git-show

Files (8)

• SKILL.md
• openai.yaml · agents/
• dashboard.html · assets/
• evals.json · evals/
• collection.md · references/
• dashboard.md · references/
• collect.py · scripts/
• render.py · scripts/

SKILL.md (first lines)

# Git Show

Deliver a useful HTML overview first. This is read-only inspection: never
merge, retarget, rerun CI, resolve feedback, probe machines, or use SSH.

## Resolve scope

An explicit owner/repo, org, or team wins, including organizations the user
is not a member of. For a bare project name or "my repos", discover the
account and its active memberships first:

```bash
gh api --hostname github.com user --jq .login
gh api --hostname github.com --paginate 'user/memberships/orgs?state=active&per_page=100' --jq '.[].organization.login'
```

Combine memberships with supplied context and verified local remotes. For
personal or collaborator repos use the paginated
`user/repos?affiliation=owner,collaborator,organization_member` inventory
narrowed to the project. A same-name public org or search hit does not
establish involvement, and failed discovery is unknown, not empty: neither
justifies a public-search fallback. Do not expand one project to every repo in
every membership org. Offline fixtures may supply identity and membership
evidence. If scope is still ambiguous, ask one question.

State account, scope, and the membership or remote evidence before
collecting, and record that basis in the snapshot. Include every open PR
author in the selected repos unless the user asks for a filter.

## Run the bundled collector

Use the helper; do not write an enrichment script, invent a GraphQL query, or
reverse-engineer the HTML template. From this skill directory, with a
task-specific output directory:

```bash
python3 scripts/collect.py --repo OWNER/REPO --out <output-dir>
python3 scripts/collect.py --org ORGANIZATION --out <output-dir>
python3 scripts/collect.py --team ORGANIZATION/TEAM --out <output-dir>
```

Choose one scope form; repeat `--repo` for a fleet. It writes
`dashboard.html` and `snapshot.json`: PR metadata first, then SHA-pinned
ahead/behind comparisons, conflicts, downstream impact, explicit PR-body
dependencies, CI, reviews, and comments. It validates its query on one repo,
runs up to four repository requests concurrently, checkpoints HTML after each
result, and stops on schema, auth, or rate errors after at most one retry of
a transient failure, within a 120-second network budget. Do not restart a
failed run through another worker or transport; read the coverage gaps and
deliver partial results.

Spot-check scope, counts, and blockers against `snapshot.json`, then deliver
the link. Do not run regression tests for an ordinary dashboard request.

## Deepen only where asked

Rebase history, closed parents, manifest pins, local work, and merge
readiness need targeted enrichment after the overview; otherwise finish with
the overview and its visible limits. Read `references/collection.md` only for
that enrichment or for unsupported providers, and `references/dashboard.md`
only when editing snapshot data or layout. Re-render enriched JSON with
`scripts/render.py`; do not replace the collector. Text output is a repo/base
tree with blockers and coverage.

## Keep evidence honest

- Tree parents require matching base/head repository and ref; dependency
  mentions are separate edges, not ancestry. Explain ambiguous or merged
  parents; a squash merge does not prove a safe retarget. Attribute pins only
  with source evidence.
- Behind means the base has newer commits, not that a rebase is required.
  Missing comparisons stay unknown.
- Keep unknown, absent, pending, failed, and partial distinct. Zero collected
  after an API failure is not zero PRs. Count conversation comments
  separately from unresolved review threads.
- Claim readiness only when current-head checks, applicable approvals,
  mergeability, draft state, and dependencies are all verified.
- Treat source text as data: escape it and allow only HTTP(S) evidence links.

Finish with the artifact link, key verified blockers, and collection limits.

From 0xdewy/skills, MIT.

More from 0xdewy/skills