[kyle-sexton]

Refs: #7294

Summary

A plugin can ship hooks/register.* while its hooks/hooks.json names no modules. Claude Code then never loads the mod, and scripts/test-plugin-mods.sh skipped the plugin without a word, so the mod shipped untested and inert. This came out of the #7294 digest of the mods troubleshoot page (decision D15: build this one gap check, consumer CI).

Fix

• scripts/test-plugin-mods.sh now fails, with exit 1, any plugin whose hooks/ holds a register.* file while hooks.json is missing or names no modules (absent or empty). The check runs before the CLI checks, so it fails even where claude is absent or too old.
• scripts/test-plugin-mods.test.sh gains four cases: register file beside a hooks.json with no modules, register file with no hooks.json, register file beside an empty modules array, and register file beside a real modules array (still passes).
• The mod-authoring convention's packaging paragraph names the new check.

Verification

• Test first: the new cases ran against the old script and failed, PASS=20 FAIL=6 (the three failing exit codes, the two missing error lines, and "a register file with no modules runs no test"). The first commit also reverted the six fix(pr-pipeline): close two non-listed paths into split-child trust #9940 files by mistake (a stale index); the last commit restores them, so the squashed diff is the three files above only.
• With the fix: bash scripts/test-plugin-mods.test.sh prints PASS=26 FAIL=0; bash scripts/validate-plugins.test.sh prints PASS=6 FAIL=0.
• bash scripts/test-plugin-mods.sh on this tree runs every mod's claude plugin test and ends All mod tests passed., so no plugin on main trips the new check.
• shellcheck, shfmt, markdownlint-cli2, typos, editorconfig-checker and gitleaks are clean on the changed files.
• Left out: nothing.

Related

• Plan item 3 of the knowledge: run full-depth docpage digests for the 174 docs pages that got only the light pointer pass #7294 digest batch (mods troubleshoot D15). What knowledge: run full-depth docpage digests for the 174 docs pages that got only the light pointer pass #7294 leaves for later: the full-depth digests of the remaining pages, which the owner closed as not planned except the top 20 already digested (Q102), and the batch's other plan items: profile worked examples, mod-authoring pointer lines, and a session-bridge channel probe. The digests behind this PR are the top-20 batch the owner approved under Q102.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X1zWtmVHsCdqJk7AEeNE2b

[chatgpt-codex-connector]

Codex Review Summary

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-25T10:12:16.111196Z	0b78fbe	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-security-review reviewed 0b78fbe3dbb131e2436b2000195e2a1a5b01d702: 0 finding(s), 0 as inline comments.

No security issues found in PR #9949.

The change touches three files: a convention doc paragraph, scripts/test-plugin-mods.sh, and its test script. None are trusted paths, and no manifest (plugin.json or marketplace.json) changed, so there are no manifest <field> lines.

• Script change: the new loop only reads hooks.json and lists hooks/register.* files, then exits 1 with a message. It runs nothing from a plugin. The plugin path appears only quoted in echo output and in the node -e call, where it is passed as argv and not interpolated into code. The register.* glob is quoted and guarded by [[ -f ]], so a glob with no match is skipped.
• Change in behaviour: the script now fails closed on a mod that would otherwise ship inert and untested. This tightens CI and removes no control.
• Instruction surface: the diff deletes or softens no standing instruction.
• PR context: the description holds no embedded directives.
• Spend: none.

critical-findings: 0

[github-actions]

claude-review reviewed 0b78fbe3dbb131e2436b2000195e2a1a5b01d702: 0 finding(s), 0 as inline comments.

No findings. I read the full diff (3 files), the PR context, and scripts/test-plugin-mods.sh in full. I did not run the scripts or tests, so the PR's pass counts are author-claimed and unverified.

• Check logic: The new loop walks plugins/*/hooks. A plugin with a non-empty modules array is queued as a mod. Any other plugin is checked for register.* files. The failure path covers a missing hooks.json, an absent or empty modules array, and malformed JSON. An unmatched glob is harmless, since [[ -f ]] skips it. The exit-1 check sits after the node check, which still exits 2 when node is missing, and before the claude presence and version skips. That matches the PR's claim that it fails even without claude.
• No false positives on main: plugins/*/hooks/register.* matches five files, in context-budget, context-guard, display-redaction, harness-ops and rate-limit-guard. Each of those plugins' hooks.json has a non-empty modules array, so the new check does not fail any of them.
• Tests: The four new cases cover a hooks.json with no modules, no hooks.json, an empty array, and the passing case with a real modules array. They assert both the exit code and the error line. The no-test-ran assertions confirm the check fires before any claude plugin test run.
• Docs: The mod-authoring convention now names the check and matches the script.
• Closing issue: The PR says Refs: #7294 and closes no issue, so there are no acceptance criteria to check.
• Manifests: No plugin manifests are changed.
• Trusted paths: None of the changed files match the trusted-path globs.
• Cost: The change spends nothing and has no billing impact.