Github Agentic Workflows

GitHub Agentic Workflows

Procedures

Step 1: Identify the repository state

1. Inspect the workspace for .github/workflows/, .github/agents/, existing .lock.yml files, and any gh aw usage.
2. Execute node skills/github-agentic-workflows/scripts/find-gh-aw-targets.mjs . when a Node runtime is available.
3. Run gh aw version before making compiler-sensitive decisions so the workflow authoring path matches the installed CLI behavior.
4. If the repository contains multiple candidate workflows, prefer the workflow the user named or the one closest to the active issue, pull request, or automation surface.
5. If the repository has no GH-AW setup and the task is to create or maintain agentic workflows, read references/authoring.md before editing.
6. If the task is limited to standard deterministic GitHub Actions YAML without agentic markdown workflows, stop and explain that this skill does not apply.

Step 2: Choose the working mode

1. Classify the task as one of: repository setup, new workflow authoring, workflow revision, workflow installation from another repository, security review, or failure debugging.
2. Read references/examples.md when the task needs a starting pattern for scheduled reports, issue or PR triage, orchestration, or agent handoff.
3. Read references/security-and-operations.md when the workflow needs safe outputs, network policy, authentication, lockdown, threat detection, or run observability.
4. Read references/troubleshooting.md when the workflow fails to compile, install, authenticate, execute safe outputs, or access tools.

Step 3: Author or revise the workflow source

1. Keep the workflow source of truth in .github/workflows/<workflow-name>.md.
2. Use assets/workflow.template.md as the base shape when creating a new workflow.
3. Choose the smallest viable trigger surface and repository role scope that satisfies the task.
4. Keep permissions: read-only unless the workflow truly needs broader GitHub Actions permissions outside the agentic section.
5. Prefer safe-outputs: for comments, issues, labels, PRs, agent assignment, and orchestration instead of granting direct write access to the agent.
6. In safe-output workflows, instruct the agent to call noop when no action is required.
7. Keep tools: and toolsets: minimal and specific to the task.
8. Default to engine: copilot unless the task explicitly requires another engine and the repository is already prepared for that engine's authentication model. Use engine: pi for experimental multi-provider routing or copilot-sdk: true for Copilot SDK mode. Third-party engines (cursor, kiro, aider, crush, opencode) are importable via gh aw add.
9. Configure network: with least privilege. Prefer ecosystem identifiers such as node, python, or github over individual registry domains when the compiler supports them.
10. Set max-ai-credits: when the workflow runs on a schedule or could trigger large agent turns, to cap per-run AI Credits (AIC) spend before it drifts upward.
11. If strict mode and the installed CLI reject custom domains that the workflow still needs, prefetch external sources in deterministic setup steps and pass local files into the agent instead of broadly relaxing the firewall.
12. Do not rely on ${{ steps.<id>.outputs.* }} placeholders reaching the agent-visible markdown body in real runs. If prompt instructions depend on runtime values, write them into a deterministic local file during setup and tell the agent to read that file.
13. Use imported or reusable workflows only when the repository genuinely benefits from shared logic or orchestration.
14. For recurring work across a dynamic set of inputs, prefer a reusable GH-AW worker plus a deterministic YAML wrapper for discovery and matrix fan-out.
15. When a reusable GH-AW worker is called from a matrix, do not leave it on the default shared workflow-level concurrency group. Set an explicit concurrency group keyed by the matrix input or prompt identity so parallel legs are not cancelled by GitHub's one-running-one-pending concurrency behavior.
16. Recompile the workflow after frontmatter, imports, or other compile-time configuration changes.
17. If only the markdown body changed and the workflow is edited directly on GitHub.com, do not recompile solely for body text changes.
18. Treat .github/aw/ as transient GH-AW runtime and compiler scratch space during local compile, validate, or trial flows unless the workflow intentionally uses checked-in files from that path.

Step 4: Configure repository prerequisites and authentication

1. Read references/authoring.md before first-time repository setup.
2. Run gh aw init when the repository is not configured for GH-AW authoring and the user wants persistent setup.
3. Configure engine secrets with gh aw secrets bootstrap or gh aw secrets set.
4. Use COPILOT_GITHUB_TOKEN or the copilot-requests Actions permission for Copilot engine authentication.
5. For Copilot runs, use a fine-grained PAT in COPILOT_GITHUB_TOKEN; a general gho_... OAuth token may pass secret checks but still fail real Copilot execution.
6. When using gh aw trial, verify that the host repository also has the required engine secret. Secrets from the logical or source repository are not copied automatically into the temporary or reusable host repo.
7. Re-check gh aw version after extension upgrades or reinstall paths so the repository guidance and compiler behavior stay aligned.
8. If a deterministic wrapper calls a reusable worker, make the caller workflow permissions at least as broad as the nested worker's requested actions, contents, and pull-requests scopes or the run can fail before agent execution starts.
9. Use a GitHub App or custom GitHub token when the workflow needs cross-repository reads or writes, Projects access, remote GitHub tool mode, or advanced safe outputs.
10. If the repository is public and the workflow will inspect untrusted external content, preserve lockdown and approval controls unless the task is explicitly a low-risk public workflow.

Step 5: Validate, compile, and execute

1. Run gh aw fix --write when the workflow uses deprecated fields or the compiler points to codemod-able drift.
2. Run gh aw validate --strict before treating a workflow as ready.
3. When the CLI version changed or a workflow depends on newer frontmatter behavior, run gh aw compile --verbose so warnings and generated-version changes are visible.
4. Run gh aw compile after validation succeeds and commit both the .md source and the generated .lock.yml file.
5. Use gh aw run <workflow> --dry-run to validate remote dispatch behavior, especially when you suspect branch, trigger, or workflow-discovery problems.
6. Use gh aw trial ./.github/workflows/<workflow>.md when you need to execute a local workflow source before pushing it. The explicit ./ path matters; without it, the CLI may parse the argument as a repository spec instead of a local file.
7. Use gh aw run <workflow> for direct execution only after the workflow source or lockfile actually exists on the target branch.
8. Use gh aw status, gh aw logs, gh aw audit, and gh aw health to review state, failures, cost, tool usage, and success trends after changes.

Step 6: Operate and improve professionally

1. Review whether the workflow can be simplified, split into smaller workflows, or converted into shared components.
2. Prefer staged safe outputs for initial rollout of workflows that create issues, comments, or pull requests at scale.
3. For recurring maintenance, pin or review the installed gh-aw version deliberately and update workflows through gh aw update or gh aw upgrade instead of ad hoc copy-paste.
4. Keep workflow prompts specific, bounded, and auditable. Move broad strategy into reusable imports or agent files when multiple workflows need the same rules.

Error Handling

• If gh extension install github/gh-aw fails, use the standalone installer path documented in references/troubleshooting.md.
• If compilation fails, fix frontmatter syntax, deprecated fields, imports, or permission mismatches before continuing.
• If compiler behavior does not match the docs you are reading, trust the installed gh aw version and validate against that version before rewriting the workflow shape.
• If gh aw trial rejects .github/workflows/<name>.md as an invalid repository spec, retry with an explicit local path such as ./.github/workflows/<name>.md.
• If a reusable worker succeeds in isolation but the wrapper run fails at startup, inspect caller-workflow permissions: inheritance before changing the worker logic.
• If a matrix of reusable GH-AW workers cancels most legs immediately, inspect the called workflow's concurrency group first. A shared workflow-level group will cancel pending legs even when cancel-in-progress is false.
• If a trial run fails on the activation step with Validate COPILOT_GITHUB_TOKEN secret, inspect the host repository with gh secret list -R <host-repo>. The source repository's secrets are not inherited by the trial host.
• If gh aw trial --force-delete-host-repo-before fails, confirm you have admin rights on the host repository and that the current GitHub auth token includes delete_repo scope.
• If the workflow prompt still shows unresolved runtime placeholders during execution, move those values into a generated local context file and have the agent read that file explicitly.
• If local gh aw compile, gh aw validate, or gh aw trial commands create .github/aw/ files such as actions-lock.json or logs, treat them as transient byproducts and remove them before commit unless the repository intentionally keeps them.
• If safe outputs do nothing, verify that staged mode is intentional and that the prompt explicitly instructs the agent to call noop when no write action is needed.
• If gh aw mcp inspect fails on a compiled scheduled workflow source with a fuzzy schedule parsing error, treat that as an inspection-path limitation first. Re-run gh aw compile, prefer installed-workflow or run-log based debugging, and do not assume the workflow itself is invalid if validation and compile already passed.
• If the compiler rejects engine fields or tool entries that look valid from older examples, trust the installed schema. In gh aw v0.58.3, engine.max-turns is not supported for Copilot, bash must be true, false, or an allowlist, and bare edit: / web-fetch: keys are accepted where boolean values are not.
• If URLs are sanitized as (redacted) or tools cannot reach required services, tighten and expand network.allowed deliberately rather than disabling the firewall.
• If Copilot inference fails with a configured token, verify that the PAT owner actually has Copilot license and inference access.
• If public-repository workflows miss external contributor content, confirm whether GitHub lockdown mode is blocking that content before changing workflow logic.