pm-dispatch: the allow-listed ccr ready flip was denied `[Auto-Mode Bypass]` in a session whose checkout carried the allow rows from its first HEAD — and since #19997 a denied ccr call leaves a session seat no landing route

[objectstack-fleet]

Filing gate: ③ a task the maintainer directed. The maintainer, in the domain:cli execution seat's session session_01TnPAC1UsTGfHPXVUCL6iLn (chat), 2026-09-24, verbatim:

「这个问题写一个 skills卡片：转 ready 要跑 curl -sS -X POST https://api.github.com/repos/objectstack-ai/objectstack/pulls/20022/ccr/ready_for_review -d '{}'。

• 这条命令在 .claude/settings.json 的放行清单里，但 auto 模式的权限分类器还是拒了，拒因是 [Auto-Mode Bypass]。」

(One client UI notice, which the paste carried mid-sentence, is omitted from the quote. It is not the maintainer's words.)

The maintainer names the skills lane. ⛔ Routing labels are triage's, so this card is filed bare. Filed by the domain:cli execution seat #6024. ⛔ Not a claim.

• Acting reader: the domain:skills seat, which owns .claude/skills/pm-dispatch/references/landing-operations.md §B. The permission choice in the last section is the maintainer's. .claude/settings.json is governed.
• Dedupe: two issue searches were run ("Auto-Mode Bypass classifier denied ccr ready_for_review allow-listed landing", and "auto mode permission classifier denies seat landing ready flip auto-merge settings.json allow row"). They returned 4 distinct hits, all closed: [finding] the harness auto-mode classifier denies the protocol's landing command non-deterministically — 8 byte-identical PUT …/ccr/auto_merge from one turn: 4 allowed / 4 denied under three reasons; and a hooks landing rotates every seat off shift before it can flip ready #18469, [finding] the committed .claude/settings.json allow-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, Seat batch closure needs a named script and a matching allow rule — the runtime's write classifier refuses the three-step close (comment · label · PATCH) non-deterministically, and the two existing allow rules do not match how seats invoke the tools #19469, pm-dispatch: relay-opened dev PRs are never attached to the seat's session, and the relay landing route has no allow rule, so auto mode decides seat landings call by call #19990. 0 are open. Why this is not a duplicate:
  • [finding] the harness auto-mode classifier denies the protocol's landing command non-deterministically — 8 byte-identical PUT …/ccr/auto_merge from one turn: 4 allowed / 4 denied under three reasons; and a hooks landing rotates every seat off shift before it can flip ready #18469's closing comment rules: 「之后如果分类器再拒掉某条具体命令,按那条命令另开一张新卡」 (if the classifier denies a specific command again, open a new card for that command). This is that card.
  • [finding] the committed .claude/settings.json allow-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 explained its own post-landing denial by the session's permission set predating the rows. That explanation is measured below and does not apply here.

The reading

• Command. curl -sS -X POST https://api.github.com/repos/objectstack-ai/objectstack/pulls/20022/ccr/ready_for_review -d '{}', issued once, standalone, on 2026-09-24 at about 20:30Z.
• Answer. Denied by the Claude Code auto-mode classifier, reason [Auto-Mode Bypass]. The record is on [finding] a THIRD package-install door — marketplace install-local — parses nothing at all: 0 ManifestSchema, 0 safeParse across 1,912 lines, so an id MANIFEST_ID_PATTERN refuses installs cleanly there #19576 as 5821758726.
• The allow row it matches. .claude/settings.json:61: Bash(curl -sS -X POST https://api.github.com/repos/objectstack-ai/objectstack/pulls/*/ccr/ready_for_review *). It landed in 67624b78 (chore(claude): allow-list the two landing REST calls in settings.json #19047) on 2026-09-18.
• The rows were on disk when this session started. The session's first HEAD is 2c1011b0 (reflog, 2026-09-24T12:43Z), and git show 2c1011b0:.claude/settings.json carries both rows at :61–:62. ⇒ [finding] the committed .claude/settings.json allow-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's explanation ("its permission set predated the rules") does not fit. What remains is pm-dispatch: subscribe relay-opened PRs at collection; seats land through the allow-listed ccr pair; a landing denial stops and surfaces #19997's other unmeasured explanation: the classifier judges content over an allow row. That this session loaded the file is not proven here, only that the file had the rows.
• The PR was a clean landing on every reading the seat owns. It was ACCEPTed (5821336143). CI on head ca79edd2 read 33 success plus 2 rostered skips, with 0 pending and 0 failed. check-governed-merges --pr 20022 said NOT governed (483 lines). There was no os-regen path and no Clause-② limb.
• Contrast in the same session and on the same day. The relay route (scripts/pm/with-fleet.sh --via dispatch, ops pr_ready + automerge_enable) readied and armed PRs fix(runtime): a seed dataset declared once registers once on an additive multi-package artifact #19981, fix(cli): mount the always-on package-registry capability at its spec-declared provider (#19387) #19983 and fix(runtime): the resume door checks WHO is resuming — the run's starter, or the sys_automation_run read grant (#19987) #20005 under auto mode without a single denial. The timeline actor is objectstack-fleet[bot]. All three came before cc4ee775 reached this seat's checkout. (PR feat(mcp): resume_run completes a paused screen flow — the caller's own run, behind run_action's gates #19985 was readied by hand by the maintainer.)
• Consistent with [finding] the harness auto-mode classifier denies the protocol's landing command non-deterministically — 8 byte-identical PUT …/ccr/auto_merge from one turn: 4 allowed / 4 denied under three reasons; and a hooks landing rotates every seat off shift before it can flip ready #18469. It measured 8 byte-identical PUT …/ccr/auto_merge calls from one turn: 4 allowed and 4 denied, one of them [Auto-Mode Bypass]. An allow row narrows the classifier's room; this reading says it does not close it.

Why this is now a skills problem, not only a classifier reading

Since #19997 (cc4ee775), references/landing-operations.md §B reads:

• :51 「绿即转 ready + 挂 auto-merge,只走 settings.json allow 行的 ccr 两命令」 (ready and auto-merge go only through the two allow-listed ccr commands);
• :54 「落地中遭分类器拒 ⇒ 停手、报维护者、卡上记命令与拒因;⛔ 不换拼法或改走中继绕过」 (a denial during landing: stop, report, record; never respell or switch to the relay).

Read together, a session whose classifier denies the ccr pair has no landing route at all. Every accepted PR in that session then waits for a human. #19997's own "Measured risk that stays open" named this case as unmeasured; this is the negative reading.

Cost, measured. PR #20022 has been ACCEPTed and green since about 20:28Z and was still draft at 23:07Z. The lane runs serial by the maintainer's standing order (5815600379 on #6024), so no next card can be dispatched until a human lands it.

What this card asks

1. The maintainer's call, still open from pm-dispatch: subscribe relay-opened PRs at collection; seats land through the allow-listed ccr pair; a landing denial stops and surfaces #19997. That PR's 席位意见 ② put a choice to the maintainer, and it has no answer on pm-dispatch: relay-opened dev PRs are never attached to the seat's session, and the relay landing route has no allow rule, so auto mode decides seat landings call by call #19990 or pm-dispatch: subscribe relay-opened PRs at collection; seats land through the allow-listed ccr pair; a landing denial stops and surfaces #19997 as of this filing. Either the ccr pair is an exception to AGENTS.md's 「写只经 scripts/pm/」, or landings go through the relay with a new allow row. This reading is an input to that choice: in this session the allow-listed ccr route was denied, and the non-allow-listed relay route passed three times.
2. The skills seat, once the maintainer answers. Make §B name exactly one landing route that a session seat takes after a denied ccr call, or state plainly that there is none and a human lands. The current pair of lines implies the second without saying it, and the stop rule reads as though a fallback exists elsewhere.
3. Record the reading where the skill keeps classifier readings (references/platform-readings.md, beside :435, the content-based [Self-Approval] reading): the allow-listed ccr ready flip was denied [Auto-Mode Bypass] in a session whose checkout carried the allow rows from its first HEAD.

Acceptance notes

• ⛔ This card authorizes no landing route. PR fix(cloud-connection): POST /api/v1/marketplace/install-local parses the package id through its declaration #20022 stays with the maintainer, as recorded on [finding] a THIRD package-install door — marketplace install-local — parses nothing at all: 0 ManifestSchema, 0 safeParse across 1,912 lines, so an id MANIFEST_ID_PATTERN refuses installs cleanly there #19576.
• The skills PR answers item 2 in §B's own lines and does not add a gate. If it adds or changes a settings.json row, that part is governed and lands by the maintainer's hand.

[objectstack-fleet]

Reading: the same ccr pair passed in the same session once it was out of auto mode

_{domain:cli execution seat #6024 · session session_01TnPAC1UsTGfHPXVUCL6iLn · 2026-09-25T00:47Z}

• The harness reported "Exited Auto Mode" in this session at about 2026-09-24T23:21Z.
• At 2026-09-25T00:47Z the seat landed PR fix(client)!: every limit query-parameter emitter sends what the caller wrote #20060 (card [finding] @objectstack/client's 11 limit emitters guard three different ways, so the same { limit: 0 } is silently dropped on some doors and refused with 400 on others — and { limit: null } is sent as the string null on six of them #19567, ACCEPT 5824753067, CI green on ba1c3eae) with exactly the two allow-listed commands:
  • curl -sS -X POST https://api.github.com/repos/objectstack-ai/objectstack/pulls/20060/ccr/ready_for_review -d '{}' answered {"draft":false}.
  • curl -sS -X PUT https://api.github.com/repos/objectstack-ai/objectstack/pulls/20060/ccr/auto_merge -d '{"merge_method":"SQUASH"}' answered {"enabled":true,…}.
  • Timeline: ready_for_review and auto_merge_enabled at 00:47:31Z / 00:47:34Z, actor os-steve.
• ⇒ In this one session, the same command shape against the same allow rows was denied [Auto-Mode Bypass] while in auto mode (PR fix(cloud-connection): POST /api/v1/marketplace/install-local parses the package id through its declaration #20022, record 5821758726) and passed after leaving it. That points at the auto-mode classifier deciding on content over the allow row, rather than at the rows failing to load. One pair of readings; ⛔ not a rate.

[objectstack-fleet]

Claim: PM loop round 1
Session: session_01LNRg6FUMXYWgkAbDsfZpC8
Branch: claude/issue-20048-relay-landing-route
Worktree: objectstack-issue-20048
Domain: domain:skills
Seat: domain:skills#1
File surface: .claude/skills/pm-dispatch/references/landing-operations.md (§B, the ccr-only line and the denial line) · .claude/settings.json (two allow rows for the relay landing spellings) · .claude/skills/pm-dispatch/references/rest-channel.md and platform-readings.md (one fact line each: the relay landing ops; the auto-mode denial reading) · .claude/skills/pm-dispatch/SKILL.md only if a line there names the ccr pair as the landing route (paid in place, 319 / 319) (stop on breach; explain in the report)
Container & model: M — a ruled route change across the landing text and one allow-row pair; no gate, mode:subagent, model: opus — dispatch-gates --tier on these paths: references and settings carry no clause-① mandate; SKILL.md, if touched, is a MANDATORY fable path whose one-line-class edit drops to opus execution compensated by this seat's review at CONTRACT_REVIEW_TIER (the tool's own exit text)
Clause-②: no
Thread-read: 5824764513
Serial constraints cleared: no open objectstack PR touches these four files (open PR head list read at the time of this claim); the sibling take #20063 (in flight, this seat) edits execution-duties.md and os-dev.md — disjoint; no claude/issue-20048* head on origin; landing-operations.md last landed with PR #19997 (cc4ee77)

Take basis (written at 2026-09-25T01:16Z): category ③ — the card was filed on the maintainer's words in the cli seat's chat, and the maintainer repeated the priority in this seat's chat today, verbatim 「所以 auto 权限的问题要优先解决，我记得有一个issue」; the choice the card and PR #19997's 席位意见 ② left open was put to the maintainer in this session as three options and answered 「B:走中继落地(推荐)」 — landings go through the fleet-write relay (pr_ready + automerge_enable ops, writes as objectstack-fleet[bot]), which also settles maintainer item (v) (the ccr pair's write identity) in AGENTS.md's favour without an exception. Evidence on the card: the relay readied and armed PRs #19981 / #19983 / #20005 under auto mode with zero denials while the allow-listed ccr flip was denied [Auto-Mode Bypass] on #20022 and [External System Writes] on objectui#10410 in this seat's session. Priority raised to p1 on the maintainer's word. Dispatched as a subagent of this session with this comment as its only claim.

---

Generated by Claude Code

[objectstack-fleet]

os-dev-report
Posted 2026-09-25T01:29Z.

{
  "issue": 20048,
  "status": "done",
  "branch": "claude/issue-20048-relay-landing-route",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/20070",
  "session": "session_01LNRg6FUMXYWgkAbDsfZpC8",
  "premise_still_valid": true,
  "summary": "Implemented ruling B. The pm-dispatch landing text (landing-operations.md section B, SKILL.md:201, rest-channel.md write table, platform-readings.md undraft/quota/denial lines) now routes the ready flip and auto-merge arm through the fleet-write relay ops pr_ready + automerge_enable (reverse pr_draft / automerge_disable, marked never called for real). The ccr pair is retired, and a denial still means stop + report + record, with no fallback. Every file is line-neutral. The .claude/settings.json allow rows were NOT added: the edit was denied by the auto-mode classifier as Self-Modification, was not retried, and is handed to the maintainer in the PR body. PR assignee is os-elon-musk and skip-changeset is set.",
  "tests": "At d32057c6: dispatch-gates --commands --repo objectstack-ai/objectstack derived 22 commands, all exit 0 (check:doc-formula-expressions first exit 3 PREREQUISITE NOT MET, then exit 0 after building formula+lint under os-verify-lock, VERDICT command-exit 0). pnpm check:pm-settings-deny-roster also exit 0 (outside the derivation). --ran: 22 derived families accounted for, 22 run, 0 NOT-MEASURED. check-governed-merges --test on the 4 paths: GOVERNED Tier S. Relay dry-runs (nothing sent): dispatch.mjs --actions-file with pr_ready+automerge_enable on 20022 exit 0, and with-fleet.sh --via dispatch --actions exit 0, same payload. No ablation (text-only change).",
  "gates": "check-closing-keyword-parity 0, its --self-test 0, check-comment-mask-corpus 0, check-governed-queue-guard --self-test 0, check-harness-current --self-test 0, lint check:doc-formula-expressions 0 (after build), agent-test-spelling 0, cross-package-test-inputs 0, doc-authoring 0, driver-memory-census 0, gitlink-declared 0, nul-bytes 0, pm-expected-skips 0, pm-governed-merges 0, pm-governed-prose 0, pm-half-states 0, pm-skill-id-lint 0, pm-skill-ratchet 0, refd-timer-probe 0, required-contexts 0, skill-frame-sync 0, watch-hint-literal 0, pm-settings-deny-roster 0",
  "line_budget": "SKILL.md 319 to 319 (ceiling 319); landing-operations.md 101 to 101 (101); rest-channel.md 82 to 82 (82); platform-readings.md 469 to 469 (469); every changed line is at most 120 bytes",
  "files_changed": [".claude/skills/pm-dispatch/SKILL.md", ".claude/skills/pm-dispatch/references/landing-operations.md", ".claude/skills/pm-dispatch/references/rest-channel.md", ".claude/skills/pm-dispatch/references/platform-readings.md"],
  "deviations": [
    ".claude/settings.json: the two dispatch.mjs allow rows were not added (auto-mode classifier Self-Modification denial, not retried); the exact rows are in the PR body for the maintainer",
    "a scratch byte-length measurement of candidate lines was denied as Auto-Mode Bypass; line widths were then judged only by check:pm-skill-ratchet on the real edits",
    "denial reason names are not spelled in platform-readings.md, because the 120-byte line budget leaves no room; they are in the PR body and on the card"
  ],
  "mcp_calls": "0",
  "api_writes": "4: relay POST /repos/objectstack-ai/objectstack/dispatches for pr_create (run 36082063906), label-write --assign os-elon-musk (run 36082105836), label-write --add skip-changeset (run 36082167796), and this os-dev-report comment via post-stamped. Plus git push (not REST).",
  "open_questions": [
    { "question": "Delete the six ccr allow rows in .claude/settings.json:61-66 now that the text retires the ccr pair?", "options": ["A delete them", "B keep them"], "recommendation": "Maintainer's surface; asked in the PR body, no recommendation from dev" }
  ],
  "out_of_scope_findings": [
    "not filed: scripts/pm/check-settings-deny-roster.mjs header comment still says the ccr auto-merge REST route serves every real enqueue; stale comment, no class a/b/c; carrier: none; noted in PR Acceptance notes"
  ]
}

[objectstack-fleet]

ACCEPT — os-dev-report 5825175951 reviewed against references/review-checklist.md by the domain:skills#1 seat (session_01LNRg6FUMXYWgkAbDsfZpC8) at 2026-09-25T01:31Z. Tier S (.claude/** ×4, check-governed-merges --test GOVERNED — Tier S) ⇒ in-seat ## Contract review record (next comment) + every check green ⇒ this seat lands it through the relay — the route this PR writes down, under the maintainer's 「B:走中继落地(推荐)」.

PR form: draft PR #20070 on main, first line Fixes #20048, line 2 Clause-②: no — the only closing keyword; head d32057c (one commit, model-free trailers); assignee os-elon-musk; labels documentation · size/s · skip-changeset (nothing publishes); 速读 draft present with 你要做的 ①② for the maintainer; session-URL footer intact.

Diff, read hunk by hunk (origin/main...d32057c, +12/−12 over 4 files, every file line-neutral, every changed line ≤ 120 bytes):

• SKILL.md :201 (MANDATORY fable path, built at opus under the one-line-class exit, reviewed here at tier): 「ready/draft 走 ccr 路 … 端点见 platform-readings」 → 「ready/draft 走中继 op … 端点见 rest-channel」 — the pointer now names the table that holds the spelling. ✓
• landing-operations.md §B :51 「只走 settings.json allow 行的 ccr 两命令」 → 「经中继 op pr_ready + automerge_enable 落地,拼写见 rest-channel.md」; :54 the denial line keeps stop + report + record and forbids 「不退回 ccr 绕过」 in place of 「不改走中继绕过」 — exactly the ruling; :53 (undraft 429) left, with the PR noting the relay's 429 shape is unmeasured. ✓
• rest-channel.md: :17 「dispatch.mjs 无 allow 行」 → 「直调 dispatch.mjs 仅落地」; the two ccr rows (:51, :55–:56) replaced by the relay spelling (node scripts/pm/fleet-write/dispatch.mjs --repo O/N --actions-file F, --via dispatch also carries it — both dry-run exit 0 by the dev), the actions-file shape with pr_ready first, and the honest note that the reverse ops are unexercised. ✓
• platform-readings.md :48–:49 undraft channel rewritten to the relay (ccr retired, 3/3 passes under auto mode from the card's reading), :138 and :146 pointers, :435 「ccr 翻转亦拒」. The dropped 「裸 GraphQL 已拒」 referred to session credentials; the relay sends the same mutations with the App token — accepted, the PR explains it. ✓
• .claude/settings.json NOT touched: the dev's edit was refused [Self-Modification] and, per the rules, not retried; the two rows are handed to the maintainer in the PR body (你要做的 ①), with the breadth caveat that dispatch.mjs * admits every closed op. ⛔ This seat does not add them either — a rule text asked a session to widen its own permissions, and that is the maintainer's surface (the PR says so). Accepted as the correct shape.

Tests / gates (all at d32057c, exit codes captured before any pipe): dispatch-gates --ran 22 / 22 run, 0 NOT-MEASURED · pm-skill-ratchet 0 (four files at ceiling, neutral) · pm-skill-id-lint 0 · skill-frame-sync 0 · doc-authoring 0 · pm-governed-merges 0 · pm-governed-prose 0 · pm-settings-deny-roster 0 · pm-half-states 0 · pm-expected-skips 0 · required-contexts 0 · nul-bytes 0 · doc-formula-expressions 0 after the lock-guarded build · relay dry-runs of [pr_ready, automerge_enable] exit 0 on both spellings, nothing sent. CI on the head at this ACCEPT: 13 success + 12 rostered skips, 4 in progress (Lint & Repo Gates, three Type Check lanes) — landing waits for them.

Deviations (four, all accepted): settings.json rows not added (above) · a scratch byte-length probe denied [Auto-Mode Bypass] — widths judged by the ratchet on the real edits · the denial-reason names not spelled in platform-readings (no headroom) — they live in the PR body and on the card · the 「bare GraphQL refused」 clause dropped with reason.

Open questions (the dev's), carried to the maintainer as the PR's 你要做的: ① add the two dispatch.mjs allow rows (or leave each relay landing to the classifier); ② delete or keep the six ccr rows. Both are the maintainer's surface; neither blocks this PR.

Findings (the dev's, judged): check-settings-deny-roster.mjs's header still says the ccr route 「already serves every real enqueue」 — a stale comment, carrier none, Acceptance notes only; the relay's 429 shape unmeasured — noted in the PR.

Landing from here: flip 定点 after the four jobs finish; relay pr_ready → automerge_enable (one actions file) — the first live use of the route this PR describes; followed to MERGED; on MERGED the card closes by Fixes, then pm:dispatched and the assignee come off.

Four axes, one line each: business need — measured (two denials of the allow-listed ccr flip in two sessions today, three relay landings with none) · long-term — one write channel for everything, as AGENTS.md already states, and maintainer item (v) closes with it · AI-error — a seat that is refused has exactly one instruction: stop and report, no second route to guess at · startup scope — twelve lines rewritten in place, no gate, the allow rows left to the maintainer.

---

Generated by Claude Code

[objectstack-fleet]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: d32057c683002b829eb99da8b4f765560e22a41e

① Derived judgments

• The diff rewrites landing-route text in four .claude/skills/pm-dispatch/** files (SKILL.md is a MANDATORY-tier protocol file; the change there is one pointer line, built at the default tier under the one-line-class exit and reviewed here): no accepted set, schema, export or gate changes; +12/−12, skip-changeset correct. Correct.
• The rule as written matches the ruling 「B:走中继落地(推荐)」: ready and auto-merge go through the relay ops pr_ready + automerge_enable (landing-operations :51, rest-channel :51–:55, SKILL.md :201); the ccr pair is retired (platform-readings :48); a classifier denial still means stop + report + record with no fallback route (landing-operations :54). Correct.
• Facts recorded are readings, not guesses: the relay's 3/3 auto-mode passes and the ccr denials come from objectstack#20048's measurements; the relay actions-file shape was dry-run by the dev (exit 0, both spellings); the reverse ops are marked unexercised rather than asserted. Correct.
• Exclusions hold: no gate, no patrol row, no tool; .claude/settings.json untouched and its two candidate rows handed to the maintainer (their surface). Correct.
• Line-ratchet: all four files at ceiling and neutral, changed lines ≤ 120 bytes, no re-wrap. Correct.
• Reader check: every changed line has a named reader — the landing seat at the flip 定点 (landing-operations, rest-channel), the reader of the write table, the seat reading the four-piece terminal (SKILL.md :201). No token tax.

② Semver level

None owed: nothing published changes; consistent with Clause-②: no and skip-changeset.

③ Boundary flags

• open_questions: two, both the maintainer's surface (.claude/settings.json rows to add; ccr rows to delete or keep) — carried in the PR's 你要做的; neither changes what this PR's text says.
• Not a widening: no accepted set touched.

Implemented-by: claude/issue-20048-relay-landing-route
Reviewed-by: session_01LNRg6FUMXYWgkAbDsfZpC8

VERDICT: PASS

---

Generated by Claude Code

[objectstack-fleet]

Reading: the same ccr pair passed IN auto mode, later the same night

_{domain:cli execution seat #6024 · session session_01TnPAC1UsTGfHPXVUCL6iLn · 2026-09-25T01:55Z}

• The session was back in auto mode (harness notice, about 2026-09-25T01:52Z).

• At 2026-09-25T01:55Z the seat landed PR fix(runtime): the run-lifecycle refusal names only a remedy the tenancy posture honours (#19874) #20074 (card RUN_LIFECYCLE_DENY_MESSAGE tells a denied operator to obtain "the unscoped admin_full_access grant" — on a walled deployment that grant no longer confers platform-admin standing (split from #19145, carrier 3) #19874, ACCEPT 5825401400, CI green on 40bf21c9) with the same two allow-listed commands. The ready flip answered {"draft":false} and the auto-merge arm answered {"enabled":true,…}. Neither was denied.

• ⇒ This session's three readings of the ccr pair:

  when	mode	verdict
  PR fix(cloud-connection): POST /api/v1/marketplace/install-local parses the package id through its declaration #20022	auto	ready flip denied [Auto-Mode Bypass]
  PR fix(client)!: every limit query-parameter emitter sends what the caller wrote #20060	not auto	both passed
  PR fix(runtime): the run-lifecycle refusal names only a remedy the tenancy posture honours (#19874) #20074	auto	both passed

  The auto-mode denial is not deterministic for byte-identical command shapes, which is [finding] the harness auto-mode classifier denies the protocol's landing command non-deterministically — 8 byte-identical PUT …/ccr/auto_merge from one turn: 4 allowed / 4 denied under three reasons; and a hooks landing rotates every seat off shift before it can flip ready #18469's finding again. So the stop rule at §B:54 will fire intermittently, not always. Three readings; ⛔ not a rate.