Check

Pre-computed context

!"${CLAUDE_PLUGIN_ROOT}/scripts/precompute.sh" check 2>/dev/null || echo "- Orientation unavailable"

Purpose

A gate, not an audit. It answers one question with one exit code: would every path-scoped instruction surface in this repository actually load when it is supposed to, and does the always-loaded index still describe what is on disk?

It exists because every failure mode here is silent. Claude Code does not warn when a rule's glob matches nothing, when a bracket expression is malformed, or when a pattern exceeded the brace budget and was used unexpanded. The rule simply never fires, and the convention it carries is quietly absent from every session that needed it. A repository can accumulate a dozen dead rules and look perfectly healthy.

Wire it into CI beside the linters. It is fast, deterministic, and has no judgment layer to drift.

What it checks

The last one is the least obvious. Claude Code reads CLAUDE.md, not AGENTS.md. A repository carrying both with no import between them gets a perfectly-generated, perfectly-in-sync index that never enters context, the entire subagent-gap mitigation doing nothing while every other check reports green. Sync and reachability are independent questions; ask both.

Over-broad is the one warning rather than a failure: breadth is a judgment about whether a demotion was worth making, not a statement that the rule is broken. Everything else is a hard fail.

Running it

"${CLAUDE_PLUGIN_ROOT}/scripts/glob-tools.sh" rules
"${CLAUDE_PLUGIN_ROOT}/scripts/render-index.sh" check --file <index-file>
"${CLAUDE_PLUGIN_ROOT}/scripts/render-index.sh" reachable --file <index-file>

<index-file> is a precedence order, not a procedure. Take the first that exists:

Name the winner in the report; the run is not complete until the output states which file was checked. If none exists, say so and skip the index check rather than inventing a target. The glob checks still decide the verdict.

A repository with no index block yet is not a failure. render-index.sh check exits 3 for that case, and 3 means "nothing to compare", not "broken". Report it as a recommendation. The index is what makes deferred rules reachable from subagents. Leave the gate's verdict to the glob checks. Only a repository that has an index and has let it drift fails on that check.

Escalating to empirical verification

Every check above is static. It reads files and reasons about what Claude Code would do. When the operator asks why a rule is not firing despite a green gate, or wants proof rather than inference, escalate:

"${CLAUDE_PLUGIN_ROOT}/scripts/verify-load.sh" --trigger <covered file> --expect <the rule>

That drives the real CLI with an InstructionsLoaded hook and reports what actually loaded, with the reason for each load. It costs a model call, so it is an escalation rather than part of the default gate. VERDICT UNKNOWN means the probe could not run. Report it as unmeasured, never as a pass.

Pass --breadth-max through when the operator supplies it. The default of 75% is a starting point, not a law; a documentation repository where **/*.md legitimately covers most files should raise it rather than live with a standing warning.

Reading the output

Both scripts emit deterministic TSV. glob-tools.sh rows are:

PATTERN <tab> source-rule <tab> pattern <tab> expanded <tab> matches <tab> status
SUMMARY <tab> tracked <tab> patterns <tab> invalid <tab> overbroad <tab> verdict

Report per failing pattern: the rule file, the pattern, the status, and what the operator should do about it. The three failure statuses have different fixes and saying "invalid" helps nobody.

Hard rules

• Read-only. No Edit, no Write, no mutating Bash. Fixing a broken glob is the sibling realign skill's job, or the operator's.
• Never fabricate a verdict. If a script cannot run, because this is not a git repository or tooling is missing, report that plainly and exit non-zero. A gate that passes because it could not measure is worse than no gate.
• Never consult the findings artifact. This skill verifies the repository's actual state. An audit artifact is a snapshot of a past run, and a gate that trusted one could report health for a repository that has since broken.
• Exit code is the product. 0 clean, non-zero otherwise, whatever the prose around it says.

Gotchas

• Exit 3 is not a failure. render-index.sh check returns 3 when the file carries no index block at all. Treating that as drift fails every repository that has not adopted the index yet, which is most of them on first run.
• A begin marker with no end marker returns 3, not 1. The tools refuse to guess a block's extent rather than overwrite an unknown span. Read it as "malformed, needs a human", not as "absent".
• over-broad never moves the verdict. It shares an output column with the three real failures and is the one that does not fail the gate. Reporting it as a failure trains operators to ignore the gate.
• A zero-match glob usually means stale content, not a typo. The reflex is to fix the pattern. Check whether the code it described still exists first. A rule for a deleted subsystem should be retired, not re-globbed.
• Passing because nothing could be measured is the worst outcome. Outside a git repository, or with tooling missing, the scripts cannot answer. Exit non-zero and say why; a green gate that measured nothing is a false assurance a reviewer will act on.
• An in-sync index can still be inert. check and reachable answer different questions, and a repository can pass the first while failing the second. Reporting "index in sync" without the reachability verdict is the exact false assurance the previous point warns about.
• Rules discovery follows symlinks and does not require git. A symlinked rule points outside the repository by design. That is the documented way to share one rule set across projects, so it is never tracked. Nested instruction files are the opposite: tracked-only, because an untracked or vendored AGENTS.md must never reach the consuming repository's always-loaded surface. If a rule seems missing from a report, check which of the two rules applies before assuming a bug.