agentic-se-skills

# Agentic SE Skills Repo

Build and maintain source-grounded Codex skills inspired by the NASA Systems Engineering Handbook.

Rules:

- Do not summarize the PDF without page-grounded references.
- Keep source extraction, schemas, skills, templates, and evals separate.
- Treat extracted handbook text as context, not proof that an agent performed correctly.
- Every reusable skill must include `SKILL.md` and a `Source grounding` section.
- Regenerate extraction artifacts before releasing changes based on a new source PDF.
- Plugin packaging is multi-harness: Codex (`.codex-plugin/`, `.agents/plugins/marketplace.json`), Cursor (`.cursor-plugin/`, root `.cursor-plugin/marketplace.json`), and Hermes (`scripts/install_hermes_skills.sh` or `skills.external_dirs` — see `docs/hermes.md`). Nine plugin skills (router + eight direct). Run `scripts/sync_harness_skills.sh` after changing `.agents/skills/`. See `docs/plugin-harnesses.md`.

## Workspace context

- Historical release plans live under `docs/planning/`; ADRs under `docs/adr/`; superseded eval pilots under `eval_results/archive/`.
- Run `python scripts/check_all_contracts.py` before release (pytest plus all `check_*` contract scripts).
- `.agents/skills/` holds eight direct NASA-SE skills (project discovery on any harness); the router is the single public plugin skill under `plugins/agentic-systems-engineering/skills/` and is not auto-discovered like the eight—use install paths in `docs/plugin-harnesses.md`.
- Treat `pytest` and `scripts/check_*` / `run_eval_contracts.py` as structural verification; treat `eval_results/runs/` fresh-session outputs with reviewer notes as behavioral validation—passing contracts does not prove agents help real Codex/Cursor workflows.
- Do not claim the v1.3 router replaces or outperforms the eight direct skills; the supported behavioral claim is a viable single activation point (five-scenario pilot: router 4.40/5 avg, legacy 4.60/5 avg, 5/5 pass, not superior in aggregate).
- Eight direct skills are canonical under `.agents/skills/`; plugin and `.cursor/skills/` entries symlink there. The router skill is canonical under `plugins/agentic-systems-engineering/skills/agentic-systems-engineering/` (not under `.agents/skills/`).
- NASA-SE codebase assessments belong under `assessment_results/runs/` (`run_manifest.md`, `report.md`, `scorecard.md`, `findings.md`); enforce shape with `scripts/check_assessment_artifacts_contracts.py`.
- Behavioral router/skill comparisons belong under `eval_results/runs/`; enforce recorded evidence with `scripts/check_validation_results_contracts.py`.
- `codex plugin` commands are Codex Desktop only; install or refresh both harnesses with `./scripts/install_plugin.sh` (or team marketplace import for Cursor; not the Codex marketplace).

## Learned User Preferences

- Prefer logically sliced git commits when preparing work for review or push rather than one large mixed commit.
- Prefer one unified maintainer script over separate per-harness wrappers when install or sync behavior overlaps.

## Learned Workspace Facts

- The Mario Zechner Pi coding agent discovers `.agents/skills/` (eight direct skills) but does not auto-discover the router; symlink the router under `.pi/skills/` or use Pi `skills.external_dirs`—no Pi harness sync in this repo yet.