launching-workflow-routines

Define the routine as a Markdown file with frontmatter and a Nodes section in a routines/ registry, then ask the agent to run it by name. The launcher validates the definition, shows the run plan, gets launch approval, and executes each node sequentially with verified artifacts.

Run scripts/validate_routine.py on the routine file with python3. It checks frontmatter fields, node schema, tiers, gates, input references, output uniqueness, and recursion rules, printing error lines that block launch and warnings that do not.

Dry-run mode resolves and validates the routine, then shows the full run plan with nodes, tiers, gates, and inputs without writing anything. Run mode additionally asks for Gate 0 approval, creates a timestamped run directory, executes every node, and writes a verified run report.

No. The validator rejects any node whose executor is the launcher skill itself, ends in .md, or matches a sibling routine filename, and the runtime re-checks this at dispatch. Executor skills that are themselves multi-stage runners are allowed because they are skills, not routines.

The node's on_fail policy applies: stop marks all remaining nodes skipped and jumps to the final report, while continue records the failure and proceeds. There are no automatic retries; the only re-execution is one human-approved redo at an after-gate.

No. It stays vendor-neutral by using tier labels (small, mid, frontier) instead of model names, and declares compatibility with claude-code, codex, copilot, gemini, and antigravity. Nodes dispatch to sub-agents when the host supports them, otherwise inline.