> 0 hits; all 20 flow/edge conditions are bare strings (examples/app-showcase/src/automation/flows/index.ts:39,153,212,281,282,325,438,691,808,924,1004,1099,1201,1258,1259,1501,1502,1532,1582,1696).	REWRITE-AS-CONSTRUCT - one sentence: "A condition slot takes a plain CEL string; the cel`...` tag is the same string with author-time validation. Both parse - pick one per file."	-10	(1) yes - an agent currently cannot tell which of two spellings the package endorses (2) no (3) yes	HIGH	no
AUTO-G-02	G	SKILL.md:59, 313, 712 vs corpus	Both record-change examples declare type: 'record_change'. The corpus overwhelmingly declares type: 'autolaunched' with the same start-node binding: 16 flows in examples/app-showcase/src/automation/flows/index.ts pair type: 'autolaunched' with a record-* triggerType, versus 1 using type: 'record_change' (:1683). Not a falsehood - the engine derives the binding from the start node regardless (packages/services/service-automation/src/engine.ts:2169-2189 resolveTriggerBinding) and a lint rule does key on record_change (packages/lint/src/validate-flow-trigger-readiness.ts:604-631) - but the package never says the two coexist, and its own eval uses autolaunched (evals/approvals/test-revise-loop.md:27) against its own examples.	ADD one clause to the record_change Flow-Types row ("autolaunched + a record-* start binding behaves identically; record_change additionally opts into the trigger-readiness lint")	+20	(1) yes - resolves a live contradiction between the skill's examples and its eval (2) no (3) yes, one table cell	MED	no
AUTO-H-01	H, I	evals/approvals/test-revise-loop.md (1,329 tok) vs SKILL.md:382-442 (875 tok)	2,204 tokens - 15% of the authored package - go to ADR-0044 send-back, a surface with 1 real usage (examples/app-showcase/src/automation/flows/index.ts:248, the single approval_revise node). The eval restates the SKILL section rather than testing beyond it: the canonical shape (eval 23-50 ~ SKILL.md:420-431), the four lint findings (eval 59-65 ~ SKILL.md:433-441), the wait-vs-approval_revise amendment rationale (eval 61 ~ SKILL.md:404-409), the same showcase pointer (eval 52-53 ~ SKILL.md:440-441). By contrast notify (20 usages) gets ~15 tokens. All lint ids in the eval are real (grep -rn IDENT packages/lint/src -> 1 non-test hit each for the four flow-approval-revise-* plus flow-error-label-not-fault, flow-date-equality-filter), so no rubric row would teach a grader to fail a correct answer.	DELETE the eval's Expected-Output fence and the Common-Mistakes rows that restate SKILL.md; keep Scenario + Validation Criteria	-700	(1) no - a grader needs criteria, not a second copy of the rule (2) yes (3) the Validation Criteria list survives as the construct	HIGH	yes (restructures a shipped file)
AUTO-H-02	H	evals/README.md:11 + package-wide	The eval suite covers exactly one of the package's ~40 rules - the one with the fewest real usages. Nothing tests the rules that gate the corpus's actual output: single-brace vs double-brace interpolation (SKILL.md:842-859), filter as a where map vs UI filter triples (SKILL.md:135-138), unguarded decision edges running in parallel (SKILL.md:343-346), runAs: 'system' for readonly writes (SKILL.md:166-176). Sibling comparison: no other package ships a prose eval fixture (ls skills/*/evals -> README-only except ui's 1,102-token JSON).	ADD - after AUTO-H-01's trim, one test-decision-edges eval (already listed as planned at evals/README.md:24)	+250 (paid by AUTO-H-01)	(1) yes - grades the rule with the widest blast radius (2) no (3) yes, an eval IS a construct	MED	yes
AUTO-I-01	I	package = 14,511 authored tokens	3rd-largest SKILL.md in the bundle (ui 25,441 > pm-dispatch 14,549 > data 13,892 > platform 12,984 > automation 12,768). Distinct real-usage patterns found in F: 14 (record-change->notify, approval chain, per-group/quorum sign-off, dynamic/expression approvers, send-back loop, scheduled digest, timeRelative sweep, screen wizard, subflow composition, loop/map batch, parallel fan-out, try_catch resilience, connector/http integration, inbound webhook) plus 2 uncovered collections (jobs, webhooks). 900 authored tokens per pattern, but distributed 41% to one pattern family. Sum of DELETE/MERGE rows above: -2,105; sum of paid ADDs: +980; net ** -1,125 (-7.8%)** while raising covered patterns from 12 to 16.	-	-1,125 net	-	HIGH	-

Incidental falsehoods (by-product, not the job)

where	claim	implementation truth
SKILL.md:24-25, SKILL.md:35	"This skill covers ... and ETL pipelines" / "You are designing an ETL pipeline for data synchronisation."	The L2 ETL layer (9 defs, 27 exported names) was removed in spec 17 with no replacement surface - packages/spec/src/migrations/entries/semantic/17.etl-pipeline-layer-retired.ts:7-20; corroborated at packages/spec/src/shared/retry-policy.zod.ts:86,138. The package declares compatibility: Requires @objectstack/spec 17.x. Filed as AUTO-A-01.
SKILL.md:791-792	Best Practice: "Always handle errors. Add boundary_event nodes for timeout and error scenarios."	boundary_event has 0 authorings in the whole corpus and the repo's own waiver calls it a non-authored BPMN-interop form whose "author-facing equivalents ... are the ADR-0031 try_catch container ... and wait timers" (examples/app-showcase/src/coverage.ts:299-300); packages/spec/src/automation/control-flow.zod.ts:42-46 calls the trio "author-invisible interchange representations". Filed as AUTO-F-01.

Not re-verified: the 2026-08-31 factual sweep (program #13658) is treated as done. The two rows
above surfaced while measuring dimensions A/G and F respectively.

---

Generated by Claude Code