@zigars/skills

@zigars/skills ships client-consumable skills for zigars-aware development workflows. It is separate from @zigars/mcp: the MCP package starts the local zigars server, while this package distributes agent instructions that tell clients how to use zigars effectively.

The package has no install side effects. It does not configure an MCP client, copy files into a user profile, or claim that skills are part of the base MCP protocol.

Usage

List shipped skills:

npx -y @zigars/[email protected] list

Print the package skill directory:

npx -y @zigars/[email protected] path

Print the package root for clients that load plugin or extension directories:

npx -y @zigars/[email protected] root

Print one skill directory:

npx -y @zigars/[email protected] path zigars-evidence-contract

Clients that support filesystem skills can copy or reference the printed skill directory according to that client's documentation.

Client Support

The same skills/<skill-name>/SKILL.md directories are packaged for multiple agent clients:

• Codex/OpenAI-compatible clients can consume the skill folders directly and read the optional agents/openai.yaml metadata.
• Claude Code can load the package root as a plugin. The package includes .claude-plugin/plugin.json, and Claude discovers the skills/ directory from the plugin root.
• Gemini CLI can load the package root as an extension. The package includes gemini-extension.json, and Gemini discovers the skills/ directory from the extension root.

For local testing after installing or checking out this package:

claude --plugin-dir "$(npx -y @zigars/[email protected] root)"
gemini extensions link "$(npx -y @zigars/[email protected] root)"

Those commands are examples for clients that support local plugin or extension directories. They are not run by this package automatically.

Shipped Skills

• zigars-compile-error-triage: route a failing Zig build to its primary error via structured diagnostics and confirm the fix with zigars_validate_patch.
• zigars-comptime-diagnose: localize the runtime-tainted operand behind a Zig comptime error and propose a compiler-verifiable fix.
• zigars-zon-hash-sync: repair a build.zig.zon hash mismatch in one preview-first edit, then rerun the failing build.
• zigars-io-016-migration: audit and migrate std.io, std.posix, std.fs, std.net.Address, and std.time call sites to Zig 0.16's std.Io interface.
• zigars-incremental-validation: pick the cheapest sufficient validation depth for a change and report what is left unproven.
• zigars-toolchain-pin-and-doctor: pin Zig, ZLS, and optional backends for reproducible builds on this machine.
• zigars-handoff-resume: capture or replay session context so work can resume across agents, clients, or sessions without redoing discovery.
• zigars-evidence-contract: audit zigars evidence before making final claims about safety, validation, release readiness, or backend support.
• zigars-safe-refactor: plan and validate risky Zig source changes with impact, edit-policy, patch-session, and test-selection discipline.
• zigars-dependency-steward: add, update, and audit Zig dependencies for provenance, license, SBOM, scanner evidence, and lock state (hash repair belongs to zigars-zon-hash-sync).
• zigars-zig-version-migrator: handle the non-IO parts of a Zig version bump (build.zig API, package metadata, non-IO stdlib re-paths, version-scoped docs) and route specialized concerns to their dedicated skills.
• zigars-ci-forensics: interpret CI logs, annotations, JUnit, SARIF, and matrix failures without replacing raw artifact authority.
• zigars-release-claim-auditor: keep release notes, semver decisions, backend support, and package artifact claims tied to citable evidence.
• zigars-runtime-crash-forensics: preserve and analyze Zig panic, sanitizer, core dump, LLDB, target-only, and crash-reproduction evidence.
• zigars-performance-regression-investigator: compare benchmark, coverage, profiler, flamegraph, Samply, Tracy, and performance-budget evidence.
• zigars-ffi-abi-guardian: review C interop, ABI layout, memory layout, alignment, unsafe operations, and cross-target binary assumptions.
• zigars-memory-fuzz-forensics: investigate allocator leaks, Valgrind or heaptrack output, fuzz crashes, corpus health, and minimization evidence.
• zigars-cross-target-artifact-auditor: inspect binary size, symbols, DWARF, QEMU, target runtime, and native-vs-cross-target artifact evidence.
• zigars-docs-example-steward: keep README commands, snippets, autodoc, local std/langref claims, examples, and docs drift evidence-based.

Maintainer Notes

Keep skills under skills/<skill-name>/. Each skill should contain a concise SKILL.md, optional agents/openai.yaml, and optional one-level references. Keep .claude-plugin/plugin.json and gemini-extension.json in sync with package.json when package metadata changes. Do not duplicate zigars' MCP tools in skills; route agents to the connected MCP server and keep deterministic behavior inside the server.

Validate before publishing:

npm test
npm run pack:dry

Also run the skill validator from the client skill tooling you use when it is available.