Register remind-ums-on-scrutiny.py after #2262 - #2265
Conversation
README's activation gate forbade binding the hook in its authoring PR. #2262 left it in KNOWN_UNREGISTERED. This follow-up adds the UserPromptSubmit inject entry, drops the allowlist row, and updates the catalog and fragment so they describe a live guard. Closes #2263. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
This comment has been minimized.
This comment has been minimized.
Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
Summary of ChangesFallback self-review after claude-review skipped (quota/credential) on earlier HEADs of this PR. Inspected This PR registers Local verification: FindingsNo actionable findings identified. Catalog, README row, and Verdict: Ready for mergeReviewed-Commit: 49d818d Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
Summary of ChangesThree commits on
Verification I ran in the worktree: Findings
Non-diff note, per the assignment-premise check: the brief names Verdict: Needs more workReviewed-Commit: 49d818d Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
Reword the scrutiny-hook note to name the sibling scripts and to say admission overlap is deliberate. State that this hook is bound in hooks.json. Refresh the remaining allowlist comment. Note that --fix binds only the manifest. Refs #2263. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
Drop the fragment binding-status sentence. Trim the duplicated allowlist header. Break the README allowlist sentence at the clause. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
|
Addressed the Claude Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
Use 2026-08-25 PDT for every measurement this branch records, matching the wraps on #2265/#2266. Retire the leftover "only cross-vendor reviewer" sentences in the retired Antigravity skills. Name opencode as outside the Copilot/codex pairing, with its OpenRouter caveat. Drop the forward "below" pointer. Justify passing Claude as vendor independence from the author, not as the intent-independence floor. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
…crutiny-hook-32a3
This comment has been minimized.
This comment has been minimized.
…crutiny-hook-32a3
This comment has been minimized.
This comment has been minimized.
|
#2294 merged to The collision is This session will not push to Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
|
ARDI round dispositions for this round (adversarial subagent reviews of f327801, tree carried to ecd7602 by main-sync only): Deferred: fetch-pattern over-fire (a Grep/Read whose arguments name a comment-fetch counts as a review-read) - declared deliberate in the hook docstring, predates this PR (#2262); sentinel-file accumulation in the shared tempdir - bounded per session, same shape as the sibling reminder hooks. Resolved: branch staleness behind main - synced; the paired hook suite was re-run green (45/45) after the sync, and check-hook-catalog stays consistent (42 registered / 43 documented / 1 remaining allowlist entry, The dispatched review verdict on f327801 was Ready for merge; no PR-owned file changed since. Merge remains blocked on a clean automated claude-review verdict for the current head (runs currently 429 on the account weekly limit); human review requested. Posted by opencode (AI agent) --- not written by a human. |
The allowlist mapped unregistered hooks to a tracker issue but never checked that the issue was still open. That is how #1717 stayed allowlisted after it closed. Fail on a closed mapped issue so a closed activation tracker cannot keep a hook silently inert. When the issue cannot be fetched, print SKIP and do not fail -- that is the documented offline path, not a silent pass. Tests inject states via HOOK_CATALOG_ISSUE_STATES so they never hit the network. Retarget remind-ums-on-scrutiny.py from closed #2261 to open #2265 (the registration follow-up) so the live catalog stays green without registering the hook in this PR. Closes #2302. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
…crutiny-hook-32a3 # Conflicts: # scripts/check-hook-catalog.py
This comment has been minimized.
This comment has been minimized.
The injected `closed` token never parses JSON, and production #2265 is open, so a regression that mapped a 200 `closed` body to None would keep the suite green and SKIP after that tracker closed. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
|
Cursor Grok 4.6 is working on this --- resolving the Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
The merge of origin/main into this branch conflicted on the allowlist dict: this PR drops remind-ums-on-scrutiny.py, and #2294 already dropped no-incomplete-check-enumeration.py. The merge kept the #1717 row, which is now registered on main, so check-hook-catalog.py failed. Restore main's keyed-per-entry comment and leave the dict empty. Keep this PR's hooks.json registration of remind-ums-on-scrutiny.py. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
|
fix validation |
|
Fallback adversarial self-review after Claude review skipped (quota/credential) on this PR. Inspected Summary of ChangesThis branch registers The merge of FindingsNo actionable findings identified. Catalog, event, and matcher now agree: Verdict: Ready for mergeReviewed-Commit: 5cadd82 Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
|
Adversarial review of Summary of ChangesThis branch registers FindingsNo actionable findings identified. Checks that survived inspection:
Verdict: Ready for mergeReviewed-Commit: 5cadd82 Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
|
ARD for the adversarial review of 5cadd82: no findings. Nothing to Address, Rebut, or Defer. Validate on that HEAD is pass (Ezra's "fix validation": the leftover Posted by Cursor Grok 4.6 (AI agent) --- not written by a human. |
…2272) * start: Cursor Task Claude adversarial-reviewer routing (closes #2270) Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * cai: Cursor Task dispatches adversarial-reviewer; prefer Claude A skipped GitHub claude-review is a different channel from Cursor Task. When the conductor is not Claude, pass a listed Claude model so the review is cross-vendor on blind spot as well as independent of intent. Closes #2270. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Address Claude review of Cursor Task routing Hedge that Cursor does not filter the persona tools list. Name the Task model parameter. Cite self-review-fallback for the cross-vendor half. Scope the dispatch to a repo that ships the persona. Move the fragment pointer below the CLI pre-push-guard paragraph. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Address remaining review findings on Cursor Task routing Qualify the self-review-fallback exclusivity claim to the CLI axis so a non-Claude Cursor conductor is not told that codex is the only self-dispatchable cross-vendor reviewer. Name foreground dispatch, hedge that Cursor Task still grants Write schemas, and drop the restated pass-model paragraph plus the forward pointer at the tool_result section. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Hedge Cursor Task routing claims the re-review falsified Drop the omit-model vendor assertion the same section called unmeasured, stop citing #1921 for a which-path question that issue does not track, and stop calling codex the only self-dispatchable CLI: opencode is another. Name the 2026-08-25 dispatch instead of "this session", and state that this conductor's Task schema listed run_in_background and did not list isolation. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Separate self-review dispatch from a cross-vendor second reviewer A Claude Task child of a Grok conductor is independent of the author, not of a GitHub claude-review primary, so it does not belong in the Copilot/codex pairing. State that Cursor's adapter skips the pre-push guard, name the Task-to-Agent mapping that trips the worktree warning, and keep the dispatch recipe in memories/cursor.md rather than restating it in the always-loaded fragments. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Check HEAD around the Cursor Task child, not git status git status is clean over unpushed commits, so it cannot show a child that committed. Record HEAD before the dispatch and compare it after. Qualify AGENTS.md: Cursor's adapter skips the pre-push guard. Name opencode as a CLI whose skill excludes this work, not as "only"'s exception. File #2276 for the unmarkable isolation warning. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Date Cursor Task measurements in Pacific and sweep sibling "only" claims Use 2026-08-25 PDT for every measurement this branch records, matching the wraps on #2265/#2266. Retire the leftover "only cross-vendor reviewer" sentences in the retired Antigravity skills. Name opencode as outside the Copilot/codex pairing, with its OpenRouter caveat. Drop the forward "below" pointer. Justify passing Claude as vendor independence from the author, not as the intent-independence floor. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Retrieve Reviewed-Commit from the child report, not the Task JSON The Task tool_result is identity-only, so a HEAD comparison that names Reviewed-Commit has to say how to get that line: a harness paste of the report, or batch-fetch-details. Compare the push refspec when it is not HEAD. Hedge the Desktop guard skip: only the project adapter path omits it. Drop the tools: causal claim the two persona copies cannot support. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Fail closed when Reviewed-Commit cannot be retrieved The posted PR comment is a record, not a gate. If the harness paste and batch-fetch-details both lack Reviewed-Commit, do not push. Hedge the AGENTS.md Task claim on the conductor listing Task. Keep git status as the dirty-tree check beside the HEAD comparison. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Require Ready for merge and includeTranscripts on the fetch A Needs more work report whose fingerprint matches HEAD is not a push. batch-fetch-details needs includeTranscripts true or it writes no transcript. File #2281 for the persona Write-schema leak. Unify the Task-lists-adversarial-reviewer precondition. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Scope the skipped pre-push guard to Cursor Cloud, not any Task tool A Claude Code session whose subagent tool is Task still has the guard. Name Cursor Cloud in the sentences. Enabling Desktop third-party Claude hooks beside the project adapter runs the native guard and denies every push. Drop "and nothing else" from the Copilot/codex pairing. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Fail closed on a Cursor Task miss and a HEAD mismatch The Desktop-lockout sentence was an unverified claim: native Claude hooks key on Bash, and Cursor's Shell path is not that adapter. Strip it from AGENTS.md. Commit before dispatching, parse the last line-start Verdict then the first Reviewed-Commit after it, and treat a child error as the CLI-fallback case rather than as a forbidden override. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Fail closed on an unclosed fence and name the hand comparison A truncated report with an open fence is no verdict to the guard and was a matching Ready under the hand recipe. State that rule, drop the "instead" in AGENTS.md, and stop forbidding the override for a reason the fragment already rejects. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * State nested-fence close rules and drop the incomplete refuse list A nested fence of different lengths is the shape the guard's scanner exists to catch. Pairing by count reads Needs more work as Ready. Name the close rule, and treat any push that is not a single named ref as uncovered rather than enumerating a subset of the guard. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Resolve shipped refs with dry-run and scope the skip to Cloud A matching HEAD sha does not cover tags or extra branches a config flag would add. git push --dry-run is the resolution step the guard exempts from review. Scope the adapter skip to Cursor Cloud so "active hook path" cannot be read as covering a paired Desktop native path. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Fail closed on a dry-run miss and drop the Cloud-only skip The adapter skip is a property of the adapter, not of Cloud. A dry-run must use the same argv as the push, and an empty or failed dry-run is not coverage. On Cursor Cloud the override prefix is inert even after a Task error. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Blank the fingerprint search and compare the dry-run source ref A fenced example sha that names HEAD is the hole the guard already measured. Search verdict and fingerprint on the same blanked text. A new-branch dry-run has no sha; the source ref, left of the arrow, is what ships. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Restore the override list and record the branch name The Cursor Cloud inert-prefix sentence had stolen the override case list. Put it after the list. Record the branch name so a new-branch dry-run whose source is HEAD is covered by the sha, and require a paste to be the child's own message. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Make the transcript route the push gate A paste of an author-composed Summary/Findings/Verdict block is the wrap this file already records. batch-fetch-details discharges the check. Prefix-match the fingerprint, skip a trailing Approved, and treat Everything up-to-date as shipping nothing rather than as a mismatch. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Break the Cursor Cloud override sentence after the semicolon PR-event new-line-breaks failed on a 81-character added line in skills/push/SKILL.md that packed two clauses around a mid-line semicolon. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Align the Do-bullet dry-run clause with the no-sha carve-outs The summary bullet refused any dry-run that did not print the fingerprint sha, which is the first push of a new branch and an Everything up-to-date retry. Name the hand comparison the transcript route discharges, and qualify the mechanism section as Claude Code. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: match Cursor Task review recipe to VERDICT_LINE and adapter scope Address the Claude Opus Task child of ea38ed0: allow an optional ATX prefix on the hand-parsed Verdict line, treat the transcript as the source of the comparison rather than a discharge of it, scope the inert ALLOW_UNREVIEWED_PUSH prefix to wherever the adapter skips the guard, and qualify the pre-push claim on CLAUDE.md and both persona copies. Defer a parse_report CLI wrapper to #2299. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Address adversarial-review findings on the Cursor Task route # Conflicts: # memories/cursor.md # skills/push/SKILL.md * fix: pin Verdict emphasis position and name the adapter in fragments Address the Claude Opus Task child of 1004f34: optional ** on a Verdict line is only between the colon and the value, an unrecognized final verdict fails closed, the transcluded fragment names Morrison-Lab/ai-config, the push-skill pronoun names the guard, and the Copilot contrast no longer repeats the billing blurb. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: restore two-space indent on the Task-route Do-bullet The parallel Address commit left four continuation lines at three spaces, which nested them under the bullet rather than continuing it. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * Widen the no-override scope to Cursor desktop as well as Cloud * fix: qualify the Cursor prefix-inert claim as Cloud-only Address the Claude Opus Task child of 2c0ed03: the adapter skip makes ALLOW_UNREVIEWED_PUSH inert on Cloud, but desktop third-party Claude hooks still run the native guard, so that prefix is the escape there. Pin Reviewed-Commit to 7-40 hex, read dry-run stderr, put the transcript subagent first, and restore dispatch-pairing scope. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: compare only the new tip on a Cursor Task dry-run Git prints old..new on a fast-forward. Treating every displayed sha as shipped would reject a normal push. Compare only the right-hand tip. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: handle git-push `...` dry-run ranges and #2241 provenance The d8d5bc3 Claude child found the dry-run gate split on `..` inside a forced-update `...` summary, skipped lowercasing the fingerprint, treated an unrecognized Verdict: line ambiguously, and left the adapter-skip claim undated at four sites. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: key the prefix on the hook system and call parse_report The 0ae0db8 Claude child found Cloud home Claude settings after bootstrap, so "no home settings" was a false mechanism, and the hand-specified VERDICT_LINE recipe duplicated parse_report(). Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: require HEAD to still be the recorded sha after the child The 852aaf9 Claude child found the fingerprint gate's "or HEAD" reading admitted a stale pre-dispatch sha, cited _rev_parse for a check that lives in verify_review, and left pr-on-claim with no action once the prefix was withheld. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: require a clean HEAD-matching parse_report on compact copies The 51aa928 Claude child found AGENTS.md and skills/push treating parse_report() as enough, a stale "Cloud has no ~/.claude" line, and an exclusive "the one running" split that both-sources falsifies. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: name install-hooks as Cloud settings writer Address the Claude Opus child of 9488dac (bc-1ddb6849): stop attributing Cloud ~/.claude/settings.json to bootstrap.sh, qualify the adapter prefix as for the adapter's sake, split parse_report into decoder and verbatim-subagent routes, drop restore/again for #2241, phrase the persona call as required, name the last heading-bearing assistant text, and keep the parse_report gate once in skills/push. Rebut 40-55 char wraps (suggestion-weight; check-new-line-breaks is clean). Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: carve out empty pr-on-claim and qualify Cloud settings provenance Address the Claude Opus child of f22a0f5 (bc-e67d9844): authorize the empty pr-on-claim push when there is no report, require both parse_report routes to produce the report body, scope the import as the Cursor Cloud recovery instrument and name #2255 pre-push-review.py / #2309, separate the Cloud settings.json measurement from how that file got there, add a Do-Confirm refusal-gate checklist, and name the adapter checkout in both personas. Rebut restating the parse_report gate across compact copies (required by the prior child of 9488dac). Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: make the pr-on-claim carve-out decidable and scoped Address the Claude Opus child of 2da6c8c (bc-e24f4da3): drop the false claim that provenance lives in #2299, decide the empty-branch carve-out with git diff ...HEAD and require disclosure, scope that carve-out in AGENTS.md and keep the Claude Code prefix, name Morrison-Lab/ai-config's adapter in AGENTS.md, and state adapter inertness unconditionally. Filed #2310 for the remaining git-decidable gates. Rebut mid-phrase wrapping (suggestion-weight; the blocking new-line-breaks gate is clean). Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: scope Cursor review gates to the push checkout Address the Claude Opus child of 751875a (bc-5f5b3190): run all six refusal gates in the checkout whose push follows, recover the report through a file, name the adapter checkout in the push skill, add checklist pause points and killer items, and treat per-ref up-to-date and ref-deletion dry-run forms as not this procedure. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: correct deletion claim and tighten Cursor review gates Address the Claude Opus child of 3466ef7 (bc-d5e046e3): drop the false verify_review-allows-deletion claim, state carve-out exemptions on every vacated gate, relabel the gate list Read-Do with inline killer items, require the verdict-line form on Cursor as well as Claude Code, default ALLOW_UNREVIEWED_PUSH off until a native deny, name tree-equality vs net-zero other commits, record the #2241 landing sweep, and decode transcript.json once. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: grade Cursor reviews only from transcript.json Address the Claude Opus child of 92be0ac (bc-63006d82): parse_report always runs on route (a)'s transcript file, split :branch vs --delete handling, drop intra-passage native-deny restatements, and name parse_report as the fingerprint reader in both persona copies. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: decide the pr-on-claim carve-out with two git commands A Cursor Task child of f55e60c showed `_argv_push` excludes `--delete` rather than ignoring it, that the transcript is not the source for dry-run gates, and that the empty-branch carve-out had no positive test. Import parse_report from the worktree hook. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: import parse_report from the worktree, not ~/.claude A Cursor Task child of 95b55f4 showed ~/.claude/hooks is a directory symlink into the primary checkout, not a copy, and that writing the recovered report inside the push checkout trips the empty-status gate. Keep compact copies; drop parent-obligation lines from the reviewer persona. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: scope parse_report import to an ai-config worktree A Cursor Task child of b39eaac showed the worktree-hook import is unfollowable outside ai-config, that one matching dry-run line was enough to pass gates 5-6, and that skills/push dropped the missing-fingerprint refusal. Fetch only an adversarial-reviewer child. Add CLAUDE.md to the #2241 sweep. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: fail-closed empty-commit carve-out and scoped hook import Address Cursor Task child bc-74f98435 of a0124f7. Drop the undefined Route (a) label. Import ~/.claude/hooks only from a non-ai-config checkout. Test the parse_report script, not the hooks/ directory. Require both carve-out git commands to succeed. Put provenance on gate 2. Name the dry-run and source-ref checks. Scope "all six" to the git commands. Rebut collapsing AGENTS.md compact copies. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: last-assistant report selector and disclosure on post Address Cursor Task child bc-16017d04 of dbf359d. Return the structured report as the call's own message. Post the recovered file then the disclosure marker. Import ~/.claude/hooks only when that file exists on a non-ai-config checkout; otherwise CLI review. Add pr-on-claim.md to the #2241 sweep. Take the last assistant text; do not scan backward. Drop the circular step-0 link. Name the file pair for the inode measurement. Narrow the gate-list ordering claim. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: unify last-nonempty assistant selector and name messages Address Cursor Task child bc-7e680d49 of 6d665f7. Take the last non-empty assistant text; empty thinking and tool_calls records are not candidates. Iterate json.load's messages list. Fix the pr-on-claim ALLOW_UNREVIEWED_PUSH=1 cause (reviewer-call gate, not shipped-commit). Gate 3 consumes gate 2's tuple. Name the verdict-line form as parse_report's referent. Say the recovered file is both the parse_report input and the posted comment body. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: post recovered file only; quiet empty-commit carve-out Address Cursor Task child bc-497f3009 of 882ed05. Always recover the transcript file; a harness paste may corroborate and is not the posted body. Drop the ~/.claude/hooks import fallback (unfresh copy). Compact copies name the transcript recovery and the always-true reason not to import that path. Use git diff --quiet HEAD^ HEAD for the empty-commit carve-out. Drop the adjacent tuple-is-the-push-gate duplicate. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: name parse_report's accepted verdict forms Address Cursor Task child bc-a728241f of 8be8093. parse_report accepts Needs work, an optional heading, and spaces around the colon; wrapping the whole verdict line is no verdict. Call parse_report from the worktree hook on the recovered report. On a failed dispatch, write the CLI review to /tmp and parse that file. Anchor the settings.json measurement in docs/cursor-hook-mapping.md. Scope the push-skill recovery to Cursor Cloud Task. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: require .get on transcript text and a bare fingerprint Name that thinking records usually omit text, parse origin/main's parse_report when the diff touches the hook, show Reviewed-Commit as a bare line, and label the refusal-gate checklist as Read-Do. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: prefix empty-claim push and close parse_report's else Put ALLOW_UNREVIEWED_PUSH=1 on the pr-on-claim copy-paste push, restore the persona fail-closed else, name the worktree hook script in compact copies, state the two Read-Do dependencies, and say the recovered report file is the ARD input. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: move claim-push prefix out of the shared block Keep the agent-neutral git push in the pr-on-claim recipe, show the Claude Code prefix beside it, treat a missing fingerprint as a CLI-review case, tell the reviewer to emit nothing after the fingerprint, and scope the push-skill carve-out to the pushing checkout. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> * fix: name the -u dry-run upstream line and decoder heading check Treat Would set upstream as neither a mismatch nor other refs. The decoder, not parse_report, decides the four headings. Scope adapter-plus-native pairing to desktop. Name both-conditions-passing as the empty-claim carve-out. Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com> --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: dem-extra1 <dem-extra1@users.noreply.github.com>
…crutiny-hook-32a3
|
Claude finished review — View run Code reviewThis is a small, mechanical follow-up PR that binds
Verification performed:
I noted one pre-existing stale comment in No issues found. Checked for bugs and CLAUDE.md/lab-manual compliance. VerdictReady for merge Reviewed commit: 18628be |
3 participants
Closes #2263.
#2262 authored
hooks/remind-ums-on-scrutiny.pyand left it inKNOWN_UNREGISTEREDper README's activation gate. This follow-up binds it:UserPromptSubmitinject entry tohooks/hooks.jsonKNOWN_UNREGISTEREDshared/workflow/run-ums-proactively.mdwith the live plugin /--fixactivation sentenceLocal verification on
fee256c6:check-hook-catalog.pyreports 42 registered and 1 known-unregistered;test_check_hook_catalog.py11/11; hook suite 45/45;check-links.pyclean; markdownlint 0 errors.install-hooks.py --fixcannot bind this hook until this PR merges.Refs #2261 / #2262.