Skill Manifest Gen — cross-harness discovery, blanket-covered

Every aria-skill is one executable. Claude Code discovers it via SKILL.md; other harnesses need their own discovery file. This generates all three from the same sources of truth so they can't drift:

Sources of truth: SKILL.md frontmatter (name + description) + the skill's script argparse (→ the input parameter schema). Nothing is hand-authored, so a regenerate always matches the skill.

Run it

# one skill (skips if manifest/ already exists)
python3 skill-manifest-gen/gen_manifests.py --skill-dir safe-restart

# blanket: every skill in the repo (existing manifests left untouched)
python3 skill-manifest-gen/gen_manifests.py --all --repo-root .

# CI drift gate: report any skill missing a manifest or whose description is stale
python3 skill-manifest-gen/gen_manifests.py --all --check        # exit 2 if any

# deliberately regenerate a curated one
python3 skill-manifest-gen/gen_manifests.py --skill-dir foo --force

Safety / fidelity

• Never clobbers. In write mode an existing manifest/ is skipped unless --force — curated manifests are often hand-tuned past what argparse exposes (richer param docs, corrected defaults), so a blanket --all is safe to run.
• --check is the blanket enforcement. It's wired into aria-skill-test, so a new skill that ships without manifests (or one whose SKILL.md description drifts from its manifest) FAILS the regression harness. That's how "nothing goes missing again" is guaranteed, not just fixed once.
• Generated manifests are a correct starting point; refine param descriptions by hand when useful, then they're protected by the no-clobber rule.

Notes

• manifest/ is discovery glue only — the script remains the behavior truth.
• Pairs with marketplace-publish (which uploads the VS Code extension / app-store builds) and aria-skill-test (the gate).

License

Apache 2.0