restricted-tool-triage

A safety guide for working in gh-aw workflows, which are automated GitHub Actions workflows with a restricted list of allowed tools. It explains how to handle denied commands without exhausting the workflow's limited denial budget.

In plain words
What is it for?
Use it when running or debugging gh-aw workflows with narrow shell or tool permissions, especially after a permission-denied response.
Why use it?
Trying commands that the workflow does not allow can terminate the session. The guide helps identify restrictions before using tools and respond correctly to denials.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/github/gh-aw/restricted-tool-triage
Any agent
npx skills add github/gh-aw --skill restricted-tool-triage
Clone the repo
git clone --depth 1 https://github.com/github/gh-aw

Made for: Claude Code, Codex.

Per session 40 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,086 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00040 $0.01086
Opus 5 $0.00020 $0.00543
Sonnet 5 $0.00008 $0.00217
Haiku 4.5 $0.00004 $0.00109

Measured 3d ago against content hash cf843438e4c2, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

restricted-tool-triage scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.github/skills/restricted-tool-triage/SKILL.md · 49 lines

How it starts

The opening of the file, as written. The whole thing — 49 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Restricted Tool Triage

Use this skill whenever you (the agent) are executing inside a gh-aw workflow whose frontmatter declares a narrow tools: allowlist (e.g. a short bash: [...] list, a scoped MCP toolsets, or no read/shell at all) and you hit — or are at risk of hitting — a "permission denied" / tool-denial response from the harness.

Why this matters

gh-aw enforces a hard, non-renewable denial budget per session (commonly 3 denied tool calls). Once the threshold is reached, the harness emits guard.tool_denials_exceeded and aborts the entire session immediately — no further turns, no partial credit, no chance to recover. Treat every tool denial as spending down a scarce budget, not as a way to probe what's allowed.

Triggers

  • A tool call returns "permission denied by workflow tool permissions" or similar.
  • You are about to try a shell/read/write command and are unsure if it's in the declared tools: allowlist.
  • The workflow frontmatter shows a short/explicit bash: list, restrictive MCP toolsets, or omits edit/bash entirely.

Procedure

  1. Read the allowlist first, before acting. Before issuing any shell/file/MCP command, check the workflow's declared tools: block (frontmatter bash: [...], edit:, MCP toolsets:, etc.) if visible in context, or infer it from the first denial message, which echoes the exact denied command. Do not assume general-purpose shell access is available just because the environment looks like a normal shell.

  2. On the first denial, stop and pivot — do not retry variants. A denial is not a request to try a slightly different phrasing of the same disallowed command (e.g. don't go from git status to git status --short to git diff --stat as three separate attempts). Instead:

    • Identify the capability you actually need (e.g. "see which files changed").
    • Map it to a tool/command explicitly present in the allowlist (e.g. use git diff --name-only if git diff:* is allowed but git status is not; use the already-available MCP toolset instead of raw read/shell for file or repo introspection).
    • If no allowed tool can achieve the capability, stop attempting workarounds for that capability and route around it (skip the sub-task, or note the limitation in your output) rather than spending more of the denial budget.

Read the full file on GitHub · 49 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 49 lines · 40 tokens per session scan A cf843438e4c2

Subscribe to this mod's changes

restricted-tool-triage is a skill published in the GitHub repository github/gh-aw (5,084 stars, last pushed today), licensed MIT. It adds 40 tokens to every session and 1,086 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories

watch-commit

Watch a GitHub commit for CI status changes using the gh-watch extension. Use when the user wants to monitor a commit's CI checks, wait for a build to finish, or track CI progress on a specific SHA.

justincampbell/gh-watch · 48 tokens

ci-cd

CI/CD pipelines for GitHub Actions and Docker-based deployments — lint, test, build, push, deploy stages; environment secrets; branch-gated workflows; health-check rollbacks; and PM2/systemd service restarts.

LuuOW/meridian-mcp · 48 tokens

deploy

Use when ready to ship — runs pre-push gates (lint, typecheck, build, tests, security sweep), commits, releases, and pushes. Standalone, never auto-invoked. Push always requires explicit confirmation. Trigger with /hyperflow:deploy, "ship it", "ready to push", "release", "cut a release", "deploy".

jeremylongshore/tons-of-skills-marketplace · 76 tokens

argot-setup-ci

Wire argot into a repository's GitHub Actions as a non-blocking configured check on every pull request — a job summary plus code-scanning annotations. Use when the user wants argot "in CI", "on PRs", "as a GitHub Action", or asks to "set up argot CI". Distinct from argot-setup (local checking) and argot-review-pr…

get-tmonier/argot · 94 tokens

fixing

Diagnose why a PR's CI failed — compare the CI definition against the sandbox, classify the failure, and make the minimal repair. Use when a PR is red and you must work out why before changing anything.

nearform/lastlight · 46 tokens

codebase-argus

Portable Codebase Argus agent playbook for evidence-first multi-agent review of GitHub pull requests, CI failures, GitHub Actions logs, GitHub App webhook review, /argus PR comment commands, autofix branch planning, OpenAI/Claude/Gemini/Codex provider tribunals, downstream merge/rebase work, and long-lived fork sync…

AaronZ345/codebase-argus · 82 tokens