fix(instruction-placement): align /memory claims, consolidate cutover records, neutralize owner-decision records - #5285

PR body contract — issue linkage

This PR body conforms to the issue-linkage contract. Nothing to do.

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Claude finished @kyle-sexton's task in 2m 21s —— View job

Code review — PR #5285

• Gather context
• Invoke /review:code-review per custom instructions
• Post inline findings (none — see below)
• Summarize in this comment

Scope reviewed: all 13 changed files (git diff origin/main...HEAD), per the code-review lane's correctness/maintainability/CLAUDE.md-alignment criteria.

Findings: none that clear the high-signal bar. No inline comments posted.

What I checked in detail:

• hooks/index-drift.test.sh: the new idx_body_case helper and its two call sites correctly exercise the hook body past the hot-path guard (payload targets .claude/rules/csharp.md, which matches the guard's */.claude/rules/*.md case). It uses file-fed stdin (<"$dir/body.in") rather than a pipe, consistent with the file's own Shell test suites pipe a payload into hooks that exit before reading stdin, so pipefail fails them intermittently #4458 note about pipe/SIGPIPE issues elsewhere in the suite. $repo's "in-sync" state at the point of the no-stale call is genuinely in sync (rendered earlier in the file via render-index.sh write), so the assertion is sound, not accidentally passing.
• hooks/hooks.json: confirmed the new README/setup claim that every hook row launches via node hooks/exec-bash.mjs — there's exactly one row, and it does.
• skills/migrate/reference/sources.md: cross-checked the new "Current fleet grade" claim (condition 2, medley's v1.0.231 → CLI 2.1.278, "at or above 2.1.277") against the unmodified release-map table above it — consistent. The consolidated "Canary host (CI canary: verify a lone AGENTS.md loads on a fresh-install claude-code-action run #4282)" and "Install-dependent loader tests (Install-dependent loader tests for Cursor, Grok Build, and Muse Code #4283)" sections read as genuinely pending-decision framing (no more "Option A taken"), matching the PR description.
• skills/{audit,check,delta}/SKILL.md: the new ## Next sections comply with .claude/rules/skill-bodies-state-current-rules.md — placed immediately before ## Gotchas, exactly one /instruction-placement:realign invocation line plus an explanatory sentence.
• CHANGELOG.md / plugin.json: version bump (0.15.21 → 0.15.22) and changelog entries match the diff's actual content.
• Dated four-part records touched by this PR (README, check, setup, migrate, verified-mechanics.md) now point at sources.md's "The minimum CLI version" instead of restating the CLI-floor specifics inline — reduces duplication of a volatile fact rather than adding drift risk.

Not independently verified: the auto-mode permission classifier for this lane declined to run shell/python commands beyond read-only git (bash scripts/check-changelog-parity.sh …, python3 -c "json.load(...)", etc. were all denied). I did not re-run check-changelog-parity.sh, validate-plugins.sh, check-skill-portability.sh, the plugin's own test suites, or cutover-check.sh. The PR body's Verification section describing those runs is author-claimed and unverified by me.

Claude finished @kyle-sexton's task in 2m 19s —— View job

Security review complete

• Gather context (PR diff, changed files)
• Check skip-gate conditions
• Review diff for security-relevant surface, incl. instruction-surface deletions
• Post findings

No security issues found.

Scope reviewed: all 13 changed files (git diff origin/main...HEAD). This PR is documentation/test-only: /memory claim corrections across migrate, check, setup, verified-mechanics.md, README.md, render-index.sh, a node prerequisite check added to setup, ## Next pointers, a test that now exercises the index-drift hook body instead of exiting at the hot-path guard, and a large prose consolidation in skills/migrate/reference/sources.md. No script logic, hook logic, workflow files, or permission/config surfaces changed.

Instruction-surface deletion lens (specifically checked): sources.md drops the Canary host decision (#4282) and Install-dependent loader tests parked (#4283) sections' "Option A (taken) / Option B (declined)" framing — including the explicit "Do not unarchive claude-lane-sandbox" / "Do not create a throwaway host" / "Do not install those tools" bullets — in favor of "pending an owner decision on #4282/#4283" language. Two reasons this isn't a finding:

• The candidate actions this record used to name (touching another repository, standing up a fleet host) are already gated by this repo's root AGENTS.md "When to stop and when to keep going" stop-and-ask list, which any session in this tree loads independently of sources.md, plus Claude Code's baseline confirm-before-CI/infra-change behavior. The mechanism that actually blocks the action survives; only a redundant restatement was removed.
• The removed framing was itself the finding the PR is fixing, not a guardrail worth keeping: it was an agent's own prior "Option A taken" decision written into a reference doc as if authoritative on an owner call, and the PR body says so directly (#4282 and #4283 are owner decisions the agent had taken on its own). Replacing it with explicit "pending an owner decision" text is a correction toward the agent-authority class, not away from it.

Parsed-record integrity: cutover-check.sh (unchanged in this diff) only parses three sources.md sections — ## The minimum CLI version, ## The CI canary, and ## `claude-code-action` release to installed CLI version — via fixed awk/grep section extraction. None of those headers or their parsed fields (version floor, canary run id, release-map table) were restructured here; only fetch dates/byte counts were refreshed, consistent with a real re-fetch. The consolidated/removed sections (Accessible fleet graded, Standing refresh (#5163), This-repo cutover run (#4281)) are narrative history the script never reads, so the remove-shims gate's actual enforcement is unaffected.

GitHub Actions hardening (triggers, permissions:, pin provenance) is out of scope for this lane and none of the changed files are workflow files regardless.

Claude finished @kyle-sexton's task in 1m 56s —— View job

Task list

• Gather context (PR diff, changed files, custom instructions)
• Invoke /review:security-review per custom instructions and follow it exactly
• Post findings

Security review complete

No security issues found.

Scope reviewed: the full PR diff, git diff origin/main...HEAD (13 files), with particular attention to what changed since the prior two security-review runs on this PR — the branch was re-synced with origin/main and one new commit landed, 46b8896, which splits the delta skill's ## Next section into two outcome-dependent bullets (findings-moved → realign; quiet-run → none) in response to Codex's P2 comment.

New commit checked under the instruction-surface lens: this change adds routing prose only — it does not remove, narrow, or soften any standing rule, permission, or guardrail. The ## Hard rules section immediately above (Never resurrect a declined finding, Never suppress silently, Never re-classify an unchanged finding, etc.) is untouched. No finding.

Confirmed unchanged from the prior reviews on this PR (still holds against the current head):

• sources.md's consolidation of the #4282/#4283 "Option A taken / Option B declined" framing into "pending an owner decision" text removes a redundant restatement, not the controlling guardrail — the actions it used to name (unarchiving a sandbox repo, standing up a fleet host, installing untested tools) are still gated by this repo's root AGENTS.md stop-and-ask list, which loads independently of sources.md.
• cutover-check.sh (unchanged in this diff) only parses three fixed sources.md sections via awk/grep; none of those headers or their parsed fields were restructured, so the remove-shims gate's enforcement is unaffected by the narrative consolidation.
• hooks/index-drift.test.sh's new idx_body_case cases build hook stdin from fixed, non-attacker-controlled paths (.claude/rules/csharp.md) — no injection surface.
• No script logic, hook logic, workflow files, or permission/config surfaces changed anywhere in the diff; the remainder is documentation and a version/changelog bump.

GitHub Actions hardening is out of scope for this lane and no workflow files are part of this diff.

Claude finished @kyle-sexton's task in 3m 53s —— View job

Code review — PR #5285

• Gather context
• Invoke /review:code-review per custom instructions
• Post inline findings (none — see below)
• Summarize in this comment

Scope reviewed: full diff at HEAD 46b8896 (gh pr diff 5285, all 13 changed files), per the code-review lane's correctness/maintainability criteria. This head adds one commit past the last review (bdc59f3): 46b8896, touching only skills/delta/SKILL.md.

Findings: none that clear the high-signal bar. No inline comments posted.

What changed since the last review, and what I checked:

• skills/delta/SKILL.md — this commit resolves Codex's P2 finding (unconditional ## Next pointing at the mutating realign skill even on a quiet run with nothing to apply). The ## Next section now has two outcome bullets: "Findings moved" → /instruction-placement:realign, "Quiet run" → none. This matches .claude/rules/skill-bodies-state-current-rules.md's "two to four bullets of <outcome>: /plugin:skill [args]." form for outcome-dependent successors, and is consistent with the file's own "a quiet run is the expected outcome, not a failed one" gotcha a few lines above.
• Re-checked hooks/index-drift.test.sh's new idx_body_case cases against hooks/index-drift.sh's hot-path guard (*/.claude/rules/*.md) — both fixture paths correctly reach the hook body rather than short-circuiting, so the "in-sync → empty stdout" and "drifted → stale notice" assertions are exercising what they claim to.
• Checked for dangling references to the four sources.md sections this PR collapses (Canary host decision, Accessible fleet graded, Standing refresh, This-repo cutover run, Install-dependent loader tests parked) — only one hit, an unchanged historical CHANGELOG.md entry under [0.15.21] (Standing refresh — not a live link, no action needed).
• Verified the new ## Canary host (#4282) and ## Current fleet grade anchors (#canary-host-4282, #the-ci-canary) resolve correctly under GitHub's heading-to-anchor rules.
• hooks/hooks.json still backs the new README/setup claim that every hook row launches via node hooks/exec-bash.mjs.
• skills/setup/SKILL.md's renumbered prerequisite list (1–5, node inserted as 3) is internally consistent; no other file references the old numbering.

I considered but did not raise two points, on the reasoning below — flagging here in case it's useful:

• Codex's P1 finding (sources.md:154, condition 2 graded [MET] while the CI-canary is stale/pending) is about cutover-check.sh's condition_2() logic, which this PR does not touch — it only rewords the record from "Option A taken" (settled) to "pending an owner decision" (the PR's stated purpose). The doc now accurately describes existing, unchanged script behavior; the gap between "pin arithmetic says MET" and "canary is stale" predates this PR.
• delta/SKILL.md's new "Quiet run: none." bullet doesn't literally fit the convention's <outcome>: /plugin:skill [args]. grammar (no skill invocation) — it's the first "no successor" outcome bullet in the repo (checked; no other ## Next has this shape). Reads as a reasonable, unambiguous extension of the convention's intent rather than a defect worth blocking on.

Not independently verified: I did not re-run scripts/check-changelog-parity.sh, scripts/validate-plugins.sh, scripts/check-skill-portability.sh, the plugin's test suites, or cutover-check.sh — the sandbox's command approval declined python3/scoped git show/git status calls during this pass. The PR body's Verification section is author-claimed and unverified by me.
· branch fix/audit-instruction-placement

Claude finished @kyle-sexton's task in 2m 35s —— View job

Task list

• Gather context (PR diff at HEAD, custom instructions)
• Invoke /review:security-review per custom instructions and follow it exactly
• Post findings

Security review complete

No security issues found.

Scope reviewed: full diff at HEAD b9b722d (gh pr diff 5285, all 14 changed files). Three commits landed since the last security-review pass on this PR (at 46b8896): 112545b (records the real #4282 canary run in sources.md), 275b444 (merge of origin/main, no net content beyond what's already reflected in the PR diff), and b9b722d (empty, CI-retrigger only, git show --stat confirms zero files changed).

New commit checked under the instruction-surface lens: 112545b rewrites sources.md's ## The CI canary, ## Canary host (#4282), and ## Current fleet grade records to name a real knowledge-corpus run (36666844023, claude-code-action v1.0.235, CLI 2.1.283) in place of the previous "pending owner decision" text. This is the resolution of Codex's P1 finding on this PR, not a new instance of it: that finding warned that grading cutover condition 2 [MET] off a stale run (35475056935 at the older v1.0.231 pin) operationally decided the pending #4282 question by default. The new record replaces the stale evidence with a run genuinely taken at the current pin, so the gap the finding described no longer exists.

Parsing verified independently (not just read as prose): cutover-check.sh's section() helper (plugins/instruction-placement/skills/migrate/scripts/cutover-check.sh:205-207) bounds ## The CI canary at the next ## heading via awk, and CI_CANARY_RUN (line 222-223) takes the first run `[0-9]{6,}` match with head -n 1. Within the rewritten section, run `36666844023` (the new run) appears before the later The earlier run `35475056935` reference, so the script correctly parses the new run id — matching cutover-check.test.sh's updated assertion. No ambiguity or wrong-value risk from having two run ids in the same section.

Confirmed unchanged from the prior three reviews on this PR (still holds against the current head):

• sources.md's consolidation of the #4282/#4283 "Option A taken / Option B declined" framing removes a redundant restatement, not a controlling guardrail — the actions it used to name (unarchiving a sandbox repo, standing up a fleet host, installing untested tools) are still gated by this repo's root AGENTS.md stop-and-ask list, independent of sources.md.
• cutover-check.sh itself is unchanged in this diff; only the three fixed sections it parses (## The minimum CLI version, ## The CI canary, the release-map table) had their content refreshed, never restructured.
• hooks/index-drift.test.sh's new idx_body_case cases use fixed, non-attacker-controlled paths — no injection surface.
• No script logic (besides the parsed-data refresh above), hook logic, workflow files, or permission/config surfaces changed anywhere in the diff.

GitHub Actions hardening is out of scope for this lane and no workflow files are part of this diff.
· branch fix/audit-instruction-placement

Claude finished @kyle-sexton's task in 2m 16s —— View job

Code review — PR #5285

• Gather context
• Invoke /review:code-review per custom instructions
• Post inline findings (none — see below)
• Summarize in this comment

Scope reviewed: the full diff at HEAD b9b722d (gh pr diff 5285, all 14 changed files vs origin/main), with focus on what changed since the last full review on this lane (commit 46b8896, 2026-09-29T14:41:56Z):

• b9b722d — empty commit ("retrigger checks after stalled PR synchronize"), no content change.
• 275b444 — merge of origin/main; mechanical version rebase (0.16.0 → 0.16.1 in plugin.json), no logic touched.
• 112545b — the substantive new commit: records the CI canary: verify a lone AGENTS.md loads on a fresh-install claude-code-action run #4282 CI canary re-run (knowledge-corpus run 36666844023, claude-code-action v1.0.235, CLI 2.1.283) in sources.md's "The CI canary" and "Canary host (CI canary: verify a lone AGENTS.md loads on a fresh-install claude-code-action run #4282)" sections, updates CHANGELOG.md, and bumps cutover-check.test.sh's expected run id from 35475056935 to 36666844023.

Findings: none that clear the high-signal bar. No inline comments posted.

What I checked on 112545b, since it resolves Codex's earlier P1 (sources.md:154, "block removal while the CI-canary decision is pending"):

• cutover-check.sh's section() helper (skills/migrate/scripts/cutover-check.sh:205-206) extracts text strictly between the ## The CI canary heading and the next ## heading, so it never reads into ## Canary host (#4282). Within that section, CI_CANARY_RUN_ID (cutover-check.sh:223) takes the first run `NNNNNN` match via head -n 1. In the new "Basis" paragraph, the first such match is run `36666844023` (the new run); the older earlier run `35475056935` appears later in the same paragraph and is correctly not picked up. This matches the doc's own updated claim ("it parses the first run id in this section").
• Since the new canary run is recorded at v1.0.235 (the pin condition_2 actually resolves against the in-scope fleet), the record no longer describes a stale canary sitting behind a newer pin — the specific gap Codex's P1 flagged is closed by this commit, not routed around it.
• sources.md's "Canary host (CI canary: verify a lone AGENTS.md loads on a fresh-install claude-code-action run #4282)" section now states the owner's decision and the run outcome as settled fact ("the owner chose... and the run passed"), consistent with Kyle's own resolving comment on the thread, and no longer frames this as a still-open agent-side option A/B choice.
• The cutover-check.test.sh diff is a single literal swap (old run id → new run id) against the test's own synthetic fixture data; it doesn't touch the CLI-version assertion on the same block, which is unrelated (it's the test's mocked pin, not tied to the real fleet's CLI floor).

Not independently verified: this lane's auto-mode permission classifier declined bash scripts/check-changelog-parity.sh --check --check-order and other non-git shell invocations during this pass, consistent with the two prior reviews on this PR. I did not re-run the changelog-parity check, validate-plugins.sh, check-skill-portability.sh, or the plugin's test suites. The PR body's Verification section (including the canary run itself, 36666844023) is author-claimed and unverified by me.

Branch: fix/audit-instruction-placement