Arguments. <status|search|check-all|scan|list|quality|create> [args]. Full form: [args]. Actions: status (default), search, check-all, scan, list, quality, create. e.g., 'search Stop hook', 'check-all', 'create bug "title"'
Pre-computed context
Claude Code version: !claude --version 2>/dev/null || echo "unknown"
Variables
Arguments: $ARGUMENTS
Purpose
Claude Code ships frequently. Features that worked last week may have new bugs today. This skill serves three purposes:
- Quick health check. Snapshot of registry health and model quality (no args needed)
- Proactive search. Find known bugs BEFORE you build on top of a feature
- Issue registry. Maintain single source of truth for the Claude product GitHub issues you track, record what they block, and create follow-up work items when bugs are fixed
Data Files
The issue registry (registry.json) persists in the registry directory. By default this is the
plugin's per-machine data directory (${CLAUDE_PLUGIN_DATA}, survives plugin updates). A repository
can keep the registry inside itself, git-tracked and team-shared, with registry_dir in its
harness-ops.yaml team file (convention home, default docs/conventions); one person can set it with
the registry_dir plugin option or the personal harness-ops.local.yaml. The personal file beats the
team file, which beats the option.
Registry location. Apply to EVERY registry_manager.py invocation: resolve it once per action,
before the first registry call:
node "${CLAUDE_PLUGIN_ROOT}/scripts/registry-dir.mjs" --project "${CLAUDE_PROJECT_DIR}" --user-registry-dir '${user_config.registry_dir}'
It prints JSON: layer (default, userConfig, team or local), path (the file that set it),
data_dir, and errors.
data_diris a path → pass--data-dir "<data_dir>"exactly as printed. The script has already checked containment and that the value holds only letters, digits,.,_,-,/and\, so nothing in it runs in the shell; do not join or normalize it yourself.data_diris null → pass--data-dir "${CLAUDE_PLUGIN_DATA}". We never trust an inheritedCLAUDE_PLUGIN_DATAto be this plugin's, so the script uses one only when it names harness-ops. Which processes receive the variable is read live from plugins manifest reference: where each variable resolves (as of 2026-11-10; recheck when that table exports the variable to the Bash tool).errorsis non-empty → each names a layer, file, key and value that was dropped (an uncontained path, a character outside that set, or a value that is not a string). Show every one to the user with the layer that won instead, and direct/harness-ops:setup. A dropped team or option value resolves to the plugin data directory, never to the layer below it.- The script cannot run (no
node) → stop the registry action and say so; do not guess a directory.
| File | Purpose | Who edits |
|---|---|---|
<registry-dir>/registry.json |
All tracked Claude product GitHub issues with status, category, affected files, and what's blocked | Skill (on search/scan --add/check-all), via scripts/registry_manager.py. See the registry-location rule above |
Read context/registry-schema.md for the full registry.json schema and field enums.
Action Router
The first token of $ARGUMENTS is the action:
| Action | Description |
|---|---|
status |
Registry stats + quality snapshot. Default with no args |
search |
Search GitHub Issues for bugs on one feature |
check-all |
Check every tracked registry issue; find newly resolved ones, create follow-up TODOs |
scan |
Grep the current repo for GitHub issue references and report untracked ones (read-only); scan --add also registers them |
list |
All tracked issues grouped by status and category |
quality |
Benchmarks + status page + recent GitHub degradation reports |
create |
Draft and file a GitHub issue on anthropics/claude-code in its template format |
When the first token is not an action, run search with the whole argument as the feature name.
Action: status
Read context/action-status.md for full status process and output format. DEFAULT action when no args.
Action: search
Read context/action-search.md for full search process, flags, and output format.
Action: check-all
Read context/action-check-all.md for full check-all process and output format.
Action: scan
Read context/action-scan.md for full scan process and output format.
Action: quality
Read context/action-quality.md for full quality check process, sources, and output format.
Action: list
Read context/action-list.md for full list process and output format.
Action: create
Read context/action-create.md for full issue creation process, mandatory gates, and safety rules.
Repos Reference
| Product | Repo | When to search |
|---|---|---|
| Claude Code | anthropics/claude-code |
Default. Hooks, skills, settings, plugins, tools |
| Python SDK | anthropics/anthropic-sdk-python |
Python API client issues |
| TypeScript SDK | anthropics/anthropic-sdk-typescript |
Node.js API client issues |
| C# SDK | anthropics/anthropic-sdk-csharp |
.NET API client issues |
| MCP Protocol | modelcontextprotocol/python-sdk |
MCP server/client issues |
| Azure MCP | microsoft/mcp |
Azure MCP server issues |
Integration Points
With your work-item tracker
check-allproposes follow-up work items for resolved issues (file them with whatever tracker the consumer project uses, e.g.gh issue create, after user confirmation)searchsuggests monitoring items for open issues worth watching
With the consumer project's workarounds docs
If the consumer project documents Claude Code quirks/workarounds (e.g. in CLAUDE.md or a
rules file), cross-reference: search checks findings against it, scan surfaces documented
issues not yet in the registry, and a resolved issue's follow-up TODO includes cleaning up the
now-obsolete workaround. Skip silently when no such doc exists.
With plan review
When stress-testing a plan that builds on a Claude Code mechanism, run search for that
mechanism first. Known bugs reshape plans cheaply before implementation.
With /bugs:write (if installed)
For bugs in YOUR code (not Claude Code itself), use the bugs plugin's skill instead;
without it, write up the defect manually. This skill covers Claude product issues only.
Spoke paths
The context/ files write this skill's directory as <skill-dir>, which is ${CLAUDE_SKILL_DIR}.
Put that path in place of the placeholder before running a command or writing it into a brief. Those
files arrive through the Read tool as plain bytes, so a ${…} token in them would reach the Bash
tool unsubstituted, and the Bash tool's environment has no CLAUDE_SKILL_DIR to expand it from.
The same table says the Bash tool's environment does not receive CLAUDE_PLUGIN_DATA, which the
registry-location rule relies on. Basis: the plugins reference,
https://code.claude.com/docs/en/plugins/manifest-reference#where-each-variable-resolves, verified
2026-10-07; recheck when that table adds supporting files to where a ${…} reference resolves, or
lists the Bash tool among the processes that receive the variables.
What This Skill Does NOT Do
- Does not read local telemetry. OTEL store, collector, hooks, ccusage → invoke
/harness-ops:observabilityvia the Skill tool - Does not fix bugs. Reports and tracks them
- Does not test features. Searches for KNOWN issues
- Issue creation requires explicit confirmation.
createaction always shows draft first, never auto-files - Does not monitor continuously. Re-run
check-allperiodically (e.g. from a recurring work item or schedule)