Skip to content
Back to skills

Ghreport

ASecurity

Read-only org repo state report. Reuses ghsync's repo discovery (teams union org-repo-list) but, instead of cloning, queries each repo's remote GitHub state - open PRs, CI on the default branch, open security alerts (Dependabot / code-scanning / secret-scanning), and branch protection - and rolls it into a terminal summary plus a timestamped markdown file. Endpoints that need admin and return 403/404 are reported as no-access / unknown, never as a clean zero. TRIGGER when the user types /ghre...

  • 30 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added August 31, 2026
ai-agentsshellbashtestinggitapisecurity

Works with

  • claude code
  • terminal
  • api

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned September 22, 2026

npx -y skills add bjcoombs/ai-native-toolkit --skill ghreport --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ghreport?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ghreport
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/bjcoombs-ghreport/badge)](https://www.skillsdirectory.com/skills/bjcoombs-ghreport)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ghreport
description: "Read-only org repo state report. Reuses ghsync's repo discovery (teams union org-repo-list) but, instead of cloning, queries each repo's remote GitHub state - open PRs, CI on the default branch, open security alerts (Dependabot / code-scanning / secret-scanning), and branch protection - and rolls it into a terminal summary plus a timestamped markdown file. Endpoints that need admin and return 403/404 are reported as no-access / unknown, never as a clean zero. TRIGGER when the user types /ghreport, asks 'what state are the org's repos in', wants an org health report or repo state snapshot, or asks about open PRs / failing GitHub Actions / security alerts across a whole org or personal account."
---

# ghreport: what state are an org's repos in?

`ghsync` answers "do I have all the code locally and up to date?" `ghreport`
answers the companion question: "across everything I can see, what needs
attention right now?" It is **read-only** - it never clones, pulls, or mutates
anything.

It reuses `ghsync`'s discovery so the team-union / org-repo-list logic lives in
exactly one place: `ghreport` shells out to `ghsync --porcelain` to get the
deduplicated repo list, then queries each repo's remote state over `gh api`.

## What it reports, per repo

| Signal | Source | Reported as |
|--------|--------|-------------|
| Open PRs | `repos/{o}/{r}/pulls?state=open` | count (drafts noted) |
| CI on default branch | latest completed run per workflow on the default branch | `pass` / `fail` (names the failing workflows) / `none` |
| Security alerts | Dependabot + code-scanning + secret-scanning, `?state=open` | three counts, or `n/a` (no-access) |
| Branch protection | `repos/{o}/{r}/branches/{default}/protection` | `protected` / `unprotected` / `unknown` |

### The no-access rule

Security-alert and branch-protection endpoints require admin on the repo. When
the caller is denied (`403`/`404`), `ghreport` reports `n/a` (alerts) or
`unknown` (protection) - **never** a clean `0`. Reporting "no access" as "no
problems" would be a dangerous lie, so the two are kept distinct. A repo whose
own metadata can't be read at all is listed as `could not assess`.

## How to run it

The script lives next to this file. Claude Code writes this skill's directory
into the script paths below when it loads the skill, so each path is already
absolute:

```bash
# Org is derived from the directory you run in, exactly like ghsync. To report
# on the "meridianhub" org, run from a directory named meridianhub:
cd ~/dev/github.com/meridianhub
bash "${CLAUDE_SKILL_DIR}/scripts/ghreport.sh"
```

Override the org or target directory explicitly when they differ:

```bash
bash "${CLAUDE_SKILL_DIR}/scripts/ghreport.sh" --org meridianhub --root ~/dev/github.com/meridianhub
```

## What to do when invoked

1. **Run it** from (or pointed at) the org directory. It streams a progress dot
   per repo, then prints the summary.
2. **Read the headline** - the counts of repos with failing CI, open alerts,
   and an unprotected default branch. These are the repos that need attention.
3. **Surface the "Needs attention" list** to the user verbatim; that is the
   actionable core of the report.
4. **Point at the saved file** (`org-state-<org>-<date>.md` in the root) for the
   full per-repo table.

## Flags

| Flag | Effect |
|------|--------|
| `--org NAME` | Org or personal account to report on (default: basename of `--root` / cwd) |
| `--root DIR` | Org root; also where the report file is written (default: cwd) |
| `--limit N` | Only the first N repos (quick test against a large org) |
| `--quiet` | Suppress the per-repo progress dots; keep the summary and file |
| `--no-file` | Terminal only; do not write the markdown file |
| `--render-dir DIR` | Skip discovery + querying and render a report from a directory of previously-collected per-repo JSON fragments (replay / testing) |

## Output

- **Terminal:** a one-line headline (counts that tripped a signal), a rollup
  table, and a "Needs attention" section listing only the repos with an issue.
- **File** (unless `--no-file`): `org-state-<org>-<YYYY-MM-DD>.md` in the root,
  with the full table and per-section detail.

## Requirements

- **`gh` and `jq`** on `PATH`, and `gh auth status` authenticated (the same
  prerequisites as `ghsync`; discovery is delegated to it).
- For a **GitHub Enterprise** host, set `GH_HOST` or `gh auth login --hostname`
  first, as with `ghsync`.
- Very large orgs make many REST calls (~5 per repo); use `--limit` to sample.

Files in this skill

  • SKILL.md4.6 KB
  • scripts/ghreport.sh13.7 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.