gh-fine-grained-pat

name	gh-fine-grained-pat
description	Create GitHub fine-grained personal access tokens from a declarative JSON spec by browser automation. Use when you need to mint, verify, or revoke a fine-grained PAT (release tokens, read-only auditors, CI status reporters, account-scoped tokens) — there is NO GitHub API for this, the web UI is the only path. Handles the no-expiration modal, repo picker, and permission grids that break naive automation.
allowed-tools	Read, Bash, Grep, Glob

gh-fine-grained-pat — declarative fine-grained PAT forge

GitHub exposes no API to create fine-grained PATs (community #148626) — the web UI is the only way. This skill drives that UI over the Chrome DevTools Protocol from a declarative JSON spec, so token creation is repeatable, reviewable, and anti-fragile instead of a hand-clicked one-off.

Self-Evolving Skill: This skill improves through use. GitHub's settings UI drifts — if a selector misses, the modal/picker/permission flow changed, or a permission's detail-page noun is wrong, fix scripts/, CLAUDE.md (selector map + noun map), or the spec immediately; don't defer. Only update for real, reproducible breakage.

When to use

• Mint a scoped token (release bot, CI reporter, read-only auditor, account-scoped) for storage in the SCS vault.
• Re-create / rotate a token from a checked-in spec.
• List, verify (read back settings), or revoke fine-grained tokens.

Prerequisites

• node (NOT bun — Bun's connectOverCDP times out) and playwright-core (already pinned at the repo root package.json).
• Google Chrome at the standard macOS path.
• A one-time GitHub login (persisted in a profile; see below).

One-time login (then fully automated)

A persistent Chrome profile holds the GitHub session cookie, so you log in once and every later run reuses it:

cd plugins/gh-tools/skills/gh-fine-grained-pat
node scripts/pat.mjs login      # opens Chrome; sign in (incl. 2FA); it auto-detects success
node scripts/pat.mjs doctor     # runtime + chrome + profile + auth health

The profile lives at ~/.local/share/gh-pat-automation/profile (override with GH_PAT_PROFILE_DIR). Treat it as sensitive — it can impersonate your GitHub session. It is outside the repo and never committed.

Create a token from a spec

# write the value to a 0600 file (default) — the value is NEVER printed to the terminal
node scripts/pat.mjs create specs/release-bot.json --out /tmp/.tok

# …or pipe it straight into the SCS vault (value passed in-process, never to chat)
node scripts/pat.mjs create specs/release-bot.json --vault cc-skills:gh.token

# rotate: revoke the same-named token, create a replacement, store it (sink required)
node scripts/pat.mjs rotate specs/release-bot.json --vault cc-skills:gh.token

Other verbs: list, inspect <name>, delete <name>, quit (terminate the debug Chrome by its specific PID). rotate is the safe way to roll a token in place — it revokes the old value and stores the new one in one step.

Spec format

A spec is JSON validated by schema/token-spec.schema.json (the machine-readable SSoT). Minimal shape:

{
  "name": "release-bot",
  "expiration": "none", // 7|30|60|90 | "YYYY-MM-DD" | "none"
  "repositoryAccess": { "mode": "selected", "repos": ["owner/repo"] },
  "permissions": { "repository": { "Contents": "write", "Issues": "write" } },
}

read → "Read-only", write → "Read and write". Permission keys are the exact UI labels (Contents, Pull requests, Commit statuses, Gists, …). Do not list Metadata — it is auto-required read-only.

Shipped templates (specs/)

Spec	Purpose
release-bot.json	semantic-release: Contents/Issues/PRs RW, one repo, no expiration
read-only-auditor.json	Contents read-only, single repo
ci-status-reporter.json	Commit statuses RW + Actions read
account-scoped.json	Account permission (Gists) only, no repo write
kitchen-sink.json	Superset: multiple repos, no expiration, many repo perms + account perms (exercises every UI branch)

More illustrative shapes (all-repos admin, org-owned, secrets-management) live in specs/examples/ — not auto-run by the campaign. See that folder's README before relying on one.

Security invariants

• A token value is never printed to stdout/chat — only --out <0600 file> or --vault <scope>:<dot.path>.
• Teardown kills Chrome by specific PID (lsof on the CDP port), never pkill -f (process-storm policy).
• The session-cookie profile is sensitive and lives outside the repo.

Prove it (empirical campaign)

node test/campaign.mjs          # create→verify→delete every spec + a forced-failure case, in one session

The harness namespaces test tokens zz-pat-selftest-*, auto-deletes them, asserts none leak, and confirms a real token (cc-skills-release) is untouched. Re-running needs no login (proves session reuse).

Autonomous web-auth (optional, multi-account)

GitHub's sudo mode ("Confirm access") normally needs a human gesture. The engine can clear it autonomously using a self-custodied credential — see the ADR. One-time per account:

node scripts/pat.mjs register --account <login>   # capture a passkey (virtual authenticator) + password/TOTP → gated vault
node scripts/pat.mjs patch-password --account <login>  # fix a missed password dialog (passkey kept; idempotent) [--force] [--totp]
node scripts/pat.mjs agent start                  # memory-only session agent: one Touch-ID unlock lasts the session
GH_PAT_AUTONOMOUS=1 node scripts/pat.mjs create specs/release-bot.json --account <login> --vault cc-skills:gh.token

Account is resolved from --account → the repo's origin host-alias → the spec owner → the logged-in profile. The credential is a Touch-ID-gated crown jewel stored as one blob (github-web-<login>): one tap when you go in to operate, then nothing for the session (the agent caches it in memory only). Headless/cron is intentionally unsupported (the gated tier needs that one live tap). Primary path = virtual-authenticator passkey; fallback = password + oathtool TOTP.

Troubleshooting

GitHub renames CSS classes but keeps names/roles/labels — the engine prefers role/name/label with DOM-text fallbacks and screenshots failures to /tmp/gh-pat-debug. When the UI drifts, the live selector map and the eight hard-won gotchas are documented in CLAUDE.md for fast repair.

Post-Execution Reflection

After running this skill, reflect before closing the task:

1. What broke? — A selector miss, a changed modal/picker/tab flow, or a sudo-mode stall: fix scripts/form.mjs (or browser.mjs) and update the selector map in CLAUDE.md.
2. A new permission? — Confirm its detail-page noun (create one token, read the settings page) and add it to test/campaign.mjs (NOUN) + the CLAUDE.md noun table.
3. A new token shape? — Add a specs/ (or specs/examples/) template so it's reusable next time.
4. Log it. — Append a dated line to the "Recent changes" section in CLAUDE.md. Do NOT defer — the next invocation inherits whatever you leave behind.