@maestria/agent-plugins
maestria ships one directly installable plugin with native host components and seven skills. Canonical methodology lives in packages/core/agent-directives/; generated resources and all five host manifest directories live at this package root. Git/local installations and npm archives expose the same resource paths.
Component ownership
| Package folder | Contents | Canonical input |
|---|---|---|
agents/ | One reviewer profile for every Markdown host, plus a generated Codex TOML beside it | Core specialists/ and generation/profiles.json |
commands/ | Complete FEIN, SONAR, and BLITZ workflow entry points shared by every command-capable host | Core commands/, routing, and policy |
rules/, AGENTS.md, kimi.system-prompt.md | rules/global.md (OMP) and rules/cursor/ (Cursor); the root AGENTS.md serves Devin, Codex, and Hermes, and kimi.system-prompt.md is Kimi Code's systemPromptPath copy with only its own host note | Core rules.md and integration notes |
skills/ | handoff and iteration-limits utilities, plus the bundled create-pull-request, docs-update, diagnose, recon, and architecture-plan methodology skills | One core rules.md section per utility; core skills/<name>/SKILL.md for each bundled skill (scripts/sync-bundled-skills.ts) |
generation/ | Development-only profile metadata (profiles.json); not shipped | Hand-authored mechanical input |
assets/ | Logo shared by the host manifests | Hand-authored package file |
Roles, modes, dispatch, and policy are not published as root skills. Full native profiles include the role and policy required in an isolated child. Workflow commands carry their mode, routing, policy, and argument contract without loading removed skills.
Each profile's frontmatter carries every Markdown host's keys at once, and each host reads only its own: readonly (Cursor), disallowedTools: [Write, Edit] (Claude Code, Kimi Code, ZCode), and subagents: [] (Kimi). Profiles omit model, so hosts use the main conversation's model. Primary-session routing comes from each host's always-on rule, context, or companion installation, not from a dispatcher profile.
Install
On a host with a managed flow, run npx maestria install <host> (claude-code, cursor, codex, kimi-code, omp, or hermes). In Claude Code 2.1.275 or later you can also install it from the maestria marketplace in one step:
/plugin install maestria --marketplace agustinusnathaniel/maestria
Host mappings
Supports Claude Code, Cursor, Codex, Kimi Code, Devin, ZCode, OMP, and Hermes, plus portable skill clients. Follow the public installation, usage, and compatibility pages for per-host setup and support boundaries.
See installation and the public guides. OpenCode and Pi retain their native runtime packages. The standalone CLI lives in apps/maestria-cli; the methodology skills this package bundles are authored in packages/core/agent-directives/skills/, and npx skills add agustinusnathaniel/maestria serves this package's seven skills to runtimes maestria does not package.
Development and verification
Edit core methodology or generation/ metadata, then run scripts/sync-all and scripts/check-sync. The consolidated compiler composes an explicitly owned output plan from canonical inputs and validates it before mutation. It is the single writer for this package's generated resources. Generic package sync validates every discovered configuration before any write. Cleanup retires only named generated outputs and preserves unrelated files.
Run pnpm --filter @maestria/agent-plugins test for archive, assembly, native context, and Kimi resource contracts. Packed evidence is written under ignored artifacts/. The capability audit records supported, selected, and unverified surfaces; host verification records reproducible loader probes and their limits.
This package has no executable runtime adapter, hooks, scheduled automations, CLI executable, or MCP server. Some hosts support these components; they are intentionally omitted because this package supplies methodology and has no concrete executable integration requiring them.