Repository navigation
governance(AGENTS.md PD15): carve the fleet relay's closed Version-PR refresh stroke (release.yml with refresh_version_pr: true alone) out of the reserved release act, per ruling A on #23193 #23393
Description
Filing gate ③ (a maintainer ruling to execute). Ruling A on #23193, recorded 6160857949; the maintainer's word, verbatim: 「23209 A」. Filed by domain:devx seat 1 (session_012zwsLMxQBwHUYJgPyrpEvx, off duty; filing and release only). ⛔ Not a claim.
Reader who acts: the next domain:devx seat, or the director's Tier H batch. Surface: AGENTS.md Prime Directive 15 (:300-312 on main), one sentence. Governed (Tier H), because it is rule text. It needs an at-tier contract review (CONTRACT_REVIEW_TIER) before the queue, and an authorized approval. The PR stays a draft until then.
Governing text
PD15 reserves "⛔ workflow_dispatch-ing release.yml or any other publish-capable workflow". Ruling B item 5 on #23177 (6139629628) orders the relay to refresh the Version PR. That contradiction is why PR #23209 is held (6141962120).
What done looks like
- One sentence in PD15 carves out exactly one dispatch. A seat may dispatch
release.ymlthrough the fleet relay only withrefresh_version_pr: trueand no other input, as fenced byWORKFLOW_DISPATCH_INPUTSin PR tooling(pm): the fleet relay may start release.yml with refresh_version_pr true alone, so a seat refreshes the Version PR through the relay #23209. That stroke runs only theVersion PR maintenancelane. The publish lanes require!inputs.refresh_version_pr, and the sentence cites them as the reason the stroke cannot publish. - Every other reserved act stays reserved, word for word: publish, tag, Release, image, any other
release.ymldispatch or input combination, thereleaseenvironment approval, and merging, queueing or arming the Version Packages PR. - Order: this amendment lands first. Then PR tooling(pm): the fleet relay may start release.yml with refresh_version_pr true alone, so a seat refreshes the Version PR through the relay #23209 merges
main, is re-checked, and lands. Only then is ruling B item 5 executable. - A mirror check. If PD15's wording is mirrored anywhere a gate reads (a skill file or a check script), the dev finds it, names it, and keeps the mirror in step in the same PR.
- ⛔ No new gate.
四维(业务立场)
- 长远合理性:规则原文与机制一致,只有一种说法;例外写死成一个封闭的调用,不扩展到发布。
- 实际业务拉动:Version PR 每次刷新都要维护者手点;ruling B 第 5 条要的就是这个由席位代劳。
- 防 AI 犯错:席位按原文判断能做什么。原文一句话写清例外和边界,AI 不会把「可以刷新」读成「可以发布」。
- 创业阶段不扩散:一句话,不加门禁,不加新通道。
Dedupe: #23193 (the relay fence, PR #23209) and #23177 (ruling B) are the parents. This card covers the AGENTS.md hunk only. Dedupe words: PD15 amendment, relay refresh stroke, refresh_version_pr carve-out.
Activity
objectstack-fleet commentedon Oct 14, 2026
Claim: PM loop round 1
Session:
session_01B7mrh59UAGffd7s9asUfoYAccount:
sunhaolin(the seat's linked user asGET /useranswers it; the card's assignee)Branch:
claude/issue-23393-pd15-refresh-carveoutWorktree:
objectstack-issue-23393Domain:
domain:devxSeat:
domain:devx#2Ruling-ref: 6160857949 (ruling A on #23193, the maintainer's word 「23209 A」; re-read this round)
File surface:
AGENTS.mdPrime Directive 15 only, one sentence. If a gate or a skill file mirrors PD15's reserved-act wording, the dev names it, and keeps it in step only when it is a true mirror; the report says which. Governed (Tier H): draft until an at-tier contract review and an authorized approval. ⛔ No edit torelease.yml,scripts/pm/fleet-write/**or PR #23209. ⛔ No dispatch of any release workflow. Stop on a breach and explain it in the report.Container & model:
S,mode:subagent,model: opus(dispatch-gates --tier: no path-derived mandate; default tier; rule text a seat acts on). The at-tier contract review follows at the seat's review.Clause-②: no
Responsibility:
n/a — not a defect cardThread-read: none
Serial constraints cleared:
none— none of the 9 open PRs touchesAGENTS.md, and no in-flight claim declares it. PR #23209 (#23193,pm:blockedon this card) touchesscripts/pm/fleet-write/**only and lands after this one, per the ruling's order. Read at 2026-10-14T08:05Z.Generated by Claude Code
objectstack-fleet commentedon Oct 14, 2026
os-dev-report
{
"issue": 23393,
"status": "done",
"branch": "claude/issue-23393-pd15-refresh-carveout",
"pr": "#23436",
"session": "session_01B7mrh59UAGffd7s9asUfoY (the session the relay derived from this container; subagent = parent's)",
"premise_still_valid": true,
"summary": "PD15's reserved clause now reads: workflow_dispatch-ing
release.ymlin any lane but the fleet-write relay's refresh stroke (refresh_version_pr: truealone), or any other publish-capable workflow. That is ruling A's own shape (option A in 6141962120), with no constant named, because WORKFLOW_DISPATCH_INPUTS is not on main (only PR #23209 head 961ec71 has it). On main the relay allowlist is shard-timings-refresh.yml alone, so the admitted stroke is refused with zero writes until #23209 lands, and the PR body names that dependence. One PM assumption was falsified: line_budget is not n/a. AGENTS.md has a ratchet ceiling of 1116 at headroom 0. The three-part sentence with the by-job warrant measures +3 lines with the gate's own wrapLine at the paragraph's 118 B width, so I landed the ruling's 91 B clause at 0 lines (AGENTS.md stays 1116) and moved the warrant to the PR body. The card's claim that the publish lanes require!inputs.refresh_version_prholds literally for release-integrity and stale-prompts, on the force branch for full-battery and publish (both also need release-integrity, which a refresh skips), and only through outputs for docker. There is no true mirror of PD15's list anywhere, so nothing moved in step and there is no Tier S hunk.","tests": "No package was touched, so there is no build and no test. Proof (scratchpad proof.mjs, base 4f81f31 vs head 9b68a60): the whitespace-normalised PD15 paragraph at head equals base plus exactly the inserted clause (true); control without the trailing comma answers false. Paragraph 15 to 15 lines, AGENTS.md 1116 to 1116. Line cost was measured with check-skill-line-ratchet's exported wrapLine at width 118: the base paragraph re-wraps to its current 15 lines (0 lossless headroom, 99 B last-line slack); the ruling clause (+91 B) gives delta 0; the PM-route three-part sentence (367 B) gives delta +3. Gates at head 9b68a60, every one exit 0 with the code captured before any pipe; verdict lines quoted under gates.",
"gates": "Derived by
node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack(EXIT=0; 1 path vs merge base 4f81f31, +9/-9). Each of these exited 0: check-closing-keyword-parity ('OK (3 parsers agree ...)'), the same with --self-test, check-comment-mask-corpus ('9045 files, 0 disagree'), check:agent-test-spelling ('0 violations'), check:docs-audit-scope, check:driver-memory-census ('OK'), check:gitlink-declared ('OK'), check:nul-bytes ('OK ... no raw ASCII control bytes'), check:pm-governed-merges, check:pm-governed-prose ('name all 7 registered governed surfaces'), check:pm-skill-id-lint ('87 file(s) clean'), check:pm-skill-ratchet ('AGENTS.md is 1116 lines (ceiling 1116; headroom 0)'), check:refd-timer-probe ('OK'), check:required-contexts ('5 required context name(s) pinned'), check:watch-hint-literal. Two more outside the derivation also exited 0: check-published-list-mirrors ('1 published list mirror(s) match') and check-skills-token-ratchet.--ranreconcile: '15 derived, 15 run, 0 NOT-MEASURED, 0 UNRUN' (a derived zero, every entry recorded with its exit code). NOT MEASURED: scripts/check-required-contexts.mjs --verify-required-set (it takes a value from the workflow). The 4 type-check lanes do not apply because no TypeScript was touched. CI on the PR: in_progress at report time (36 check runs: 16 success, 12 skipped, 8 in progress; Governed Surface Queue Guard and Test Core success).","line_budget": "AGENTS.md 1116 to 1116 lines against a ceiling of 1116 (headroom 0 before and after). 0 lines were added, the ceiling was not touched, and no raise was requested.",
"files_changed": [
"AGENTS.md"
],
"deviations": [
"Suggested route and card wording: the sentence carries the admitted dispatch and its uniqueness. It does not carry the by-job warrant. The warrant (job names and guards) is in the PR body. Reason: ratchet headroom 0, the warrant costs +3 lines, and a raise needs a maintainer ruling that ruling A did not give. PD15's own second paragraph also says the YAML is not a barrier, so the authority is the ruling.",
"The suggested route said 'one sentence appended'. I made an in-clause amendment instead, which is ruling A's literal shape ('the reserved act becomes ...'). The other 7 changed lines of the hunk are a forced re-flow; the words are unchanged (proved).",
"PM assumption 'line_budget n/a' is falsified by measurement (see line_budget).",
"
git fetch origin pull/23209/head:refs/os-dev-23393/pr23209auto-followed tags into the shared .git, including create-objectstack@17.7.0 and 18.0.0-next.0 from origin. My private ref is deleted at cleanup. The tags are origin's own and are left as they are.","The harness attribution reminder offered a model-bearing Co-Authored-By trailer and a two-line PR footer. I followed AGENTS.md instead: the model-free trailer pair, and one session-URL footer block on the PR body."
],
"mcp_calls": "0 (no MCP tool used)",
"api_writes": "3 seat-side writes, each POST /repos/objectstack-ai/objectstack/dispatches through the fleet-write relay, as objectstack-fleet[bot]. (1) pr_create: relay POST /repos/objectstack-ai/objectstack/pulls (draft) and POST /repos//issues/23436/assignees (sunhaolin), run 38535237958, read-back identical 9515 B. (2) labels_add via scripts/pm/label-write.mjs: relay POST /repos//issues/23436/labels (skip-changeset), run 38535330641; the read-back matches (bot-added documentation and size/s are left alone). (3) This os-dev-report comment via scripts/pm/post-stamped.mjs: relay POST /repos//issues/23393/comments. git push is not counted. There was no PR body PATCH.",
"open_questions": [
{
"question": "PD15's bold lead still says no AI seat '... or triggers a release workflow'. With the clause amended, the lead reads stricter than the list it summarises. Align it in this PR (a second governed edit beyond ruling A's literal clause)?",
"options": [
"A: leave the lead. The list ('What is reserved is the release act itself') is the definition and names the exception. A literal reader of the lead fails closed: the refresh is not done and nothing publishes. Cost 0.",
"B: 'triggers a release workflow' becomes 'triggers a release' (9 B shorter, 0 lines, no gate). The list still reserves every other release.yml lane and any other publish-capable workflow by name."
],
"recommendation": "B, at the seat's word in its contract review. Business need: ruling B item 5 needs seats to run the refresh after each fix, and #23209's own dev already read PD15 literally and stopped. Long-term: one rule text for one mechanism is the ruling's stated rationale. AI error: a bold lead that contradicts its own list invites split readings, while the list keeps full coverage. Startup focus: 0 lines and no gate. A is safe if the seat holds ruling A to the clause only."
}
],
"out_of_scope_findings": [
"none (class a/b/c): no reproducible defect, contract violation or authoring trap found. docker's lack of an input guard is documented by design in release.yml (its outputs are the gate).",
"Mirror search: no true mirror. Queries and hit counts are in the PR body Acceptance notes. Pointers left unmoved: .claude/skills/pm-dispatch/SKILL.md:34; scripts/release-github-releases.mjs:76,842; the scripts/pm/check-skill-line-ratchet.mjs:1395 comment; ADR-0125 and ADR-0082 citations; docs/releases-maintenance.md (the maintainer runbook). carrier: none. Noted in Acceptance notes and not filed.",
"PD13: ADR-0125 line 95 ('prohibitions stand as written') predates refresh_version_pr. No ADR reserves the refresh stroke, so nothing is superseded. Noted, not filed."
]
}
objectstack-fleet commentedon Oct 14, 2026
Seat answer to the report's
open_questions[0]·domain:devx#2·session_01B7mrh59UAGffd7s9asUfoY· 2026-10-14T08:32Z. ⛔ Not a verdict; the review of PR #23436 continues after this patch round.Answer: B. PD15's bold lead changes from "… or triggers a release workflow" to "… or triggers a release". Nothing else in the lead moves.
6160857949) adopts the recommendation in6141962120: "one rule text that matches one mechanism". With only the clause amended, the lead still forbids what the amended list admits, because the relay's refresh stroke does triggerrelease.yml, a release workflow. That is the same contradiction that held PR tooling(pm): the fleet relay may start release.yml with refresh_version_pr true alone, so a seat refreshes the Version PR through the relay #23209. Aligning the lead executes the ruling as given. It is not a new decision, so it is answered here and not escalated.release.ymllane and every other publish-capable workflow stays reserved by name. The dev measured B at 0 lines; it must still re-measure on the new head (AGENTS.mdceiling 1116, headroom 0).claude/issue-23393-pd15-refresh-carveout. The maintainer's quick-read on the PR names both edits. The approval is the maintainer's either way.Generated by Claude Code
objectstack-fleet commentedon Oct 14, 2026
os-dev-report
{
"issue": 23393,
"addendum": "patch round 1: PD15 lead aligned (seat answer B)",
"previous_head": "9b68a60e2f",
"new_head": "e7712a4402",
"status": "done",
"branch": "claude/issue-23393-pd15-refresh-carveout",
"pr": "#23436",
"session": "session_01B7mrh59UAGffd7s9asUfoY (the session the relay derived from this container; subagent = parent's)",
"premise_still_valid": true,
"summary": "Executed the seat's answer B (card comment 6164269656). PD15's bold lead now reads '... or triggers a release, and none merges the Version Packages PR'. The one word 'workflow' was removed, and nothing else in the lead or the file changed. Commit e7712a4 on top of 9b68a60; pushed. AGENTS.md stays at 1116 lines (ceiling 1116, headroom 0). Line 302 is shortened in place with no re-flow, and the PD15 paragraph stays at 15 lines. The PR body now names both edits under 'What changes', each with before, after and why ('one rule text that matches one mechanism', the seat's answer), and the 维护者速读 names both edits with 席位意见 left empty. I read the stored body back in full and it is identical. GitHub reports the PR mergeable, so main was not merged. No ready flip, no review request and no label write.",
"tests": "No package was touched, so there is no build and no test. Word diff vs merge base 4f81f31 at head e7712a4 (
git diff -w --word-diff=porcelain) shows exactly '-release workflow,' / '+release,' and the inserted clause '+in any lane but the fleet-write relay's refresh stroke (refresh_version_pr: truealone),'. Whole-file proof (scratchpad proof2.mjs): whitespace-normalised AGENTS.md at head equals base with exactly those two edits applied, and answers true; control without the clause's trailing comma answers false. Lines: base 1116, head 1116. The gate's wrapLine at width 118 gives a PD15 paragraph of 15 lines, 1670 B (base 1588 + 91 - 9), with 12 B of slack on the last line.","gates": "Re-derived at head e7712a4 with
node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack(EXIT=0; 1 path vs merge base 4f81f31, +10/-10). It is the same 15 commands as at 9b68a60 (diff empty). All 15 exited 0, exit codes captured before any pipe: check-closing-keyword-parity ('OK (3 parsers agree ...)'), the same with --self-test, check-comment-mask-corpus, check:agent-test-spelling, check:docs-audit-scope, check:driver-memory-census, check:gitlink-declared, check:nul-bytes ('OK ... no raw ASCII control bytes'), check:pm-governed-merges, check:pm-governed-prose ('name all 7 registered governed surfaces'), check:pm-skill-id-lint ('87 file(s) clean'), check:pm-skill-ratchet ('AGENTS.md is 1116 lines (ceiling 1116; headroom 0)'), check:refd-timer-probe, check:required-contexts, check:watch-hint-literal. Also run outside the derivation: check-published-list-mirrors ('1 published list mirror(s) match'), exit 0.--ranreconcile: '15 derived, 15 run, 0 NOT-MEASURED, 0 UNRUN' (a derived zero). NOT MEASURED: scripts/check-required-contexts.mjs --verify-required-set (it takes a value from the workflow). CI: previous head 9b68a60 converged (38 check runs: 26 success, 12 skipped, 0 failures). New head e7712a4 is in_progress at report time (41 runs: 27 success, 12 skipped, 2 in progress, 0 failures).","line_budget": "AGENTS.md 1116 to 1116 lines against a ceiling of 1116 (headroom 0). The lead edit costs 0 lines and the ceiling was not touched.",
"files_changed": [
"AGENTS.md"
],
"deviations": [
"The worktree objectstack-issue-23393 had been removed at the end of round 0, as part of the dev contract's cleanup step. I re-added it at the same path on the same local branch; the local and remote heads were both 9b68a60 and the remote was untouched. No second worktree exists. I reinstalled dependencies there, and it is removed again after this report.",
"The PR body edit was sent without a footer, following the edit-side prescription in platform-readings. The stored body is byte-identical to what was sent (10995 B) and carries NO footer: the relay's PATCH /issues/23436 appended nothing, against the +58 B bare footer platform-readings records for REST PATCH. Attribution now lives in the body prose ('Patch round 1 (session_01B7mrh59UAGffd7s9asUfoY ...)'). I did not send a second write to add a footer, because the budget was one edit."
],
"mcp_calls": "0 (no MCP tool used)",
"api_writes": "Patch round 1: 2 seat-side writes, each POST /repos/objectstack-ai/objectstack/dispatches through the fleet-write relay as objectstack-fleet[bot]. (1) issue_patch body: relay PATCH /repos//issues/23436 (the PR body), run 38536589998, read-back identical 10995 B. (2) This addendum comment via scripts/pm/post-stamped.mjs: relay POST /repos//issues/23393/comments. git push of e7712a4 is not counted. No label, ready, review or assignee write. Round 0's 3 writes are in comment 6164234572.",
"open_questions": [],
"out_of_scope_findings": [
"none (class a/b/c).",
"Platform reading, not a defect: a body-only PATCH /repos//issues/23436 on a PR, executed by the fleet-write relay as objectstack-fleet[bot] (run 38536589998), stored the body exactly as sent with no footer appended. platform-readings records a +58 B bare footer for REST PATCH on the same surface; the App-identity relay route differs. carrier: PM (platform-readings owner). Noted here, not filed."
]
}
objectstack-fleet commentedon Oct 14, 2026
ACCEPT (Tier H, awaiting an authorized approval) — PR #23436 ·
domain:devx#2· 2026-10-14T08:53ZSeat review of PR #23436 against GitHub, at head
e7712a4402on merge base4f81f31526. Reports:6164234572(round 0) and the patch-round addendum. Claim:6163826101. Ruling-ref:6160857949(A on #23193).Checklist conclusions
main. Body line 1 isFixes #23393. tooling(pm): the fleet relay's workflow_dispatch can startrelease.ymlwithrefresh_version_pr: trueonly, so a seat refreshes the Version PR through the relay (#23177 ruling B, item 5) #23193 and [Decision] 门禁过严拖慢发版:一轮 CI 暴露全部红、成本型棘轮加容差带、门禁瘦身审计 #23177 are referred to as "remains open", and no closing keyword sits beside them.Clause-②: nois at line start. The assignee issunhaolin.pulls/23436/files):AGENTS.mdonly, +10 / −10. Governed, Tier H.git diff -w --word-diff=porcelainagainst4f81f31526): exactly two edits.release.ymlbecomesrelease.yml"in any lane but the fleet-write relay's refresh stroke (refresh_version_pr: truealone),".AGENTS.mdstays at 1116 lines, the ratchet ceiling, so no raise is asked.mainhas not touchedAGENTS.mdsince the base.CONTRACT_REVIEW_TIERbecause this is governed rule text. Isolated at-tier review PASS on this head, record6164595133on the PR (Head-shae7712a4402550e361e9c849a09ebee2d036fb768,Local-runs: none).e7712a4402: 41 check-runs, 29success, 12skipped, none red.skip-changesetis right, sinceAGENTS.mdpublishes nothing.Terminal, per the governed path: this ACCEPT is the review record. In the same act the PR gets
needs-user-decisionand the final 维护者速读 comment, and review is requested from both authorized approvers. The PR stays a draft. Once an authorized approval is on record, this seat lands it: the pre-landing check, the label cleared,pr_ready,automerge_enable. Then #23209 follows, per ruling A's order.Generated by Claude Code
objectstack-fleet commentedon Oct 14, 2026
Landed — PR #23436 →
5b31fb9297·domain:devx#2· 2026-10-14T09:23Zos-zhuang, APPROVED on heade7712a4402). The maintainer readied it and armed auto-merge.mainby blob identity:AGENTS.mdonorigin/mainis19f3a79d3b, the same blob as the PR heade7712a4402. PD15 now carries the carve-out ("the fleet-write relay's refresh stroke") and the aligned lead.be50633b(@objectstack/plugin-security#build,TS2304 AnchorBindingContext; disposition6164873246). This PR was never removed. Its group rebuilt green after fix(cli,plugin-security): os migrate organization-ownership --apply resolves the built-in positions and the everyone baseline #23433 left the queue.completedbyFixes #23393.pm:dispatchedcomes off in this act.release.ymlwithrefresh_version_pr: trueonly, so a seat refreshes the Version PR through the relay (#23177 ruling B, item 5) #23193) mergesmain, is re-checked and lands. Ruling B item 5 becomes executable only after that.Generated by Claude Code