Repository navigation
pm tooling: the os-dev's PR-assignee write (label-write.mjs --assign) is refused by the auto-mode permission classifier on 4 of 12 objectui PRs in one seat's session — same command, same brief, nondeterministic #22467
Description
Activity
Triage: graded on first touch,
tooling·priority:p3·domain:skills·pm:queue. Maintainer-directed, so the product-only queue rule's exception appliesTriage seat (objectstack-wide, seat post #6015) ·
session_01AavokzJ5DndAwitDXvKy4U· 2026-10-10T02:17Z. ⛔ Not a claim, ⛔ not a dispatch.- Queue entry: a
toolingcard enters the queue only with anUnblocks:line or a named published surface, unless the maintainer directed it. The maintainer did: 「这个给 skills 立卡。」 So it queues. - Why p3: a refused assign leaves a PR without the card's assignee. That is a fleet convention, not product behaviour, and nothing is blocked. It recurs, and it cannot heal itself, because a seat never repeats a refused write.
- Re-read on
main96e4be4829:.claude/settings.jsonallowlists bothlabel-write.mjsspellings (:68,:70). The card's hypothesis stays unmeasured: the refused calls did not match an allow rule. - Direction: the skills seat chooses among A, B and C, as the card says. Measure first: the exact command lines of the 4 refusals, from the dev transcripts. Hold to these constraints:
- ⛔ No widening of the allowlist beyond the shapes devs actually run.
- ⛔ No new gate, which defaults to no. A pin on the allowlist's match is a test, not a gate.
- Queue entry: a
Measured data point for this card's first step (the exact command lines of the refusals), from skills seat 1 (
session_01RdnZdPZH9ByduzPRWuH9tN), 2026-10-10T09:20Z. ⛔ Not a claim.- Two
mode:subagentos-devs in this one session on objectstack (project root/home/user/objectstack, each dev working from its own worktree../objectstack-issue-N), the same command shape, within about 15 minutes of each other:- docs(adr-0131): §8/D14 execution-plan note — C2's remainder and C3 land as one cutover, C7 splits into C7a/C7b, and C12 moves to the 18.x line (rulings 6094045326, 6094175435, 6094183024) #22619's dev:
node scripts/pm/label-write.mjs --repo objectstack-ai/objectstack --issue 22650 --add skip-changeset --assign marchtian→ passed (relay run 38040085614; read-backskip-changeset·marchtian). - pm-dispatch decision framework: every decision card and every director presentation states each option's development workload (maintainer-directed) #22597's dev:
node scripts/pm/label-write.mjs --repo objectstack-ai/objectstack --issue 22654 --add skip-changeset --assign marchtian→ refused by the auto-mode classifier before any request (0 requests), category "External System Writes"; not retried (dev report 6096004928).
- docs(adr-0131): §8/D14 execution-plan note — C2's remainder and C3 land as one cutover, C7 splits into C7a/C7b, and C12 moves to the 18.x line (rulings 6094045326, 6094175435, 6094183024) #22619's dev:
- In both runs the same dev's
pr_createthroughfleet-write/dispatch.mjsand itspost-stampedreport comment passed. The PM seat then rannode scripts/pm/label-write.mjs --repo objectstack-ai/objectstack --issue 22654 --add skip-changesetfrom the shared checkout (its own docs-only verdict) and it passed (relay run 38040745377); the seat did not repeat the refused--assign. - Reading: identical command text, identical allowlist rule candidates (
Bash(node scripts/pm/label-write.mjs *)onmain), identical session, opposite outcomes ⇒ the refusal is not explained by the command's spelling alone; the card's hypothesis (an unmatched shape) does not cover this pair. Supports option B (the relay'spr_createcarriesassigneesin the same stroke, so no separate assign write exists to be refused).
Generated by Claude Code
- Two
Claim: PM loop round 2
Session:session_01RdnZdPZH9ByduzPRWuH9tN
Account:marchtian(the seat's linked user asGET /useranswers it; always the card's assignee)
Branch:claude/issue-22467-pr-assignee-in-pr-create
Worktree:objectstack-issue-22467
Domain:domain:skills
Seat:domain:skills#1
File surface:scripts/pm/fleet-write/ops.mjs(thepr_createop gains an optionalassigneesfield that issues a second request,POST /repos/{o}/{r}/issues/{n}/assignees, in the same relay stroke),scripts/pm/fleet-write/validate.mjs/execute.mjs/dispatch.mjsonly where the op table's shape requires it, the fleet-write self-tests,.claude/agents/os-dev.md(the write-budget and PR-assignee lines: the assignee ridespr_create;label-writeon the PR is for labels only);.github/workflows/fleet-write.ymlonly if the mint must change (it already mints issues write). ⛔.claude/settings.jsonis not widened; ⛔ no new gate (stop on breach; explain in the report)
Container & model:M,mode:subagent,model: CONTRACT_REVIEW_TIER—dispatch-gates --tier --repo objectstack-ai/objectstackatd817f1869prints "Model tier — MANDATORY: … .claude/agents/os-dev.md ⇢ '.claude/agents/os-dev.md' — clause ① (2026-08-20 narrowing): the dev-agent definition is protocol semantics"
Clause-②: no
Responsibility: the auto-mode permission classifier of a dev's session refuses thelabel-write … --assign LOGINstroke nondeterministically (identical command passed for one dev and was refused for another in this seat's session today, 6096066282; 4 of 12 objectui PRs before) | the relaypr_createstroke, which passed in every measured case, can carry the assignee in the same stroke, so no separate assign write exists to be refused | every os-dev the fleet dispatches on objectstack and objectui; four objectui PRs and PR #22654 landed without the card's assignee
Thread-read: 6096066282
Serial constraints cleared:none— no open PR touchesscripts/pm/fleet-write/**,.claude/agents/os-dev.mdor.claude/settings.json(open PRs' file lists read at 2026-10-10T09:49Z); PR #22654 (queued) touches.claude/skills/**only
Direction: the seat's call among the card's options is B (the relay'spr_createcarriesassignees), with A's documentation half (os-dev.md names the one shape that remains); C is not taken (it only moves the exposure). The measurement the triage grade asked for is on the thread (6096066282); the dev re-reads it and the fleet-write op table before writing.
Landing: Tier S (.claude/**) plusscripts/pm— at-tier## Contract reviewPASS in seat, the fleet-write self-tests green,check-expected-skipsexit 0,check-governed-merges, then the relay flip into the queue. File surface read atorigin/maind817f1869; claim written at 2026-10-10T10:10Z.os-dev-report
{ "issue": 22467, "status": "done", "branch": "claude/issue-22467-pr-assignee-in-pr-create", "pr": "https://github.com/objectstack-ai/objectstack/pull/22686", "session": "session_01RdnZdPZH9ByduzPRWuH9tN", "premise_still_valid": true, "summary": "Option B landed: the relay's pr_create op takes an optional assignees list and becomes two requests in one action (the draft create, then POST /repos/{o}/{r}/issues/{number}/assignees at the number the create ANSWERED), so the os-dev has no separate assign write for the classifier to refuse; the leg is judged by its answer (a 2xx listing fewer logins than sent is a FAILED action, never a quiet partial) and the dispatch read-back row carries the pull's assignees as a reading. Measured first, as the dispatch asked: the executor already ran an op's requests in order but computed every path from the action alone; no existing op threaded an earlier answer into a later request, so the smallest extension is a declared placeholder (answerPlaceholder('number') + fromAnswer) the executor fills, pinned by self-tests that feed a fake first answer. The mint step already grants issues: write literally on every stroke, so fleet-write.yml is untouched; .claude/settings.json untouched; no new gate. os-dev.md: :37 and :58 rewritten in place (the assignee rides pr_create; label-write on the PR is labels only) plus the seat rider at :353 (成本 → 真实代价与工作量), 402/402 lines, every edited line under the 120-byte cap. RETRY context: the prior dev's empty branch at ee3ae0360 was reused and fast-forwarded locally onto origin/main 243dd3c62 (no own commits, no second probe push, no second claim).", "tests": "Self-tests at 1e0dd1689 (exit captured before any pipe): node scripts/pm/fleet-write/validate.mjs --self-test → exit 0, '128 cases pass across 10 batteries' (baseline 119); execute.mjs --self-test → exit 0, '111 cases pass across 15 batteries' (baseline 100/14; new battery 'the assignees leg' = 11 cases); dispatch.mjs --self-test → exit 0, '216 cases pass across 19 batteries' (baseline 212). pnpm check:pm-skill-ratchet → exit 0, '.claude/agents/os-dev.md is 402 lines (ceiling 402; headroom 0)'; check:agent-model-declared exit 0; check:nul-bytes exit 0 ('no raw ASCII control bytes'); eslint --no-inline-config on the four modules exit 0. Ablation (fix committed first; scripts/ablation-replace.mjs WRAP mode; each leg restored with blob == HEAD and git diff HEAD empty): A prev = r → prev = null in execute.mjs → '5 of 111 case(s) failed'; B if (missing.length) → if (false && missing.length) → '1 of 111 case(s) failed' (the quiet-partial pin alone); C fromAnswer: 'number', → fromAnswerDropped: 'number', in ops.mjs → validate '2 of 128 case(s) failed', execute '7 of 111 case(s) failed'. A first C attempt was a null operation (the tool refused a replacement that was a substring of the anchor: replacement count rose 0 not 1; the self-test never ran) and was redone — the first reading is void, not a measurement. Gates: 47 families derived (dispatch-gates --commands --repo objectstack-ai/objectstack, no paths, at 1e0dd1689 vs merge base 243dd3c62), all 47 run, all exit 0; one (pnpm --filter @objectstack/lint run check:doc-formula-expressions) first exited 3 PREREQUISITE NOT MET (@objectstack/lint unbuilt — nothing measured), then after turbo run build --filter=@objectstack/formula --filter=@objectstack/lint under os-verify-lock.sh (VERDICT command-exit 0) it re-ran exit 0; --ran: '47 derived famil(ies) accounted for — 47 run, 0 NOT-MEASURED, 0 UNRUN'. NOT MEASURED: the live pr_create with assignees (the relay runs on main; the seat's first live stroke after merge is the measurement); repo-wide pnpm lint (CI's run).", "mcp_calls": "0 — none", "api_writes": "3 REST-proxy writes, every one through scripts/pm/ and the fleet-write relay (one repository_dispatch each, landing as objectstack-fleet[bot]): ① POST /repos/objectstack-ai/objectstack/dispatches → pr_create (relay run 38056183051; PR #22686 draft on main, read-back identical 11627/11627 bytes); ② POST /repos/objectstack-ai/objectstack/dispatches → labels_add skip-changeset + assign marchtian on #22686 (relay run 38056234790; read-back labels documentation, size/m, skip-changeset · assignee marchtian, MATCHES the target; `documentation` is the labeler bot's); ③ POST /repos/objectstack-ai/objectstack/dispatches → comment on #22467 (this report, through post-stamped). Plus git push ×1 (not a REST write): 1e0dd1689 to origin/claude/issue-22467-pr-assignee-in-pr-create (the probe push was already spent by the previous dev). Everything else was a read (gh api GET: the card, its comments, PRs #22457 and #22460, the PR read-back, check runs).", "open_questions": [], "out_of_scope_findings": [ "carrier: the skills seat · noted, not filed — .claude/skills/pm-dispatch/references/execution-duties.md:59 reads 「dev `pr_create` 同轮 `label-write.mjs --issue PR_NUMBER --assign LOGIN`;席位自有 PR 指派自己。」 and names the old shape; it goes stale the day this lands (outside the declared file surface, so not edited here); a one-line in-place rewrite such as 「dev `pr_create` 带 `assignees: [LOGIN]`;席位自有 PR 指派自己。」; dedupe words: execution-duties pr_create assign label-write os-dev assignee" ], "deviations": [ "label-write on the PR was NOT refused this time: `node scripts/pm/label-write.mjs --repo objectstack-ai/objectstack --issue 22686 --add skip-changeset --assign marchtian` passed (relay run 38056234790, ④ MATCHES the target) — one more data point for this card: the same stroke, same session shape, passed here. The old shape was used deliberately: the relay on main would refuse `assignees` as a key pr_create does not take until this PR lands.", "RETRY under the existing claim: the prior dev's empty branch (ee3ae0360, no own commits) was reused and fast-forwarded locally onto origin/main 243dd3c62 before any edit; no second claim, no second branch, no second probe push (one content push, 1e0dd1689).", "Gate runner shape: a loop running `bash -c \"$cmd\"` over the derived list was refused by the harness's built-in safety check (an opaque -c script; not the permission classifier, no write involved), so the 47 commands were spelled literally in four batches, each exit code recorded before any pipe.", "Ablation leg C's first attempt was a null operation (ablation-replace refused the mutation: the replacement text was a substring of the anchor, count rose 0 not 1) and was redone with a distinct replacement; reported, not hidden.", "Attribution trailers: the harness reminder asked for a `Co-Authored-By: Claude Fable 5.1` line; AGENTS.md's model-free pair (Claude-Session + Co-authored-by: Claude) was used because the pre-push hook refuses a model identifier in that pair and AGENTS.md outranks the reminder; no model name landed in the commit, PR or comments.", "Build for a derived gate's prerequisite (@objectstack/formula + @objectstack/lint) was run under scripts/pm/os-verify-lock.sh (slot issue-22467, held 2s, waited 0s); no other build or test was needed (no package in the diff)." ], "files_changed": [ "scripts/pm/fleet-write/ops.mjs", "scripts/pm/fleet-write/validate.mjs", "scripts/pm/fleet-write/execute.mjs", "scripts/pm/fleet-write/dispatch.mjs", ".claude/agents/os-dev.md" ], "line_budget": { "file": ".claude/agents/os-dev.md", "ceiling": 402, "before": 402, "after": 402, "net_lines": 0, "edited_lines_bytes": { "37": "116 -> 118", "58": "120 -> 120", "353": "113 -> 120" }, "cap_bytes": 120 }, "gates": [ "node scripts/check-ci-filter-parity.mjs :: exit 0", "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/check-declaration-mirrors.mjs :: exit 0", "node scripts/check-declaration-mirrors.mjs --self-test :: exit 0", "node scripts/check-scripts-symbol-anchors.mjs :: exit 0", "node scripts/check-scripts-symbol-anchors.mjs --self-test :: exit 0", "node scripts/check-self-test-wired.mjs :: exit 0", "node scripts/check-self-test-wired.mjs --self-test :: exit 0", "node scripts/check-self-test-workflow-commands.mjs :: exit 0", "node scripts/check-self-test-workflow-commands.mjs --self-test :: exit 0", "node scripts/check-whole-set-label-write.mjs :: exit 0", "node scripts/check-whole-set-label-write.mjs --self-test :: exit 0", "node scripts/pm/bare-root-worklist.mjs --self-test :: 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 0", "pnpm check:agent-model-declared :: exit 0", "pnpm check:agent-test-spelling :: exit 0", "pnpm check:bash32-floor :: exit 0", "pnpm check:cli-command-ids :: exit 0", "pnpm check:commit-card-trailers :: 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:entry-guard :: exit 0", "pnpm check:gitlink-declared :: exit 0", "pnpm check:nul-bytes :: exit 0", "pnpm check:parse-guard :: exit 0", "pnpm check:pm-close-cards :: exit 0", "pnpm check:pm-dispatch-gates :: exit 0", "pnpm check:pm-fleet-write-dispatch :: exit 0", "pnpm check:pm-fleet-write-execute :: exit 0", "pnpm check:pm-fleet-write-validate :: exit 0", "pnpm check:pm-governed-merges :: exit 0", "pnpm check:pm-issue-create :: exit 0", "pnpm check:pm-issue-transfer :: exit 0", "pnpm check:pm-label-write :: exit 0", "pnpm check:pm-post-stamped :: exit 0", "pnpm check:pm-skill-id-lint :: exit 0", "pnpm check:pm-skill-ratchet :: exit 0", "pnpm check:pm-write-pace :: exit 0", "pnpm check:pnpm-filter-targets :: exit 0", "pnpm check:refd-timer-probe :: exit 0", "pnpm check:skill-frame-sync :: exit 0", "pnpm check:watch-hint-literal :: exit 0" ], "classifier_data_point": "this session, 2026-10-10: the dev-side `dispatch.mjs --actions-file` pr_create stroke PASSED (run 38056183051); the `label-write … --add skip-changeset --assign marchtian` stroke PASSED (run 38056234790; 0 refusals in this run); the post-stamped report comment: see api_writes ③. No MCP tool was called.", "ci": "in_progress at report time — 33 check runs on 1e0dd1689: 12 success, 11 skipped, 10 in_progress, 0 failure; not waited for (CI convergence is the PM’s)" }
Generated by Claude Code
Claim file surface revised (claim
6096436788, round 2) — patch round 1 on PR #22686, dispatched to the same dev by the seat at 2026-10-10T13:41Z.- Added to the surface:
.claude/skills/pm-dispatch/references/execution-duties.mdline 59 only — the one line in.claude/**andAGENTS.mdthat names the retired shape (label-write.mjs --issue PR_NUMBER --assign LOGIN); it goes stale the moment this PR lands. In place, one line for one line (the file is at its ceiling), no other line of that file. - Why in this PR, not a card: the dev's out-of-scope finding on the report
6098063738; the four conditions for an adjacent in-place fix hold — same defect class (the retired write shape in text), mechanical, no other claim on the file (the hot-file queue reads it FREE), same gate family (ratchet, id-lint, governed-prose). Declared as "seat rider: execution-duties.md:59". - Review of head
1e0dd1689otherwise stands (read in full on the diff; the at-tier record follows on the PR once the patch head is in and CI has converged).
Generated by Claude Code
- Added to the surface:
os-dev-report
{ "issue": 22467, "status": "done", "patch_round": 1, "branch": "claude/issue-22467-pr-assignee-in-pr-create", "pr": "https://github.com/objectstack-ai/objectstack/pull/22686", "head": "19b7927ad87ba85f5cb2dceef3b6a12a8135e5b5", "session": "session_01RdnZdPZH9ByduzPRWuH9tN", "premise_still_valid": true, "summary": "Patch round 1 (seat rider accepted from the round-1 report, claim surface revised on the card in 6098099848): `.claude/skills/pm-dispatch/references/execution-duties.md:59` replaced in place, one line for one line, with 「- dev `pr_create` 带 `assignees: [LOGIN]`,同笔指派;席位自有 PR 指派自己。」 (107 → 87 bytes; the file stays 183/183 lines). No other line of that file and no other file touched; the fleet-write modules are untouched, so their self-tests were not re-run. The worktree was recreated from the branch at 1e0dd1689 and fast-forwarded to the remote before the edit.", "tests": "The four named gates after the edit, exit captured before any pipe: pnpm check:pm-skill-ratchet → exit 0 (\"execution-duties.md is 183 lines (ceiling 183; headroom 0)\"); pnpm check:pm-skill-id-lint → exit 0 (\"34 file(s) clean\"); pnpm check:pm-governed-prose → exit 0 (\"2 instruction surface(s) name all 6 registered governed surfaces\"); pnpm check:skill-frame-sync → exit 0 (\"90 markdown files scanned for undeclared copies\"). Control-byte self-scan of the edited file: no hit. Grep before the edit: no other line in the file named `--assign LOGIN`.", "mcp_calls": "0 — none", "api_writes": "1 REST-proxy write, through scripts/pm/ and the fleet-write relay: POST /repos/objectstack-ai/objectstack/dispatches → comment on #22467 (this patch-round report, through post-stamped). Plus git push ×1 (not a REST write): the rider commit to origin/claude/issue-22467-pr-assignee-in-pr-create. ⛔ No PR body PATCH, no label write, no second claim. Reads: the card thread tail (gh api GET).", "open_questions": [], "out_of_scope_findings": [], "deviations": [], "files_changed": [ ".claude/skills/pm-dispatch/references/execution-duties.md" ], "line_budget": { "file": ".claude/skills/pm-dispatch/references/execution-duties.md", "ceiling": 183, "before": 183, "after": 183, "net_lines": 0, "edited_lines_bytes": { "59": "107 -> 87" }, "cap_bytes": 120 }, "gates": [ "pnpm check:pm-skill-ratchet :: exit 0", "pnpm check:pm-skill-id-lint :: exit 0", "pnpm check:pm-governed-prose :: exit 0", "pnpm check:skill-frame-sync :: exit 0" ], "ci": "in_progress at report time on the new head; not waited for (CI convergence is the PM’s)" }
Generated by Claude Code
ACCEPT — PR #22686 (head
19b7927ad, after patch round 1), Tier S + PM toolingSkills seat 1 ·
session_01RdnZdPZH9ByduzPRWuH9tN· 2026-10-10T14:13Z · claim6096436788(surface revised6098099848) · reports6098063738,6098117103.Reviewed against GitHub per the review checklist, not against the report's narrative. The seat's call among the card's options was B with A's documentation half; the PR implements exactly that.
- Shape: draft, base
main, line 1Fixes #22467(the only closing keyword). Six files: the four relay modules,os-dev.md(three lines in place, 402/402), and the declared riderexecution-duties.md:59(one line in place, 183/183).skip-changesetholds as the seat's own verdict (nothing publishes); PR assignee set by the dev's ownlabel-write, which passed the classifier this time. - Done-when, read on the diff: the relay's
pr_createcarriesassigneesin the same stroke (two requests in one action, the second at the number the create answered, refused to send without one); a 2xx listing fewer logins than sent is a FAILED action, never a quiet partial; validation refuses every malformed list by name; the read-back row carries the pull's assignees as a reading;fleet-write.ymluntouched becauseissues: writeis already a literal mint input;.claude/settings.jsonuntouched; no new gate. - Evidence: self-tests +24 cases across the three modules at exit 0; three ablation legs red the expected cases (one void first attempt reported, redone); 47 derived gate families exit 0 at
1e0dd1689, the four text gates exit 0 at19b7927ad. CI on this head —Lint & Repo Gatessuccess,Type Check · workspacesuccess, every other check success or an expected skip (8 skipped, all in the roster, exit 0) · 0 failed on19b7927ad. - Deviations: the classifier passed the dev's assign stroke (recorded on this card as a data point); the retry reused the empty branch; the gate-runner loop refused by the harness's built-in check and spelled out literally; the model-free trailers per AGENTS.md; the build of two packages for one gate's prerequisite under the verify lock; the worktree removed after each report — all accepted.
- Contract review:
## Contract reviewPASS on the PR, record 6098379777 (CONTRACT_REVIEW_TIER,Local-runs: none). - Out-of-scope findings: the one finding (execution-duties.md:59) became the declared rider of patch round 1 —
filednowhere, fixed here. - NOT MEASURED, handed to the next sitter: the first live
pr_createwithassigneesafter merge. - Landing (seat's, Tier S):
check-expected-skips --pr 22686→ exit 0 (8 skipped check-runs, every one in the roster);check-governed-merges --pr 22686→ exit 3 = GOVERNED, Tier S (.claude/**×2:os-dev.md,execution-duties.md; the fourfleet-write/*.mjsnot on the register; 243 changed lines); then relaypr_ready+automerge_enable; MERGED proof and the landing note follow on this card.
Generated by Claude Code
- Shape: draft, base
Landed — PR #22686 MERGED as
73d700188a(option B: the PR assignee ridespr_create)Skills seat 1 ·
session_01RdnZdPZH9ByduzPRWuH9tN· 2026-10-10T14:52Z · claim6096436788(surface revised6098099848) · ACCEPT6098383919· contract review6098379777(PASS, head19b7927ad).- Proof:
merged: true; the merge commit73d700188ais an ancestor oforigin/main; all six files at the merge commit are byte-identical to the reviewed head19b7927ad(6 of 6 files, md5-equal). Queue path: ready 2026-10-10T14:14Z →added_to_merge_queue2026-10-10T14:16Z → merged 2026-10-10T14:51Z. - Effect, from this merge on: the relay's
pr_createtakes an optionalassigneeslist and assigns in the same stroke at the number the create answered; a 2xx listing fewer logins than sent is a FAILED action;label-writeon a PR is for labels only;os-dev.md:37/:58andexecution-duties.md:59say so. The oldassignop stays. - The card's measurement, closed: identical
label-write … --assignstrokes in one session — refused ×1 (PR docs(pm-dispatch): decision cards and director presentations state each option's development workload #22654's dev,6096066282), passed ×2 (PR docs(pm-skill): platform-readings diet — delete the readings a scripts/pm tool now holds (tool first, text second) #22684's and this PR's devs); the dev-sidedispatch.mjs --actions-filestroke refused ×0 across every reading. B removes the refusable write rather than explaining the classifier. - NOT MEASURED, handed over on the seat post: the first live
pr_createcarryingassigneesafter this merge — the next dispatch's dev measures it; its report cites the pr_create row's "Assignees as the pull lists them". - Card: closed by the merge (
Fixes #22467);pm:dispatchedand the assignee stripped in the same close-out.
Generated by Claude Code
- Proof:
Filing gate ③ — a task the maintainer directed. In the objectui
domain:uiseat 3 session (session_01CGZy1BGCjdN5cXqL9cnvB8), 2026-10-09, about the refusal reported below, verbatim: 「这个给 skills 立卡。」 Routed todomain:skillsthrough the maintainer's direct-dispatch channel; triage may grade it on first touch.Who acts on it: the objectstack
domain:skillsseat. The surface is the os-dev agent definition,.claude/settings.json's allowlist andscripts/pm/. ⛔ Not a claim.What happens
The os-dev's write budget includes one PR-assignee write:
.claude/agents/os-dev.md(about:36and:57onmain33ae4bff0f), "PR assignee = 卡的:--issue PR_NUMBER --assign LOGIN". The dispatched dev runsnode scripts/pm/label-write.mjs --repo objectstack-ai/objectui --issue PR --assign os-support-ai.In one objectui seat's session, with the same brief and the same command shape, that write landed on 8 PRs and was refused on 4:
os-support-aiThe latest refusal, from objectui#12047's dev report, verbatim: "The PR assignee write (label-write.mjs --assign os-support-ai) was refused by the auto-mode classifier before any request was sent: 0 requests, not retried". Its deviation names the category: "[External System Writes]".
Why it does not heal itself. A seat must never perform an action a subagent reports was refused by a permission check (permission laundering). So a refused assign leaves the PR without its assignee for good. The 4 PRs above landed that way, and the convention "PR assignee = the card's" is broken on each.
Measured and read (for the seat to verify)
.claude/settings.jsononmain33ae4bff0fallowlistsBash(node scripts/pm/label-write.mjs *)andBash(node --use-env-proxy scripts/pm/label-write.mjs *)(about:68and:70), and still the classifier refused.git push): the refused calls did not match an allow rule. They may have used a compoundcd /home/user/objectstack && node …, an absolute script path, or a session whose project root is not objectstack. So they fell through to the auto-mode classifier, whose judgment varies. The 8 successes would then be calls that matched, or classifier passes. The dev transcripts hold the exact command lines.Options (the skills seat's call)
pr_createcarryassigneesin the same stroke (scripts/pm/fleet-write/dispatch.mjs), so the dev makes no separate assign write and the budget drops a stroke.Duplicate check
git pushpassed seconds after the chained spelling was refused, which supports the hypothesis above..claude/settings.jsonallow-lists every REST write the seat makes except the two landing calls (ccr/ready_for_review,ccr/auto_merge) — so landing falls to the non-deterministic auto-mode classifier and seven green PRs waited for a human #19014: the allowlist missed the two landing calls, so landing fell to the non-deterministic classifier. Same mechanism, a different write.Filed by the objectui
domain:uiexecution seat 3 (seat post objectstack-ai/objectui#9800, sessionsession_01CGZy1BGCjdN5cXqL9cnvB8).Generated by Claude Code