SKILL.md
Verify Bundle Sync
This skill exists solely to work around an upstream Claude Code symlink bug (anthropics/claude-code#53948) that requires plugins/dev-workflow-bundle/skills/<name>/ to be a real directory copy of skills/<name>/ rather than a symlink. It is a project-local skill (lives under .claude/skills/verify-bundle-sync/, not registered in .claude-plugin/marketplace.json). When the bug is fixed and the bundle layout returns to symlinks, delete this skill directory, the .claude/dev-workflow.md test_commands entry, the dev-workflow-triage (d4) sub-step, and the .claude/rules/project.rules.md bullet that document this workaround.
The skill compares each bundle member's canonical directory against its bundle copy and reports drift. It is detect-only — it never modifies any files.
Process
Accepts an optional --base-commit <sha> argument (ignored — the scope is structural, not changeset-dependent). Running with no arguments behaves identically.
Run the following directly in the main thread (no subagent dispatch is needed — the check is lightweight: one jq invocation plus one diff -rq per bundle member).
- Load the bundle membership list from
.claude-plugin/marketplace.json:
``bash bundleskills=$(jq -r '(.plugins[] | select(.name == "dev-workflow-bundle") | .skills[]) // empty' .claude-plugin/marketplace.json 2>/dev/null) # If $bundleskills is empty after this guard, halt and emit the Return contract response # (Layer 1 prose Status: EXECUTION_ERROR + Layer 2 fenced JSON with status: "error", # reason: "marketplace.json missing, malformed, or dev-workflow-bundle plugin entry absent") # immediately — see § Return contract. ``
The // empty is the array-enumeration null-fallback idiom: when the entry / array is absent, it yields a zero-length stream (no literal null\n leaking to stdout). This is a different concern from the canonical scalar // "unknown" pattern documented in .claude/rules/project.rules.local.md § jq の null 文字列フォールバック, which targets scalar values. The post-pipeline [ -z "$output" ] guard catches array absence, jq non-zero exit, and file-not-found uniformly.
- For each bundle member entry
./skills/<name>:
- Resolve canonical=skills/<name>/ - Resolve bundlecopy=plugins/dev-workflow-bundle/skills/<name>/ - Verify both directories exist with test -d "$canonical" && test -d "$bundlecopy". If either is missing, exit immediately with EXECUTIONERROR and report which path was missing. - Run diff -rq "$canonical" "$bundlecopy". Capture stdout. If exit code is non-zero AND stdout is empty, treat as EXECUTIONERROR (tool failure). If stdout is non-empty, treat every output line as a drift entry — each line is one of: - Files <canonical-path> and <bundle-copy-path> differ → type: "differ" - Only in <canonical-dir>: <file> → type: "onlyincanonical" - Only in <bundle-copy-dir>: <file> → type: "onlyin_copy"
- Aggregate the result:
- All entries drift-free → SUCCESS (e.g. 6 bundle skills verified, 0 drift) - Any entry has drift → TESTFAILED. Include the per-entry drift list and a remediation hint of the form cp -R skills/<name>/. plugins/dev-workflow-bundle/skills/<name>/ for each affected member - jq failed / diff missing / marketplace.json unreadable / per-entry path missing → EXECUTIONERROR
EXECUTIONERROR is deterministic within a run: marketplace.json absence, missing tooling (jq / diff), and missing path entries do not become resolved during the same run, so retrying the same invocation will not change the outcome. Callers that retry on EXECUTIONERROR (such as dev-workflow Phase 9's retry handler) will simply burn through their retry budget producing the same error each time — that wastes a few extra invocations but is harmless.
Return contract
The skill emits its result in two layers in a single response so that both prose-reading callers (such as dev-workflow Phase 9) and JSON-parsing callers (such as dev-workflow-triage (d4)) can extract the verdict mechanically.
Layer 1 — Prose summary (first, at the top of the response):
Status: SUCCESS | TEST_FAILED | EXECUTION_ERROR
<one-paragraph human-readable summary>
<if TEST_FAILED: per-entry drift list with remediation hint lines>
<if EXECUTION_ERROR: reason and which step failed>
Layer 2 — Fenced JSON verdict (last, at the end of the response):
````
{
"status": "ok" | "drift" | "error",
"checked_count": <int>,
"drift_files": [{"skill": "<name>", "path": "<relative-path>", "type": "differ|only_in_canonical|only_in_copy"}],
"reason": "<optional, required when status=error>"
}
````
Mapping between the prose status token and the JSON status field:
Prose Status: |
JSON status |
|---|---|
SUCCESS |
ok |
TEST_FAILED |
drift |
EXECUTION_ERROR |
error |
checkedcount: number of bundle member entries actually inspected (8 at the time of writing —ask-peer,dev-workflow,extract-rules,rules-review,tidy,prose-polish,mobpro,kabeuchi). If the list could not be loaded (EXECUTIONERRORfrom Step 1), set this to0.drift_files[]: drift / one-sided-presence entries, populated only forstatus: "drift". Empty array forokanderror. Thepathvalue preserves the raw line as it appeared indiff -rqoutput so that downstream rendering does not need to re-derive it.reason: required onstatus: "error". Short, ≤ 80 characters. Examples:marketplace.json missing,dev-workflow-bundle plugin entry absent,jq not in PATH,canonical missing: skills/ask-peer.