AgentStack
SKILL verified MIT Self-run

Onboard

skill-cookys-autopilot-onboard · by cookys

>

No reviews yet
0 installs
17 views
0.0% view→install

Install

$ agentstack add skill-cookys-autopilot-onboard

✓ scanned · ✓ verified — works with Claude Code, Cursor, and more.

Security review

✓ Passed

No issues found. Passed automated security review. · v0.1.0 How review works →

  • Prompt-injection patterns
  • Secret / credential exfiltration
  • Dangerous shell & filesystem operations
  • Untrusted network calls
  • Known-malicious package signatures

What it can access

  • Network access No
  • Filesystem access No
  • Shell / process execution No
  • Environment & secrets No
  • Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

Are you the author of Onboard? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

Onboard — calibrate autopilot to a consuming project

Turns a repo with no autopilot config into one that speaks its own build commands, coverage thresholds, doc layout, and ecosystem-standalone dispatch chains. The deterministic half is scripted (detect + scaffold); the judgment half is yours (map this project's domain to skills, derive doc⇄code drift domains, name security surfaces). The golden reference output is the hand-built ~/projects/hangar-bridge/.claude/.

> Premise: ecosystem-standalone. The scaffolded chains default to autopilot-only (the assumed > baseline is autopilot + codeforge + mnemos, no third-party plugins). superpowers is added only as an > explicit, commented fallback when the project actually has it installed. Never scaffold a superpowers-first > default.

Available Scripts (prefer over LLM judgment for the mechanical half)

| Script | Role | |--------|------| | [scripts/project-detect.js](../../scripts/project-detect.js) | Detect (pure read → JSON): package manager, commands (+lint_is_noop), per-package coverage thresholds, doc convention (doc/ vs docs/), workspace/packages, default_branch, protected-path candidates, project paths, installed plugins. Run it FIRST; it owns every mechanical fact. | | [scripts/scaffold-config.js](../../scripts/scaffold-config.js) | Scaffold (mechanical write): copies+fills the 9-file config set into /.claude/, fills detected values, writes autopilot-only chains when superpowers=false, and updates .gitignore to exclude runtime state while KEEPING *-config.md tracked. Idempotent; --force to clobber a hand-edit; --dry-run to preview. |

Workflow

1. Detect (mechanical — run the script, read the JSON)

node /scripts/project-detect.js  | tee /tmp/onboard-detect.json

Read the JSON. Sanity-check it against the repo: does package_manager, doc_dir, the coverage table, and default_branch match reality? default_branch is only emitted when the target IS the git top-level; if it is null or wrong (e.g. the repo's origin/HEAD isn't set), note the canonical branch for step 4. Do not re-derive these facts by hand — the detector owns them.

2. Scaffold (mechanical — run the script)

node /scripts/scaffold-config.js  --detect /tmp/onboard-detect.json

Preview first with --dry-run (writes nothing, prints what it WOULD do); pass --force to overwrite a hand-edited config. This writes the 9 configs + the .gitignore runtime-state block. Inspect the JSON summary ({written, skipped, gitignore_updated}). On a re-run it is idempotent (writes nothing unchanged) and skips hand-edited files unless you pass --force. The 7 mechanical configs (next, project-lifecycle, dispatch, quality-gate, dev-flow, test-strategy, qc-gate) are now correctly filled. The 2 judgment configs (skill-routing, doc-drift) are SKELETONS with TODO(onboard) markers — step 3 fills them.

3. Enrich the judgment configs (THIS is the skill's value — scripts can't do it)

Read the target repo's code + docs, then fill:

  • skill-routing.md — map THIS project's domain keywords → the right autopilot methodology skill. Read

the source/docs for the recurring domain terms (wire format, schema, auth surface, perf-sensitive paths, test tooling) and route each to autopilot:debug / autopilot:reviewer / autopilot:test-strategy / autopilot:profiling. (Golden example: hangar-bridge routes "envelope/zod schema → debug", "channel tag/ escaping/injection → reviewer", "latency/throughput → profiling".)

  • doc-drift-config.md — one ### domain block per doc⇄code mapping: which doc file(s) describe which

code dir(s), and a focus: line naming the claims that must stay true (commands, coverage numbers, API surface, status claims). Keep the deterministic gate_command + staleness_days the skeleton already has.

  • quality-gate-config.md security surfaces — if the repo has security-sensitive code (auth, crypto,

escaping, input validation, idempotency, ordering invariants), add a ## Security-critical surfaces block naming them as always-full-review (never fast-path). Read the code to find them; do not guess.

4. Reconcile CLAUDE.md (optional, if one exists and is stale)

If the target has a CLAUDE.md, align its commands / coverage thresholds / doc paths / default branch to the detected reality (step 1). If default_branch detection was null/ambiguous, write the canonical branch you confirmed. Do NOT invent conventions — only correct drift against detected facts. If there's no CLAUDE.md, skip (creating one is out of this skill's scope).

5. Seed memory pointers (optional)

If the target warrants persistent memory, note the project memory dir (~/.claude/projects//memory/) so autopilot:learn writes to one home. Add a one-line pointer to its MEMORY.md if the project has ongoing work worth tracking. (This mirrors what was hand-done for hangar-bridge.)

6. Verify

  • ls /.claude/ shows the 9 configs; /.gitignore excludes runtime state but NOT

*-config.md (they must stay tracked — git -C check-ignore .claude/dispatch-config.md returns nothing).

  • Re-run scaffold-config.js --detect /tmp/onboard-detect.json (same detect JSON as step 2) →

summary written:[] (idempotent; your step-3/4 hand edits are preserved, reported in skipped).

  • The dispatch chains are autopilot-only (no superpowers-first default) unless the project genuinely has

superpowers installed.

Acceptance (what "done" looks like)

The /.claude/ set reproduces the mechanical values of the golden hangar-bridge output (same commands, coverage table, doc paths, protected paths, autopilot-only chains) AND carries the project-specific judgment content you added in step 3. The scaffold is re-runnable without churn.

> Note (first run after install): a newly-shipped onboard skill is not dispatchable until Claude Code > restarts (the plugin caches skills at session start). The scripts (project-detect.js / scaffold-config.js) > work immediately from the CLI regardless.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet — be the first.

Versions

  • v0.1.0 Imported from the upstream source.