2.2 KiB
name, description
| name | description |
|---|---|
| agentic-workflows | Route requests to design, create, update, debug, or upgrade GitHub Agentic Workflows (gh-aw), including workflow cost audits. Use for workflow infrastructure, not ordinary product development or code review. |
Agentic Workflows Router
Choose one primary route below. Fetch its reference from
https://raw.githubusercontent.com/github/gh-aw/v1.0.1/.github/aw/.
This version matches the repository's workflow lock and Copilot setup. Before
using another version, check the target lock's compiler_version; upgrades must
align the compiler, setup, and references together.
| Task | Reference |
|---|---|
| Design from requirements | designer.md |
| Create a workflow | create-agentic-workflow.md |
| Update a workflow | update-agentic-workflow.md |
| Debug or audit a run | debug-agentic-workflow.md |
| Upgrade or fix deprecations | upgrade-agentic-workflows.md |
| Reduce or measure token cost | token-optimization.md |
| Configure an engine | configure-agentic-engine.md |
| Create shared components or MCP wrappers | create-shared-agentic-workflow.md |
| Choose architecture | patterns.md |
| CLI commands or MCP equivalents | cli-commands.md |
Read the matched guide, then only the supporting sections needed for the task.
Do not preload the reference catalog or full manuals. Add report.md for report
design, test-coverage.md for coverage data, dependabot.md for generated
manifest updates, or skills.md for requested skill/plugin integration. For
OTEL/OTLP analysis, use upstream skills/otel-queries/SKILL.md at the same tag.
If .github/aw/instructions.md exists locally, read it as the repository overlay.
User scope and repository instructions take precedence over upstream defaults.
Edit workflow .md sources and run gh aw compile <workflow> with the matching
compiler; include the generated lock in the change. Never edit locks directly.
Keep permissions and data reads scoped to the workflow's purpose. A request to
edit or audit a workflow does not authorize dispatching runs or posting to GitHub.
For cost audits, distinguish prompt size from measured run usage, including cache
reads/writes; a shorter prompt alone does not prove lower credits or better review.