| name | gitquarry-operator |
| description | Operate gitquarry correctly for GitHub repository search, inspection, comparison, remote tree/code lookup, recipes, authentication, host selection, and script-safe output. Use when running gitquarry commands, choosing between native and discover search, selecting rank, depth, README, explain, plan/probe flags, tree/code path filters, format, progress, host, or config options, troubleshooting auth or flag conflicts, or producing operator-ready gitquarry command lines. |
| allowed-tools | Bash(gitquarry:*) |
Gitquarry Operator
Use this skill to drive gitquarry as a tool, not just to pick one benchmarked search preset.
Start Here
Load the version-matched copy from the installed CLI before choosing commands:
gitquarry skills get gitquarry
Use gitquarry agent as a shorthand for the same core guide.
Workflow
- Classify the task as
auth, search, inspect, compare, tree, code, recipe, mcp, config, or scripting.
- Verify the effective host before assuming credentials or config state.
- Start with the narrowest command that solves the task.
- Prefer native
search first. Turn on discover mode only when the request actually needs broader coverage, reranking, README evidence, or explain output.
- Prefer structured flags over stuffing GitHub qualifiers into the free-text query.
- Keep enhanced behavior explicit. Do not imply that
discover, readme, or reranking are the default path.
- Prefer
json or compact plus --progress off for scripts, CI, and agent runs.
Command Selection
- Use
gitquarry auth login|status|logout for credential work.
- Use
gitquarry search for repository discovery and ranking.
- Use
gitquarry inspect <owner/repo> when the target repository is already known.
- Use
gitquarry compare <owner/repo>... when several known repositories need side-by-side evidence.
- Use
gitquarry tree <owner/repo> when the task needs repository paths without cloning.
- Use
gitquarry code <owner/repo> <pattern> when the task needs remote code search without cloning.
- Use
gitquarry recipe run <file> when a checked-in TOML search recipe should be executed reproducibly.
- Use
gitquarry mcp only when an MCP client will launch gitquarry as a stdio server.
- Use
gitquarry config path|show when the task is about effective config state.
Tree And Code Rules
- Prefer
tree over source path when path inspection is enough and no local checkout is needed.
- Prefer
code over source path for bounded literal or regex code search inside one known repository.
- Add
--path filters, --depth, --limit, or --max-file-bytes when a broad remote scan could be noisy or API-heavy.
- Use
--reference when a branch, tag, or commit matters.
- Use
--mode regex only when regex semantics are required; literal search is the default.
Search Rules
- Start with
gitquarry search "<query>" unless the user explicitly needs enhanced discovery behavior.
- Add structured filters such as
--language, --topic, --org, --user, star ranges, or date windows before escalating to discover mode.
- Use
--mode discover only for broader candidate collection, local reranking, README-aware reranking, or explain-driven ranking analysis.
- If discover mode is used and the task does not specify a rank, remember that gitquarry itself defaults to
blended.
- If you are recommending a safer advanced preset to a human, prefer
--mode discover --depth balanced --rank quality --explain.
- Add
--readme only as an explicit second pass when evidence matters more than latency.
- Use
--plan before network execution when the task is to debug compiled query, effective mode, rank, sort, limit, or estimated request count.
- Use
--probe-path and --probe-code when search results need explicit no-clone evidence without changing result order.
- Keep probes bounded with
--probe-limit, --probe-match-limit, path filters, and --probe-max-file-bytes when scanning broad repositories.
Compare And Recipe Rules
- Use
compare only for explicit known repositories. It does not discover candidates or compute a trust score.
- Add
--tree-summary only when file-layout counts matter.
- Add
--readme only when README text is needed in the comparison output.
- Prefer
recipe run for shared, reviewed workflows in repos, docs, CI, or incident runbooks.
- Do not put credentials in recipes. Use normal host-scoped auth or token environment variables.
- Treat CLI
--host as higher precedence than recipe host.
MCP Rules
gitquarry mcp is a stdio JSON-RPC server for MCP clients; do not run it as an interactive terminal command unless you are testing JSON-RPC manually.
- Register it with clients as
gitquarry mcp, for example codex mcp add gitquarry -- "$(command -v gitquarry)" mcp.
- Use
--host before mcp for GitHub Enterprise, for example gitquarry --host https://ghe.example.com mcp.
- MCP tools return JSON command output as text content. Parse that text as JSON when structured fields matter.
- Available tools are
gitquarry_search, gitquarry_inspect, gitquarry_tree, gitquarry_code, gitquarry_compare, and gitquarry_skill.
- Credentials come from the same environment, config, and host-scoped auth paths as normal CLI commands.
Output Rules
- Use
pretty for human terminal reading.
- Use
json for structured automation.
- Use
compact for machine pipelines or logs.
- Use
csv for flat exports.
- Prefer
--progress off in non-interactive runs.
Host And Auth Rules
- Treat credentials as host-scoped.
- Use
--host when the target is GitHub Enterprise or a non-default API host.
- Prefer host-specific env vars such as
GITQUARRY_TOKEN_GITHUB_COM when scripting against multiple hosts.
- Use
GITQUARRY_CONFIG_DIR to isolate state in CI, tests, or agent runs.
- Do not assume insecure credential fallback is allowed unless
GITQUARRY_ALLOW_INSECURE_STORAGE=1 is explicitly set.
Failure Rules
- If a discover-only flag is used without
--mode discover, fix the command instead of guessing around the error.
- If a raw query qualifier conflicts with a structured flag, remove one side of the conflict.
- If
inspect input is not owner/repo, correct the repository shape before auth debugging.
- If
compare input is not owner/repo, correct the repository shape before auth debugging.
- If
recipe run fails with E_RECIPE_INVALID, fix unknown keys, malformed TOML, or invalid mapped search flags.
- If
tree or code is too broad, narrow with --path, --depth, or --limit before escalating to source path.
- If auth fails, verify token, host, and resolution order before changing search flags.
References
Read these only when needed:
references/benchmark-operator-playbook.md for the full operator playbook, command patterns, host/auth/scripting rules, and benchmark-backed discover heuristics.
../docs/commands/search.mdx when exact search flag behavior matters.
../docs/commands/inspect.mdx when the task is repository inspection rather than search.
../docs/commands/tree.mdx when the task is remote repository tree inspection.
../docs/commands/code.mdx when the task is no-clone code search.
../docs/commands/compare.mdx when the task is side-by-side explicit repository comparison.
../docs/commands/recipe.mdx when the task is checked-in search recipe execution.
../docs/commands/mcp.mdx when the task is MCP server setup.
../docs/guides/output-and-scripting.mdx when the task is CI, pipeline, or agent-safe usage.
../docs/guides/github-enterprise-hosts.mdx when the task involves non-default hosts.
When recommending commands, explain whether the operator is buying native fidelity, broader coverage, or stronger evidence, and what latency or complexity that choice adds.