| name | workflow |
| description | Pull in the standard Milo dev loop (research → plan → implement → run → review → wrap-up) at the start of a task. Use when beginning any non-trivial change so gates and definition-of-done aren't skipped. |
Workflow
Load and follow AGENT_WORKFLOW.md (repo root) for this task. It is the canonical loop; this skill just guarantees it's pulled in and enforced.
Do this now
- Read
AGENTS.md to route to the docs/skills for what you're touching.
- Open a worksheet — copy
docs/worksheets/TEMPLATE.md to docs/worksheets/<date>-<slug>.md. Record goal + plan there and keep it live (mandatory if running autonomously).
- Work the loop in
AGENT_WORKFLOW.md:
- Research → understand the real code path; fix any doc that lied. Review gate:
scripts/agent_review.sh research for non-trivial approaches.
- Plan → write it in the worksheet, terse, end with open questions. Gate:
scripts/agent_review.sh plan.
- Implement → match
CONVENTIONS.md; camelCase milo; feature = checker+lower+codegen+formatter+LSP; write targeted tests as you go.
- Run (mandatory) → compile+run real fixtures/examples;
bun test; /verify. Never claim done unrun.
- Review →
scripts/agent_review.sh implementation (different model, personas) + /code-review.
- Wrap-up → full
bun test + benchmarks; update stale docs (bump last-verified); commit worksheet + feedback with the code; tag git tag ws/<slug>.
The one rule
Do not tell the user something works until you have run it and seen it work. Report failures with their output; name any skipped step.