redgate: point hooks at the handlers that exist; sort runs in C locale - #136
Merged
Merged
Conversation
Two fixes lifted out of #115 so they do not wait on 76k lines. hooks.json referenced ${CLAUDE_PLUGIN_ROOT}/hooks-handlers/<h>.sh, but both handlers live under hooks/hooks-handlers/, and ${CLAUDE_PLUGIN_ROOT} is the plugin dir on an installed plugin (agent-compiler's own hooks.json resolves the same way). No plugins/redgate/hooks-handlers/ directory exists. So on every installed copy the PreToolUse write guard — the hook that stops a write reaching a ratified run's contract while its round is in MIDDLE — and the SessionStart announcer silently never ran. The cheap tier was green the whole time because nothing resolved handler paths against the filesystem. criteria-index.sh sorted run dirs with the host locale. Its own header promises deterministic output and --check compares against it, so an en_US host can order slugs differently from CI and report drift that is not there. LC_ALL=C pins it. New cheap-tier section 2b resolves every ${CLAUDE_PLUGIN_ROOT}/<rel> in every plugins/*/hooks/hooks.json to a file under plugins/<p>/<rel>, and fails closed if the walk finds nothing. It is resolved against the filesystem, not pattern-matched: voice's hooks.json uses the identical-looking hooks-handlers/ path and is CORRECT, because voice keeps its handler directly under the plugin root. Verified: cheap tier 1295 passed / 0 failed (1290 on main + 5 handler paths). Against main's pre-fix hooks.json the new section reports both redgate hooks as missing (1293/2). Two further mutations — an agent-compiler handler renamed away, and the walk pointed at an empty glob — each go red. Not verified here: the locale ordering difference. No en_US locale is installed in this container; LC_ALL=en_US.UTF-8 falls back to C and sorts identically, so the one-line fix rests on sort(1) semantics and #115's report of a red cheap tier on en_US hosts, not on a reproduction. The deep tier's path filter matches criteria-index.sh, but redgate ships no pier pack, so that check will report green because its leg is skipped, not because it ran. No SKILL.md, command, or skill references/ touched; demonstration discipline does not apply. No eval tier, job, or pack added, removed, or re-scoped, so docs/testing.md is unchanged. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XHYtoYrtWhTGy59GC9uJGd
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Contributor
There was a problem hiding this comment.
🟡 Changes recommended
The handler-path validation needs canonical containment checks and reliable walker error propagation.
Get a fresh assessment by requesting another Copilot review.
Pull request overview
This PR fixes Redgate hook paths, makes criteria indexing locale-independent, and adds cheap-tier handler validation.
Changes:
- Corrects both Redgate hook handler paths.
- Forces criteria sorting with
LC_ALL=C. - Adds filesystem-based hook path checks.
File summaries
| File | Summary |
|---|---|
plugins/redgate/skills/criteria-contract/scripts/criteria-index.sh |
Makes run ordering deterministic across locales. |
plugins/redgate/hooks/hooks.json |
Corrects handler locations. |
evals/cheap/run.sh |
Validates resolved hook handler paths. |
Review details
Suppressed comments (1)
evals/cheap/run.sh:82
- The Python walker runs in process substitution, so its non-zero exit status is not propagated by
set -o pipefailand the loop only checks whether it saw any output. If an earlier plugin emits valid paths and a later syntactically valid but structurally malformedhooks.jsonraises inside this walker,hook_paths_seenremains nonzero and section 2b reports no failure instead of failing closed. Capture and check the walker's exit status (or run it as a normal command) before accepting the results.
)
- Files reviewed: 3/3 changed files
- Comments generated: 1
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a9145b8edd
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… is nothing-to-check Two corrections to the section added in a9145b8, both found by running it against something other than the tree it was written on. 1. The counterfeit corpus's synthetic marketplace ships no hooks.json at all, and the section's "zero paths resolved" fail-closed turned that root red: the corpus's own calibration guard reported "baseline plugin is NOT green — corpus is miscalibrated, every rejection below is meaningless" and the required counterfeit tier failed on CI. Reproduced locally (24 passed / 1 failed) before touching anything. The section now counts hooks.json FILES separately from extracted PATHS: no files is a legitimate "nothing to check", the same PASS shape every other section uses for an absent surface; files present but zero paths extracted still fails closed, because that is the walk breaking, not the plugins being clean. 2. The resolver checked isfile() only, while the section's name promises containment. `${CLAUDE_PLUGIN_ROOT}/../shared/x.sh`, or a symlink out of the plugin, would report OK when the target exists even though the installed command resolves outside the plugin root. Copilot's finding on #136. Both paths are now realpath'd and the target must share the plugin dir as its commonpath — the same idiom the relative-link resolver further down already uses. Verified, in order: counterfeit harness on the fixed file: 25 passed / 0 failed, baseline green, all 18 counterfeits rejected by their expected gate (was 24 / 1). cheap tier on the repo: 1295 passed / 0 failed, unchanged. Five mutations against the final section, each restored after: main's pre-fix redgate hooks.json -> 2 FAIL (1293 / 2) agent-compiler handler renamed away -> 1 FAIL (1294 / 1) ../ escape to a REAL file in redgate -> 1 FAIL (1294 / 1) isfile alone said OK extraction regex broken, 3 files kept -> 1 FAIL "3 file(s) present but the walk extracted ZERO handler paths" glob pointed at a root with no hooks -> PASS "nothing to check" (1291 / 0) The regex mutation's first attempt did not apply (a sed pattern that missed the heredoc's escaping left the file untouched and the tier at 1295 / 0); that was a null test, not evidence, and is recorded here so the second, applied attempt is the one that counts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XHYtoYrtWhTGy59GC9uJGd
Owner
Author
|
Pushed Counterfeit tier was red, and it was this PR's fault. The corpus's synthetic marketplace ships no Copilot's containment finding was correct and is fixed — both paths are Five mutations against the final section, cheap tier 1295 / 0 unchanged. One null result recorded honestly: the regex-break mutation's first attempt didn't apply (a Two things this run revealed that are outside this diff:
Generated by Claude Code |
Codex's finding on #136, confirmed by reproduction before touching anything: the handler-path scanner ran in process substitution, so its exit status was discarded. A manifest that is valid JSON but not the shape the walk expects — `"command": null` — raised TypeError, the remaining manifests were never scanned, and because an EARLIER manifest had already emitted records the zero-paths guard stayed quiet. On the unmodified file, nulling voice's command made the `voice hooks.json SessionStart` line simply vanish; section 2b reported no failure and the required tier exited 0 (1294 / 0), while stderr carried the traceback nobody reads. Nulling agent-compiler's instead went red only incidentally, via the zero-paths guard, and misreported "1 file(s)" when three exist — the scan had been silently truncated. Three layers now, each verified on its own: * Every shape the walk touches is checked, not assumed. Nonsense in a valid-JSON manifest emits a MALFORMED record naming the plugin and event instead of raising, and the walk continues to the next manifest. * The scanner writes to a temp file and its exit status is captured; a nonzero one is a recorded FAIL saying an unknown number of manifests were never checked. * The temp file itself fails closed if it cannot be created, and the "nothing to check" PASS moved inside the scanner-ran branch, so no silent-scanner path can produce that green line. Verified against the final file: null command, last manifest -> FAIL "MALFORMED hook entry (a hook has no string "command")" 1294 / 1, exit 1 null command, first manifest -> same FAIL, and redgate + voice still scanned after it 1294 / 1 crash on every manifest -> FAIL "scanner exited 1" + zero-paths FAIL crash on the LAST manifest -> FAIL "scanner exited 1" 1294 / 1 (the exact earlier-records shape that used to pass) "hooks" not an object / event not a list / entry's hooks not a list -> one MALFORMED FAIL each TMPDIR unwritable -> FAIL, and no "nothing to check" green The five prior mutations unchanged: main's pre-fix redgate (2 FAIL), handler renamed (1), ../ escape to a real file (1), regex broken (now reports "3 file(s)", not 1 — the walk no longer truncates), no-hooks root (PASS). cheap tier 1295 / 0 unchanged; counterfeit harness 25 / 0, baseline green. The pre-fix reproduction and the full mutation sweep were run by a delegated agent; the fixed-file cases above, both roots, and the hooks.json files being byte-identical to HEAD were re-verified independently before this commit. Survey, not fixed here: sections 10 (per-plugin safety-pack discovery) and 12 (install-smoke coverage) feed `while read` from the same `< <(python3 …)` shape and neither asserts the enumerated count against marketplace.json, so a scanner crash after N plugins would silently run only N packs and stay green. Section 10 is the one this repo's safety story rests on. Follow-up. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XHYtoYrtWhTGy59GC9uJGd
JRichlen
pushed a commit
that referenced
this pull request
Sep 18, 2026
The merge brought in #136, which added checks. The figure this branch introduced (~19s / 1290 checks, measured 2026-09-14) was already stale against its own instruction to re-measure rather than trust it. Measured on this checkout at 2026-09-18: 13.6s, 1296 checks, 25 plugins. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb
JRichlen
added a commit
that referenced
this pull request
Sep 24, 2026
… fits instead (#134) * Research: where the Harness Knowledge Graph lands in this marketplace Placement analysis for the Harness Software Delivery Knowledge Graph, diffed against every shipped skill and against this repo's own prior verdict on graph memory in docs/research/agentic-patterns-corpus.md. Verdict: not a new plugin. Harness's payoff needs heterogeneous, non-git-native estate data with no single strongly-consistent query surface; a repo fleet already has one. But the corpus rejected graph memory on evidence quality (one 198-document paper), and Harness retires that reason — the rejection is re-based on domain fit, which is stronger and survives. Lands as evidence inside two shipped skills: - eval-ladder: their published eval methodology is a production eval ladder for a retrieval system, and names a rung we don't — "a registered relationship is not necessarily a usable relationship" (declared != populated != fresh). Their product-backed validation is independent, citable confirmation of "grade the surface closest to the harm". - fleet-playbook-curator: three of Harness's four theses were arrived at independently (canonical identity on node_id, freshness as an independent clock, index-not-CMDB). The fourth is a verified gap — the fleet manifest is a flat entity table with no edge field, so every relationship exists only as uncited prose. Records what the research could not establish, including that the 15-25x token claim is an unreplicated vendor benchmark and that the docs page was read through a search-index fetch because the origin is egress-blocked here. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Correct the fleet-playbook relationship claim: not uncited, unmodeled Codex flagged the note as asserting that relationship prose in a fleet playbook is "uncited by construction". Checked against source, and the finding is right: - SKILL.md:98 names "cross-repo interactions" as exactly the kind of thing that belongs in a playbook. - The per-claim rule covers EVERY substantive claim, relationships included: repo@sha:path plus an as-of stamp, or omitted/flagged STALE. - validate-citations.sh fails the build on a claim citing a repo not read this pass, or a path absent from that repo's gathered tree. So a relationship claim is cited, and its citation is machine-checked for traceability. Saying otherwise understated a safety mechanism the plugin actually has, and rested the edge argument on a false premise. The gap is narrower and survives restating. Relationships are not MODELED, with two consequences: - No edge is diffable. diff-fleet.sh cascades over membership and pushed_at; the staleness clock stamps head_sha per member and nothing per edge, so a relationship that stops holding raises no signal of its own. - Traceable is not supported. validate-citations.sh says so in its own comments: semantic support is the behavioral layer's job. For a single-repo claim the cited file usually is the evidence; for an edge the evidence is the join, and a claim can cite two real, genuinely read paths while asserting an edge neither supports. Cheap tier: 1290 passed, 0 failed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Drop "strongly-consistent" for "authoritative" — the argument needs the weaker claim Copilot flagged that the note asserts there is "no single strongly-consistent query surface" for a DevOps estate and then names `gh api` as one for a repo fleet. GitHub publishes no consistency guarantee for REST list endpoints, so the phrasing reads as a guarantee the note cannot substantiate. The finding is right, and the fix costs the argument nothing: what actually carries it is that ONE surface is the system of record and its join key (node_id) is stable — not any ordering or read-your-writes property. Both occurrences now say "authoritative". Added a parenthetical noting that fleet-playbook-curator's own prose calls `gh api orgs/<owner>/repos` "strongly-consistent", and that the defensible contrast it reaches for is with the Search API, which documents its indexing lag. Left that SKILL.md wording alone — out of scope for a docs note, and worth its own look. Cheap tier: 1290 passed, 0 failed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Fix six review findings in the Harness research note A full review of the note against this repo's own source turned up six claims that do not hold. All six are corrected here; the note's argument and verdict are unchanged. - diff-fleet.sh does not key on pushed_at. It joins on node_id and its content-drift bucket keys on head_sha (diff-fleet.sh:28-30) — head_sha is the diff's own key, not a separate clock beside it. The "no edge is diffable" point survives; the mechanism given for it was wrong. This error was introduced by c1d3d9a, the commit that fixed the last one. - eval-ladder already has the rung. Rung 1 (discriminating corpus) sits exactly between structural and code assertion, so what is missing is the retrieval-layer vocabulary, not a rung. - The fleet-playbook table counts now agree: four rows, four theses, and the gap named as a fifth item absent from the table. - Provenance said five sources; the list has six. The uncounted product page is load-bearing further down. - "CMDB" is Harness's vocabulary, not ours. SKILL.md:45 says "not a runbook" and says nothing about a CMDB. - The corpus now back-links here. agentic-patterns-corpus.md and its .json twin carried the retired evidence-quality rejection with no pointer, so a reader arriving via red-gate-protocol.md:745 read a dead verdict as live. The original verdicts are left intact and annotated: the evidence half is marked superseded, the domain-fit half upheld and now scoped to git-native repo data. Cheap tier: 1290 passed, 0 failed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Add companion research note on LLM-generated wikis and repo maps Asks what structure is right for the fleet-playbook edge gap, which the Harness note named but never answered. Surveys the LLM-wiki family (DeepWiki, Google Code Wiki, DeepWiki-Open, Auto Wiki), repo maps, SCIP/LSP indexes, embeddings, llms.txt, AGENTS.md, graphify and doctests. Verdict: nothing in the family closes the gap, and PR #134's verdict stands. The sharpening is worth more than the answer — on a DeepWiki page pinned to a commit that is still HEAD (zero drift), three of five rows in the architecture table are wrong, one naming a function absent from the file. Every citation resolves. Traceability was never the problem, which is the gap validate-citations.sh names in its own comments, reproduced at industrial scale. Derives an entry condition for any future edge: an edge is only worth adding if it comes with a command that re-derives it and a comparison that can go red. That disqualifies one of the three candidate edges in the companion note. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Correct two claims that verification found false Both were caught by dives verifying this repo's own assertions against primary sources, and both were load-bearing where they sat. harness-knowledge-graph.md called node_id "stable". GitHub does not say that. The GraphQL global-node-ID guide promises only that "it's best practice to persist the global node ID so you can easily reference objects across API versions", and the migration guide states that "The legacy format will be closing down and replaced with a new format" — so the identifier's string is on the record as changing, with no shutdown date. The verdict's domain-fit argument rests on this claim, so it now says what is actually supported: node_id is opaque, unambiguous and independent of the mutable full_name within a pass. It does not license joining a manifest captured before the format migration against one captured after, which is what diff-fleet.sh does across passes — a fleet straddling that boundary sees every member as removed plus added rather than renamed, and nothing in the plugin would say why. AGENTS.md:63 and evals/cheap/run.sh:3 both claimed the cheap tier runs "under a second". Measured on this checkout: 18.63s for 1290 checks across 25 plugins. Off by roughly 19x, in the file that tells every contributor and every agent what the tier costs. Replaced with the measured figure, its date, and the fact that it scales with plugin count — a number that will go stale again is worth less than the instruction to re-measure. Cheap tier green after the change: 1290 passed, 0 failed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Add knowledge-management corpus: 140 patterns, 7 dives, 57 corrections Ten scouts swept ten source clusters — retrieval and indexing, code graphs, context files, agent memory, provenance and freshness, developer portals, semantic layers and lineage, enterprise search and MCP, evaluating knowledge systems, and knowledge-management prior art outside software. Seven dives then re-verified the load-bearing claims against primary sources, organised by cross-domain convergence rather than by scout. Verdict: nobody machine-checks a knowledge claim. The field's answer is to make claims that do not need checking — derive them from a parser or compiler, execute them, or accept them as unchecked and label how they were produced. Where a semantic check is attempted the measured ceiling is ~77% balanced accuracy on prose (0.55 informedness against a 50% chance baseline), ~61 on the hardest prose split, ~58 on contested cases, and 0.17 span-F1 for a prose-trained checker on code as evidence. Three findings the survey did not expect: Fail-open is the norm and nobody says so. Every expiry and drift mechanism examined degrades to green rather than red — an env var that skips every check, a swallowed network error, a doc preprocessor silent on a missing anchor for nearly seven years, an ownership gate that goes vacuous exactly where ownership is broken. The rule "the comparison must be able to go red" is insufficient; it must also go red when it cannot be made. Citation is not transcription, proven on this corpus itself. 57 of ~140 scout claims needed correction — not for missing citations but for misread ones. The cleanest instance is external: "84% of KM programmes fail" traces through three hops of valid, resolvable citations to a 1997 article saying the rate is one third, and that the number is an estimate rather than a study. Pinning where is not pinning what, and repo@sha:path pins only where. This marketplace is further ahead than its own scouts believed and its best evidence is buried: verify-before-claim ran a negative control three times across six scenarios, got a null every time, and ships without a calibration case with the reasoning recorded — an in-house reproduction of the year's most-cited context-file null, reached before the paper it matches was revised. It lives in a YAML comment no tier or doc references. Two claims recorded but deliberately not acted on, because both are plugins/** changes that trigger the behavioral tier and the demonstration review gate: SKILL.md:65 asserts GitHub's "stable node_id", which GitHub declines to guarantee; and index.schema.json requires repo ("owner/name") with additionalProperties false and no node_id field, while validate-citations.sh matches on full_name, so a rename leaves every prior claim carrying a dead name that can resolve to the wrong repository. Cheap tier: 1291 passed, 0 failed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Re-measure the cheap tier after merging main The merge brought in #136, which added checks. The figure this branch introduced (~19s / 1290 checks, measured 2026-09-14) was already stale against its own instruction to re-measure rather than trust it. Measured on this checkout at 2026-09-18: 13.6s, 1296 checks, 25 plugins. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * fleet-playbook-curator: key citations on node_id, not the mutable repo name The plugin's thesis is that membership is "joined on node_id, never full_name". That held for the manifest and the diff and was abandoned at the claim ledger: gather-context.sh reduced each member to full_name and dropped node_id one line before it was needed, so validate-citations.sh could only match on the mutable name, and index.schema.json had no node_id field with additionalProperties:false forbidding one. Measured, not assumed. A plain rename already failed CLOSED (the old name is absent from context.json, so the claim reads as fabricated). The case that actually bites is a repo name later REUSED by a different repository: the stale citation matched the impostor's gathered surface and resolved to the wrong repository, silently, exit 0. GitHub's own rename doc warns that reusing an old name retargets the redirect. Threaded node_id through the seam: gather-context.sh carries it into every context entry, the schema admits it as an OPTIONAL claim field, and validate-citations.sh prefers it when present. A rename now keeps the citation resolving (with a note); a reused name fails closed. Ledgers written before the field validate unchanged on the old full_name match, and the validator now says so on every run instead of leaving it implicit. Two prose claims corrected in the same pass, both verified unsupported against GitHub's own documentation: - "GitHub's stable node_id" -> opaque. GitHub promises only that it is the value to persist "across API versions" and states the legacy format "will be closing down and replaced with a new format". - "strongly-consistent rate bucket" -> authoritative. GitHub publishes no consistency guarantee for REST list endpoints. This is the same claim Copilot flagged in the research note on this branch; the note's copy was fixed and the plugin's original was left behind. Regression coverage that can go red: a cite-index-hijack.json fixture plus three checks. Verified by reverting the node_id preference and watching the hijack check fail, then restoring it. Cheap tier 1300 passed, 0 failed. Touches plugins/*/skills/**/scripts/** — a deep_safety_path — so the gated pier run fires on this PR by design. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb * Move the fleet-playbook-curator fix to its own PR Reverts b4a4dc4 on this branch. The change is not being dropped: it now lives on claude/fleet-playbook-node-id-citations as its own pull request, cherry-picked onto main. This PR was opened as docs-only, and its body says so: no plugins/**, no SKILL.md, so no behavioral tier and no demonstration comment. b4a4dc4 made that statement false by adding a safety-path change to plugins/*/skills/**/scripts/**. Splitting makes the body true again, lets the research notes merge without waiting on a pier run they don't need, and gives the plugin change its own review, demonstration comment and deep tier run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DghMFX2SuCASz3fjHoYwjb --------- Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two fixes lifted out of #115 so they do not wait on 76k lines. Three changed lines, plus a cheap-tier section that would have caught the first one.
The shipped defect
plugins/redgate/hooks/hooks.jsonreferenced${CLAUDE_PLUGIN_ROOT}/hooks-handlers/<handler>.sh. Both handlers live underhooks/hooks-handlers/, and${CLAUDE_PLUGIN_ROOT}is the plugin directory on an installed plugin —agent-compiler's ownhooks.jsonresolves the same way. There is noplugins/redgate/hooks-handlers/directory.So on every installed copy of redgate, the
PreToolUsewrite guard — the hook that stops a write reaching a ratified run's contract while its round is inMIDDLE— and theSessionStartannouncer silently never ran. The cheap tier stayed green throughout, because nothing resolved handler paths against the filesystem.The locale sort
criteria-index.shsorted run directories with the host locale. Its own header promises deterministic output and--checkcompares against it, so anen_UShost can order slugs differently from CI and report drift that is not there.The guard
New cheap-tier section 2b resolves every
${CLAUDE_PLUGIN_ROOT}/<rel>in everyplugins/*/hooks/hooks.jsonto a file underplugins/<p>/<rel>, onePASS/FAILper handler, and fails closed if the walk finds nothing. It is resolved against the filesystem rather than pattern-matched on purpose:voice'shooks.jsonuses the identical-lookinghooks-handlers/path and is correct, because voice keeps its handler directly under the plugin root. A pattern rule would have flagged the healthy plugin and could not have told the two apart.Verification
main+ 5 handler paths)main's pre-fixhooks.jsonNot verified here, stated rather than implied:
en_USlocale is installed in this container;LC_ALL=en_US.UTF-8falls back toCand sorts identically. The one-line fix rests onsort(1)semantics and feat(evals): marketplace-wide agentic test framework (T01–T52) with red-team lane #115's report of a red cheap tier onen_UShosts, not on a reproduction.criteria-index.sh(plugins/*/skills/**/scripts/**), but redgate ships no pier pack, sodeep tier (pier)will report green because its leg is skipped, not because anything ran. That is the existing gate design, named here so a green is not read as proof.No
SKILL.md, command, or skillreferences/touched, so demonstration discipline does not apply. No eval tier, workflow job, or eval pack added, removed, renamed, or re-scoped, sodocs/testing.mdis unchanged and section 20 stays green.Independent of the OpenRouter credit outage blocking #131 and #133 — nothing here calls a model.
🤖 Generated with Claude Code
https://claude.ai/code/session_01XHYtoYrtWhTGy59GC9uJGd
Generated by Claude Code