本文へ移動
cccskills
無料GitHub で公開

assess-github-repo-security

Read-only audit of a GitHub repository's security posture. Gathers ref protection (rulesets AND classic branch protection), Actions token permissions, code and supply-chain features (Dependabot, secret scanning, push protection, CodeQL), and repo hygiene toggles via `gh api`, then classifies findings against essential / recommended / advanced tiers into a PASS/GAP report. Makes NO changes. Use when reviewing a repo before open-sourcing or a release, auditing a public user-owned repo whose CI auto-commits to the default branch, verifying a hardening change actually took effect, or producing a baseline security posture report for a repository.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md17.7 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Assess GitHub Repository Security

Audit a GitHub repository's security posture, read-only, and classify it against tiered best practices. This skill only reads — it never enables, disables, or changes any setting. Remediation is a separate concern (see harden-github-repo-security).

When to Use

  • Reviewing a repo before open-sourcing it or cutting a release
  • Auditing a public, user-owned repo whose CI auto-commits to the default branch (e.g. via stefanzweifel/git-auto-commit-action + the default GITHUB_TOKEN)
  • Verifying that a hardening change actually took effect
  • Producing a baseline posture report to track over time

Inputs

  • Required: Target repository as OWNER/REPO
  • Required: gh CLI authenticated with an account that has the admin repository role on the target repo (some endpoints 403 otherwise). This is a role requirement, not a need for a broad write-capable token: a fine-grained, single-repo PAT scoped to read-only Administration / Actions / Secret-scanning permissions satisfies the same admin check with far less blast radius if the audit credential is ever leaked or reused in automation
  • Optional: Branch to check for classic protection (defaults to the repo's default_branch)
  • Optional: security_events token scope — needed only to count open Dependabot alerts; the rest works without it

Set a shorthand for every command below:

R="OWNER/REPO"   # e.g. pjt222/agent-almanac

Procedure

Step 1: Preflight — auth and admin

Confirm the CLI is authenticated and the account has admin on the repo. Non-admins get partial data (403 on rulesets, Actions, and security endpoints), which silently understates the posture.

gh auth status
gh api "repos/$R" --jq '{full_name, admin: .permissions.admin, visibility}'

Expected: gh auth status shows a logged-in account; the second call prints admin: true.

On failure: If not authenticated, run gh auth login (or set GH_TOKEN). If admin: false, stop and note in the report that findings are incomplete — a non-admin cannot read ruleset internals or Actions permissions, so absence of a control cannot be distinguished from lack of access.

Step 2: Repo basics — visibility, merge toggles, collaborators

gh api "repos/$R" --jq '{visibility, private, allow_forking,
  default_branch, allow_merge_commit, allow_squash_merge,
  allow_rebase_merge, delete_branch_on_merge}'

# Direct collaborators and their effective role
gh api "repos/$R/collaborators?affiliation=direct" \
  --jq '.[] | {login, role_name}'

Expected: JSON for the repo toggles plus one line per direct collaborator with role_name (admin/maintain/write/triage/read).

On failure: A 403 on collaborators means the token lacks admin; note it and continue. Record visibility — it changes which features are free: on public repos secret scanning, push protection, CodeQL, and dependency review are free; the paid GHAS/Secret Protection products are private/org-only, so their absence on a public repo is not a gap.

Step 3: Ref protection — check rulesets AND classic (both)

A repo can be protected by a ruleset, by classic branch protection, by both, or by neither. Rulesets are the actively-developed path and are free on public user-owned repos. You MUST check both — a repo with only a ruleset returns 404 on the classic endpoint and vice versa.

b=$(gh api "repos/$R" --jq '.default_branch')

# (a) Rulesets — list, then inspect each active one's rules + bypass list
gh api "repos/$R/rulesets" --jq '.[] | {id, name, enforcement, target}'
# For each id from above:
#   gh api "repos/$R/rulesets/<id>" \
#     --jq '{name, enforcement,
#            rules: [.rules[].type],
#            bypass: [.bypass_actors[] | {actor_type, bypass_mode}]}'

# (b) Classic branch protection on the default branch (404 = none)
gh api -i "repos/$R/branches/$b/protection" 2>/dev/null | head -1
gh api "repos/$R/branches/$b/protection" \
  --jq '{required_status_checks: .required_status_checks.checks,
         strict: .required_status_checks.strict,
         enforce_admins: .enforce_admins.enabled,
         required_reviews: .required_pull_request_reviews.required_approving_review_count,
         linear: .required_linear_history.enabled,
         signatures: .required_signatures.enabled,
         allow_force_pushes: .allow_force_pushes.enabled,
         allow_deletions: .allow_deletions.enabled}' 2>/dev/null \
  || echo "no classic branch protection"

Aggregate ALL rulesets, not just the first. Several rulesets can target the same branch at once; GitHub applies them together (most-restrictive-wins), and a common pattern splits ref protection (deletion/non_fast_forward, no bypass) and a required check (required_status_checks, with bypass actors) into separate rulesets. The branch's true posture is the union of every active ruleset plus classic protection — inspect each id from step (a); never conclude from one ruleset alone. A rule absent from ruleset X may be enforced by ruleset Y.

Expected: Either a ruleset with rules such as deletion, non_fast_forward, pull_request, required_status_checks, or a classic protection object, or an explicit "none" from both.

On failure: A 404 on either endpoint means that mechanism is not configured — it is a valid data point, not an error; record "none" for that mechanism. Note the rule set from the ruleset rules[].type list and the bypass_actors. Remember: the default GITHUB_TOKEN / github-actions[bot] can never be a bypass actor by design — if the repo auto-commits AND has required_status_checks/pull_request rules with no Integration (GitHub App) or DeployKey bypass actor, flag that the bot push must be 403ing (a real finding).

Step 4: Actions security — token defaults and PR approval

# Actions enablement + allowed-actions policy (+ SHA-pinning policy)
gh api "repos/$R/actions/permissions" \
  --jq '{enabled, allowed_actions, sha_pinning_required}'

# Default GITHUB_TOKEN permissions + can-the-bot-approve-PRs
gh api "repos/$R/actions/permissions/workflow" \
  --jq '{default_workflow_permissions, can_approve_pull_request_reviews}'

Expected: default_workflow_permissions is read (hardened) or write (permissive default), and can_approve_pull_request_reviews is false (hardened) or true (a review-bypass vector).

On failure: 403 means non-admin — note it. If allowed_actions is all, any action can run (looser); selected with a pinned allow-list is tighter — but note that selected + github_owned_allowed still blocks non-GitHub actions like stefanzweifel/git-auto-commit-action unless explicitly allow-listed. sha_pinning_required may be absent/null on repos that never set the 2025-08 policy — record as "not enforced". Per-workflow permissions: blocks live in the YAML, not the API; note that the API shows only the repo default, which is a default (not a cap) for same-repo push events.

Step 5: Code and supply-chain features

# Dependabot alerts + dependency graph (204 = ON, 404 = OFF)
gh api -i "repos/$R/vulnerability-alerts" 2>/dev/null | head -1

# Dependabot security updates (auto fix PRs) — separate toggle
gh api "repos/$R/automated-security-fixes" \
  --jq '{enabled, paused}' 2>/dev/null || echo "automated-security-fixes: off/unavailable"

# Secret scanning + push protection status (security_and_analysis has NO
# code-scanning field — CodeQL is probed separately below)
gh api "repos/$R" --jq '.security_and_analysis'

# CodeQL code scanning — default setup state (its own endpoint)
gh api "repos/$R/code-scanning/default-setup" --jq '.state' 2>/dev/null \
  || echo "code-scanning default setup: not configured / no access"

# Dependabot version updates config present?
gh api -i "repos/$R/contents/.github/dependabot.yml" 2>/dev/null | head -1

# SECURITY.md — GitHub auto-detects it at root, docs/, OR .github/.
# Present if ANY of the three returns 200; only a GAP if all three 404.
for p in SECURITY.md docs/SECURITY.md .github/SECURITY.md; do
  echo "$p: $(gh api -i "repos/$R/contents/$p" 2>/dev/null | head -1)"
done

# Private vulnerability reporting — status is a 200 body {"enabled": bool},
# NOT a 204/404 toggle, so read the field directly.
gh api "repos/$R/private-vulnerability-reporting" --jq '.enabled' 2>/dev/null \
  || echo "PVR: not accessible"

# Open Dependabot alert count — needs security_events scope; may 403
gh api "repos/$R/dependabot/alerts?state=open" --jq 'length' 2>/dev/null \
  || echo "dependabot/alerts: 403 (needs security_events scope) — skipped"

Expected: vulnerability-alerts → HTTP/2.0 204 when enabled; security_and_analysis shows secret_scanning.status and secret_scanning_push_protection.status as enabled on a hardened public repo; code-scanning/default-setup .state is configured when CodeQL default setup is on; automated-security-fixes → enabled: true; the dependabot.yml probe returns 200 when present; at least one of the three SECURITY.md paths returns 200 when a policy file exists; private-vulnerability-reporting .enabled prints true when PVR is on.

On failure: vulnerability-alerts returning 404 = Dependabot alerts OFF (a gap). Remember alerts != fixes: alerts on with automated-security-fixes off means nothing is auto-remediated — flag both separately. On public repos secret_scanning is usually already enabled by default; if security_and_analysis omits the field entirely, treat it as advisory (public-repo default on) but note the API did not confirm it. code-scanning/default-setup 404/non-configured = CodeQL default setup not enabled (a recommended-tier gap on a public repo, where code scanning is free). For SECURITY.md, record a GAP only if all three paths 404 — a 200 on root, docs/, or .github/ all count as present. private-vulnerability-reporting empty/false = PVR off; a non-JSON or error response = not accessible. dependabot/alerts 403 is expected without security_events scope — record "not assessed", not "0".

Step 6: Produce the tiered PASS/GAP report

Classify every gathered fact into three tiers and mark PASS (control present), GAP (control absent), or N/A (not applicable, e.g. a paid private-repo feature on a public repo). Distinguish a required check from an advisory one: a status check that runs but is not listed in the ruleset's required_status_checks does not gate anything — it is advisory only.

Interpret the solo-maintainer lockout. Cross-reference the Step 2 direct-collaborator count with the Step 3 required_approving_review_count. On a single-maintainer repo, required_approving_review_count >= 1 (and require_code_owner_review with a sole owner) is unsatisfiable — you cannot approve your own PR — so on an auto-commit repo it is a self-lockout / functional breakage, not a PASS. Flag it as a GAP-with-caveat, and note that a ruleset (unlike classic branch protection's enforce_admins) does not auto-exempt the admin: the sole maintainer stays blocked unless explicitly added to the ruleset bypass_actors.

Do not publish the raw report into the audited repo. The PASS/GAP list enumerates exact, admin-only-visible gaps (no ruleset, GITHUB_TOKEN can approve PRs, no push protection) — a pre-remediation vulnerability roadmap. While any GAP is open, keep the findings local/private (a gitignored file, a private gist, or a draft GitHub Security Advisory); never commit it into the public repo being audited (contrast security-audit-codebase, which writes SECURITY_AUDIT_REPORT.md into the project root — do NOT copy that pattern here). Redact ruleset / App ID / bypass-actor identifiers before any public write.

# GitHub Repo Security Assessment — OWNER/REPO
Date: YYYY-MM-DD   Visibility: public   Auditor role: admin

## Essential
- [PASS] Ruleset with deletion + non_fast_forward on default branch
- [GAP]  default_workflow_permissions = write (should be read)
- [PASS] can_approve_pull_request_reviews = false
- [GAP]  actions/checkout pinned to @v4 tag, not a 40-char SHA
- [PASS] Dependabot alerts (204) + security updates (enabled)
- [GAP]  No .github/dependabot.yml (version updates off — keeps
         github-actions action pins fresh; essential supply-chain hygiene)
- [PASS] Secret scanning + push protection enabled

## Recommended
- [N/A]  required_status_checks — none (no CI gate configured)
- [PASS] CodeQL default setup: configured
- [PASS] SECURITY.md present (.github/) + PVR enabled

## Advanced
- [GAP]  sha_pinning_required: not enforced
- [N/A]  Required signed commits — would block the auto-commit bot
- [N/A]  Paid GHAS / Secret Protection — public repo, features free

## Notes
- Auto-commit bot: ruleset has NO App/DeployKey bypass actor; if
  required_status_checks were added the github-actions[bot] push would 403.
- Solo maintainer (1 direct collaborator): required_approving_review_count
  is 0 — correct; any value >= 1 would be an unsatisfiable self-lockout and
  a ruleset would NOT auto-exempt the admin.
- Dependabot open alerts: not assessed (token lacks security_events).
- This report is kept local/private — not committed into the audited repo
  while GAPs remain open.

Expected: A written report with every fact placed in a tier and marked PASS / GAP / N/A, plus a Notes section for auth gaps, advisory-only checks, and the auto-commit-bot bypass interaction.

On failure: If some endpoints 403'd, still emit the report but mark those rows "not assessed" and state the missing scope/role at the top — never record an unreadable control as a GAP (absence of access is not absence of the control).

Validation

  • gh auth status confirmed and admin on the repo verified (or the report is explicitly flagged incomplete)
  • visibility recorded (drives which features are free vs N/A)
  • BOTH rulesets AND classic branch protection were queried (not just one), and EVERY active ruleset was inspected + aggregated (not just the first)
  • Actions permissions and permissions/workflow both read
  • Dependabot alerts, security updates, secret scanning, and push protection each checked as separate toggles
  • CodeQL default setup queried via code-scanning/default-setup (not inferred from security_and_analysis, which has no such field)
  • SECURITY.md checked at all three auto-detected paths (root, docs/, .github/) — GAP only if all three 404
  • Private vulnerability reporting read from the .enabled body field (not a 204/404 toggle)
  • Solo-maintainer lockout interpreted: collaborator count cross-referenced with required_approving_review_count
  • Every finding is tiered and marked PASS / GAP / N/A
  • Report kept local/private while GAPs are open — NOT committed into the audited public repo
  • No setting was changed — this audit is strictly read-only

Common Pitfalls

  • Checking only rulesets or only classic protection: They are independent mechanisms. A repo protected by a ruleset returns 404 on branches/{b}/protection; concluding "unprotected" from that alone is wrong. Always query both.
  • Some endpoints 403 without security scope/admin: dependabot/alerts needs security_events; rulesets and Actions permissions need admin. A 403 is "not assessed", not a GAP — recording it as a gap fabricates a finding.
  • Misreading the auto-commit-bot bypass: The default GITHUB_TOKEN / github-actions[bot] can never be a ruleset bypass actor. If a repo auto-commits to a branch that has required_status_checks or a pull_request rule with no Integration/DeployKey bypass actor, the bot push is being rejected — surface it as a real finding.
  • Reading the API default as a hard cap: default_workflow_permissions is the repo default; per-job permissions: in the workflow YAML can raise or drop it for same-repo events. The API cannot show the effective per-workflow grant — note that limitation.
  • Inferring CodeQL from security_and_analysis: that object carries secret_scanning*, dependabot_security_updates, and advanced_security — but no code-scanning field. CodeQL default-setup state comes only from the code-scanning/default-setup endpoint; never claim a CodeQL PASS/GAP the skill did not actually probe.
  • Marking a solo review requirement as PASS: on a one-maintainer repo, required_approving_review_count >= 1 is an unsatisfiable self-lockout, not a protective control — and a ruleset (unlike classic enforce_admins) will not auto-exempt the admin. Report it as a functional-breakage GAP.

Related Skills

  • harden-github-repo-security - apply the fixes this audit surfaces
  • security-audit-codebase - complementary in-tree secret/dependency scan
  • configure-git-repository - foundational repo + .gitignore setup

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Execute structural metamorphosis using strangler fig migration, chrysalis phases, and interface preservation. Covers transformation planning, parallel running, progressive cutover, rollback design, and post-metamorphosis stabilization for system architecture evolution. Use when assess-form has classified the system as READY for transformation, when migrating from monolith to microservices, when replacing a core subsystem while dependents continue operating, or when any architectural change must be gradual rather than big-bang.

日本語の概要は準備中です。原文の説明を表示しています。

pjt222/agent-almanac372026年10月10日 更新

adaptic

無料

Master skill composing the 5-step synoptic cycle for panoramic synthesis across multiple domains. Orchestrates meditate, expand-awareness, observe, awareness, integrate-gestalt, and express-insight into a coherent process that produces unified understanding rather than sequential compromise. Use when a problem genuinely spans 3+ domains and the interactions between domains matter more than depth in any one, when sequential analysis feels like compromise rather than integration, or before major architectural decisions affecting multiple stakeholders.

日本語の概要は準備中です。原文の説明を表示しています。

pjt222/agent-almanac372026年10月10日 更新

Scaffold a new puzzle type across all 10+ pipeline integration points in jigsawR. Creates the core puzzle module, wires it into the unified pipeline (generation, positioning, rendering, adjacency), adds ggpuzzle geom/stat layers, updates DESCRIPTION and config.yml, extends the Shiny app, and creates a comprehensive test suite. Use when adding a completely new puzzle type to the package or following the 10-point integration checklist to ensure nothing is missed end-to-end.

日本語の概要は準備中です。原文の説明を表示しています。

pjt222/agent-almanac372026年10月10日 更新

Add Rcpp or RcppArmadillo integration to an R package for high-performance C++ code. Covers setup, writing C++ functions, RcppExports generation, testing compiled code, and debugging. Use when an R function is too slow and profiling confirms a bottleneck, when you need to interface with existing C/C++ libraries, or when implementing algorithms (loops, recursion, linear algebra) that benefit from compiled code.

日本語の概要は準備中です。原文の説明を表示しています。

pjt222/agent-almanac372026年10月10日 更新

aikido

無料

Practice aikido as a defensive martial art emphasizing harmony, redirection, and controlled resolution. Covers centering and grounding, ukemi (safe falling and rolling), core techniques (irimi-nage, shiho-nage, kote-gaeshi, ikkyo), blending and tai sabaki (body movement), randori (multiple attacker practice), weapons awareness, and applying principles off the mat. Use when learning a defensive martial art that prioritizes de-escalation, developing the ability to redirect aggression without causing unnecessary harm, building safe falling skills, or cultivating calm centeredness under physical pressure.

日本語の概要は準備中です。原文の説明を表示しています。

pjt222/agent-almanac372026年10月10日 更新

Analyze an arbitrary codebase to identify functions, APIs, and data sources suitable for exposure as MCP tools, producing a tool specification document. Use when planning an MCP server for an existing project, auditing a codebase before wrapping it as an AI-accessible tool surface, comparing what a codebase can do versus what is already exposed via MCP, or generating a tool spec to hand off to scaffold-mcp-server.

日本語の概要は準備中です。原文の説明を表示しています。

pjt222/agent-almanac372026年10月10日 更新

pjt222 のスキルをすべて見る

このスキルの問題を報告する