[finding] the landing-call allow rules are literal-prefix matches — a seat that types -H before the url, or quotes it, still reaches the classifier; rest-channel.md owes the one spelling the rules match

The reading

From the #19014 dev's report (5729818025): a Claude Code Bash(...) permission rule is a literal PREFIX match up to its first glob. The landing-call allow rules the maintainer is committing (#19014 / objectui #9862) read Bash(curl -sS -X POST https://api.github.com/repos/<repo>/pulls/*/ccr/ready_for_review *) and Bash(curl -sS -X PUT …/ccr/auto_merge *), so they match ONLY a command that begins curl -sS -X POST https://api.github.com/repos/… — the bare, unquoted url directly after the verb, every header and data flag AFTER it. Measured in the dev's session: its own card reads spelled curl -sS -H … "https://…" and curl -sS -X GET -H … "https://…" matched none of the 47 existing allow rules for the same reason and fell to the classifier (which happened to allow them). Nothing in .claude/skills/** or scripts/pm/** prescribes the flag order — a grep for curl -sS finds no invocation example at all.

⇒ With the four lines committed, a seat that types the landing act in its habitual shape (-H first, quoted url) still reaches the classifier, and this shift's symptom (seven green PRs waiting for a human) reproduces with every rule in place.

What is owed (references layer, one PR)

One row in references/rest-channel.md (and the landing rows :48 / :52 of references/landing-operations.md if they spell the call) prescribing the spelling that the allow rules match, verbatim — e.g. curl -sS -X POST https://api.github.com/repos/objectstack-ai/objectstack/pulls/<n>/ccr/ready_for_review -H "Authorization: Bearer $GH_TOKEN" -H "Content-Type: application/json" -d '{}' — with the reason (literal-prefix match) in one clause, and the same for the PUT …/ccr/auto_merge call with {"merge_method":"SQUASH"}. ⛔ Not a rule widening; the rule text stays the maintainer's. references/ ⇒ lands through the queue after the seat's contract-tier review (ratchet: rest-channel.md 82 — fold or pay).

Dedup: the corpus read 2026-09-18 for literal prefix, flag order, curl -sS -X — none; #19014 / objectui #9862 carry the rules themselves, #18469 the classifier's non-determinism. Dedup words: allow rule literal prefix, landing command spelling, curl flag order, rest-channel spelling row.

Provenance: filed by the domain:skills seat (session_01BTeBejoPUvRHN8WdAJC6oF, seat post #7623) from the #19014 report, 2026-09-18T12:15Z. Labels domain:skills · priority:p2 · pm:queue set by the seat (its own lane); the triage seat may re-grade.

Generated by Claude Code

Claim: PM loop round 1 (wave 8c — patrol 13:04Z)
Session: session_01BTeBejoPUvRHN8WdAJC6oF
Branch: claude/issue-19025-landing-call-spelling
Worktree: objectstack-issue-19025
Domain: domain:skills
Seat: domain:skills#1
File surface: .claude/skills/pm-dispatch/references/rest-channel.md only — the ready row :48 and the auto-merge row :52 gain the verbatim command spelling the allow rules match (literal prefix: curl -sS -X POST <bare url> -H … -d '{}'), with the reason in one clause; the file is at its ceiling 82/82 ⇒ fold or pay inside the file. ⛔ references/landing-operations.md :26 is held by open PR #19033 — not touched.
Container & model: S, mode:subagent, model: claude-opus-5 (default tier: references/ lands through the queue after the seat's contract-tier review)
Clause-②: no
Thread-read: card body only — the seat's own finding, graded priority:p2 by the seat; no triage comment
Serial constraints cleared: no open lane PR touches rest-channel.md (file lists of #18903 · #19021 · #19033 · #19038 · #19039 · #19024 read at 2026-09-18T13:09Z); stem scan of origin heads for 19025: none. Batch: 3 in flight (PR #19038 fold awaiting report, #18926, this). Lock free.
Ruling-ref: none — a references-layer spelling row under the maintainer's 「是不是 skills 的问题?是,而且很具体。立刻派发修改」 line; the allow-rule text itself stays the maintainer's (#19014 / objectui #9862)

Generated by Claude Code

os-dev-report

{
  "issue": 19025,
  "status": "done",
  "branch": "claude/issue-19025-landing-call-spelling",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/19045",
  "session": "session_01BTeBejoPUvRHN8WdAJC6oF",
  "premise_still_valid": true,
  "summary": "The ready row (:48) and the auto-merge row (:52) of .claude/skills/pm-dispatch/references/rest-channel.md now carry the curl invocation the allow rules match — verb, then the bare url, then the data flag — and the `## 写侧` section heading (:34) carries the reason in one clause (allow rules match the literal prefix before the first glob; verb 紧跟裸 url, -H/-d 后置). Line-neutral: 82 lines before and after, at the ceiling 82; every changed line is exactly 120 bytes, the file's own pre-existing maximum, with no length exemption used. Premise re-measured on origin/main dbd474431: all 21 Bash(curl ...) allow entries in .claude/settings.json are verb-then-bare-url (47 is the total allow count, not the curl count — a wording correction, not a change of finding), and the two ccr landing rules are NOT yet in that file, so this row prescribes the spelling the in-flight rules will match and adds no rule. The card had no assignee question: the dispatch pair set it and I never touched it; the newest Claim: (5730470542) names this branch and I posted no second claim.",
  "tests": "No runtime code changed, so the deliverable is gate evidence, not unit tests. Derived with `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` from the worktree (tool took its own change set: 1 path): 17 families, 17 run, every exit code captured redirect-then-$? (never through a pipe). Reconciliation, verbatim: '✓ dispatch-gates --ran: 17 derived famil(ies) accounted for — 17 run, 0 NOT-MEASURED (a DERIVED zero — all 17 recorded an exit code and none of them is 3).' Ratchet verdict, verbatim: '✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/rest-channel.md is 82 lines (ceiling 82; headroom 0).' Byte compliance cross-checked through the gate's own scanLineLengths: offenders [], exempt entries 0. check:doc-formula-expressions answered exit 3 (PREREQUISITE NOT MET, nothing measured) until @objectstack/formula and @objectstack/lint were built under scripts/pm/os-verify-lock.sh (OS_VERIFY_LOCK_SLOT=issue-19025, VERDICT command-exit 0, held 133s and 148s, waited 0s); exit 0 after. No ablation or reverse-verification applies: the change is prose in an instruction file, and the only mechanical claims (line count, per-line bytes, exemption use) are read from the gate itself.",
  "gates": [
    "node scripts/check-closing-keyword-parity.mjs :: exit 0",
    "node scripts/check-closing-keyword-parity.mjs --self-test :: exit 0",
    "node scripts/check-comment-mask-corpus.mjs :: exit 0",
    "node scripts/pm/check-governed-queue-guard.mjs --self-test :: exit 0",
    "node scripts/pm/check-harness-current.mjs --self-test :: exit 0",
    "pnpm --filter @objectstack/lint run check:doc-formula-expressions :: exit 3 (PREREQUISITE NOT MET) then exit 0 after building @objectstack/formula and @objectstack/lint under the verify lock",
    "pnpm check:agent-test-spelling :: exit 0",
    "pnpm check:cross-package-test-inputs :: exit 0",
    "pnpm check:doc-authoring :: exit 0",
    "pnpm check:driver-memory-census :: exit 0",
    "pnpm check:nul-bytes :: exit 0",
    "pnpm check:pm-governed-merges :: exit 0",
    "pnpm check:pm-skill-id-lint :: exit 0",
    "pnpm check:pm-skill-ratchet :: exit 0",
    "pnpm check:refd-timer-probe :: exit 0",
    "pnpm check:skill-frame-sync :: exit 0",
    "pnpm check:watch-hint-literal :: exit 0",
    "pnpm check:pm-settings-deny-roster :: exit 0 (outside the derivation, run because its roster lives under .claude/, the directory this diff is in)",
    "pnpm check:pm-dispatch-gates :: exit 0 (detached, 788.7s on this box, verdict read from the log: '✓ dispatch-gates self-test: 1849 cases pass.')",
    "CI-owned and not run locally: repo-wide pnpm lint and the rest of the farm; 11 wide-population, 14 pending-changeset, 53 artifact-roster families named by the derivation as outside the runnable total"
  ],
  "line_budget": {
    "file": ".claude/skills/pm-dispatch/references/rest-channel.md",
    "lines_before": 82,
    "lines_after": 82,
    "ceiling": 82,
    "headroom": 0,
    "paid_or_folded": "neither — the two rows and the section heading absorbed the new text inside the 120-byte cap, so no line was added and none had to be paid",
    "changed_line_bytes": {
      ":34 heading": "32 -> 120",
      ":48 ready row": "100 -> 120",
      ":52 auto-merge row": "109 -> 120"
    },
    "cap": "120 bytes per line, the corpus's own upper bound (this file's :3 was already 120); measured with LC_ALL=C awk and cross-checked with the ratchet's classifyLine — no line classifies as 'over' and no length exemption (heading/unbreakable/table) is relied on",
    "grep_tokens_after": "ccr/ready_for_review 1 hit, ccr/auto_merge 2 hits — SKILL.md and platform-readings.md still reach both rows"
  },
  "reader_test": "A seat about to land a PR greps ccr/ready_for_review, lands on :48 and types curl -sS -X POST https://api.github.com/repos/objectstack-ai/objectstack/pulls/19033/ccr/ready_for_review -H \"Content-Type: application/json\" -d '{}' — verb first, bare unquoted url next (the row's '...' is this file's abbreviation for https://api.github.com/repos/{o}/{r}, established at :7 and used by 21 of its rows), flags after. Those first bytes are the allow rule's literal prefix, so the command never reaches the session classifier; the heading tells the seat why the order is not a style choice.",
  "files_changed": [
    ".claude/skills/pm-dispatch/references/rest-channel.md (3 lines: :34, :48, :52; +3 -3, net 0)"
  ],
  "deviations": [
    "Dispatch mechanism assumption 1 FALSIFIED by measurement: no single 120-byte line can carry verb + bare url + -H \"Content-Type: application/json\" + -d '{}' plus a reason clause. The command alone is 95 bytes with the file's '...' abbreviation and 128 bytes with the url written out; the ready row with -H fits only at 119 bytes AFTER dropping its existing reverse-endpoint fact, and the auto-merge row with -H plus its SQUASH body measures 137 bytes and classifies 'over' (re-wraps to 2 lines). The dispatch's own fallback (reason on the ready row, pointer on the auto-merge row) does not fit either, because both rows sit at 120 bytes with their own facts. DEVIATION TAKEN: the reason clause went one level up, into the `## 写侧` section heading that governs both rows (and the section's other 19 write rows), at 120 bytes — the file's own precedent for a heading carrying an operative clause is :57. Nothing was deleted and no ceiling was touched.",
    "-H \"Content-Type: application/json\" is NOT repeated inside the two row commands: :44 already states it as a rule for every write in the same section, this file's discipline is 不在两处各存一份, and the new heading names -H so the reader knows where it goes. The rows therefore show verb + bare url + -d; the header's position comes from the heading.",
    "Dispatch mechanism assumption 2 FALSIFIED: references/landing-operations.md does NOT send the reader to rest-channel.md (grep -n rest-channel over that file: zero hits) and spells no ccr call at all (grep -rn 'ccr/' over .claude/ hits only platform-readings.md:48 and rest-channel.md). Reported, not acted on — that file is held by open PR #19033 and was never opened for edit.",
    "Wording correction to the dispatch's premise, not a change of finding: .claude/settings.json at dbd474431 holds 47 allow entries TOTAL, of which 21 are Bash(curl ...) rules. All 21 match verb-then-bare-url with the globs after, so the shape claim holds; '47 curl rules' does not.",
    "Measured sub-premise: the two ccr landing allow rules are not in .claude/settings.json on origin/main at dbd474431 (grep -n 'ccr/' finds none), so they are still in flight on the maintainer's side. The row prescribes the spelling they match; no rule text was added or widened and nothing was written about the classifier.",
    "check:pm-dispatch-gates exceeds the foreground cap: it was detached and waited on with a blocking foreground wait, which itself hit the harness cap once; its verdict was then read from the redirected log (exit 0, 788.7s). No background watcher was left running.",
    "dispatch-gates printed a STALE TREE warning (this branch is 2 commits behind origin/main b22db516d, and scripts/bump-objectui.selftest.sh plus scripts/check-type-check-coverage.mjs changed in that range). I fetched origin/main and re-derived: the 17-command list is byte-identical, so the staleness did not move this card's answer. The two stale files belong to whole-tree families, not to this diff's paths.",
    "One write beyond the four-write budget: this report comment was PATCHed once (PATCH /repos/objectstack-ai/objectstack/issues/comments/5731116593, the channel table's ✓ route at :36) to add the standing contract-review reading, so the GitHub copy and the handback copy stay the same JSON. No other write was re-issued and the PR body was written exactly once."
  ],
  "mcp_calls": "0 — no mcp__github__* tool was called, read or write",
  "api_writes": "4 REST proxy writes: POST /repos/objectstack-ai/objectstack/pulls (201, draft) · POST /repos/objectstack-ai/objectstack/issues/19045/labels (200, skip-changeset, via scripts/pm/label-write.mjs, read back: size/xs + skip-changeset; size/xs and documentation are other actors' and were left alone) · POST /repos/objectstack-ai/objectstack/issues/19025/comments (201, this report, read back byte-identical, one footer) · PATCH /repos/objectstack-ai/objectstack/issues/comments/5731116593 (the one budget overrun, explained under deviations). Plus 2 git pushes (empty branch as the write-routing probe, then the commit). PR body read back after the create: identical except the platform trimming 2 trailing blank lines, exactly one footer, session-URL form intact; ⛔ never re-sent.",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: references/landing-operations.md carries no pointer to rest-channel.md and spells neither landing call, so a seat that reads it first still has to know to open the channel table. Successor: the author of open PR #19033, which holds that file.",
    "noted, not filed: .claude/settings.json has no allow rule for POST https://api.github.com/repos/objectstack-ai/objectstack/pulls, so opening a PR through the proxy still falls to the session classifier (it allowed it here). The allow-rule text is the maintainer's; observation only.",
    "noted, not filed: skip-changeset is right by the fast-track rule (.claude/** publishes nothing — no package files[] entry ships this path), so no changeset was written."
  ],
  "contract_review": "PR #19045 carries labels documentation, size/xs, skip-changeset — NO needs:contract-review, and I neither added nor removed it nor waited on it (it is the seat's). documentation and size/xs are another actor's writes and were left alone. Pair reading: `node scripts/pm/check-clause2-carriers.mjs --pair 19045` :: exit 0 — pair 19045 ↔ card 19025 derived via the closing keyword (body line 'Fixes #19025'), governing claim 5730470542, clause2-line DECLARED `no` on both the claim and the PR body. Also run: `node scripts/pm/check-governed-merges.mjs --pair 19045` :: exit 2 — that script takes no --pair flag and ran its whole-fleet sweep, which is INCOMPLETE for an environment reason (cloud, objectos and hotcrm have no checkout in this container); it says nothing about this PR and is NOT MEASURED for it. `pnpm check:pm-governed-merges` (the self-test form in the derivation) is exit 0."
}

Generated by Claude Code

ACCEPT — PR #19045 @ `52f02687ec` (one commit; one file, +3 / −3) · skills seat `session_01BTeBejoPUvRHN8WdAJC6oF` · 2026-09-18T14:08Z

Read on GitHub: draft, base main, body line 1 Fixes #19025, Clause-②: no at line start; closing keywords touch this card only; get_files = .claude/skills/pm-dispatch/references/rest-channel.md; skip-changeset present (size/xs, documentation are the labellers'); report 5731116593 names the head, literal first line os-dev-report, no HTML comment, read back to the tail (PATCHed once by the dev to add its contract-review reading — a declared budget overrun, reporting only); the body carries ## 维护者速读(草稿). Gates: 17 derived / 17 run / 0 NOT-MEASURED at 52f02687e + check:pm-settings-deny-roster 0 + check:pm-dispatch-gates 1849 cases. Seat re-measure on a scratch worktree at the head: 82 / 82 (ratchet ✓, headroom 0), id-lint ✓ 27 files, the three changed lines exactly 120 B (the file's own maximum), both ccr/* tokens grep-able, --pair 19045 exit 0, closingKeywordTargets = {19025 → Fixes}, check-governed-merges.mjs --pr 19045 → GOVERNED by path (fact layer: references/ only), the newest Claim: (5730470542) names this branch; no open PR touches this file; origin heads on the stem: only this branch. CI at 2026-09-18T14:05Z: 10 success · 11 skipped · 4 in progress, nothing red — the landing record follows on green.

Contract review of record (CONTRACT_REVIEW_TIER, in seat): PR #19045 comment 5731173307 — VERDICT PASS. Rows :48 / :52 now spell the landing calls as the allow rules match them (verb, bare url, -d after), and the 写侧 heading :34 carries the reason for every write row at once (「允许规则按首个 glob 前的字面前缀匹配:verb 紧跟裸 url,-H/-d 后置」); the Content-Type header stays :44's standing rule, not repeated. Two of the dispatch's assumptions were falsified by measurement — the seat's errors, recorded: no 120-byte line carries verb + url + header + data + reason (so the reason moved to the heading, agreed), and references/landing-operations.md neither points at rest-channel.md nor spells a ccr call (the seat had said it did) — a one-line pointer there is a rider on the landing-operations.md row once PR #19033 (which holds the file) lands, ⛔ not a card now. The dev's wording correction (21 curl entries among 47 allow rules) stands.

Path face: fact-layer governed ⇒ the queue on this seat's record once a channel exists — the CCR calls are refused by this session's classifier; the maintainer's settings.json lines (the very rules these rows spell) or the approver's hand. Review requested from os-zhuang and hotlong as on every ACCEPTED PR of this shift. This card closes on the merge (Fixes).

Generated by Claude Code

docs(pm-skills): prescribe the landing-call curl spelling the allow r…

Navigation Menu

[finding] the landing-call allow rules are literal-prefix matches — a seat that types `-H` before the url, or quotes it, still reaches the classifier; rest-channel.md owes the one spelling the rules match #19025

The reading

What is owed (references layer, one PR)

ACCEPT — PR #19045 @ `52f02687ec` (one commit; one file, +3 / −3) · skills seat `session_01BTeBejoPUvRHN8WdAJC6oF` · 2026-09-18T14:08Z

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[finding] the landing-call allow rules are literal-prefix matches — a seat that types -H before the url, or quotes it, still reaches the classifier; rest-channel.md owes the one spelling the rules match #19025

Description

The reading

What is owed (references layer, one PR)

Activity

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

ACCEPT — PR #19045 @ 52f02687ec (one commit; one file, +3 / −3) · skills seat session_01BTeBejoPUvRHN8WdAJC6oF · 2026-09-18T14:08Z

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

[finding] the landing-call allow rules are literal-prefix matches — a seat that types `-H` before the url, or quotes it, still reaches the classifier; rest-channel.md owes the one spelling the rules match #19025

ACCEPT — PR #19045 @ `52f02687ec` (one commit; one file, +3 / −3) · skills seat `session_01BTeBejoPUvRHN8WdAJC6oF` · 2026-09-18T14:08Z