dancing-bear: Skill for Claude Code

Workflow — DAG Execution Engine

Parse a workflow YAML definition, compile it into a parallel execution plan, walk each group by spawning the appropriate agent type, handle human gates, run validation passes, and report results.

When to Use

• User says "run workflow X", "execute workflow X", or references a file in workflows/
• User wants to preview a workflow execution plan (--dry-run)
• User wants to resume a halted workflow from a human gate

Usage

/workflow workflows/test/coverage-report.yaml
/workflow workflows/test/coverage-report.yaml --execute
/workflow workflows/test/coverage-report.yaml --workspace out/my-run
/workflow workflows/test/coverage-improve.yaml --params min_coverage=80

Input Parsing

Extract these arguments from the user's request or skill args:

Argument	Required	Default	Description
--workflow / positional	Yes	--	Path to YAML file in workflows/
--execute	No	false	Actually run the workflow (default is dry-run/preview)
--workspace	No	out/{name}-{run_id}	Override workspace directory
--params	No	--	Trigger parameter overrides as key=value pairs

---

Execution Flow

Step 0: Parse and Compile

Parse --params arguments into a dict. Then compile the workflow, passing every parsed param — one --params flag per pair (--params a=1 b=2 is a usage error):

./bin/workflow compile <WORKFLOW_PATH> --format json \
  [--params key=value] [--params key=value]...

Compile enforces the workflow's trigger.param_rules/required before any substitution. If compile or init-workspace exits non-zero, stop and report; never substitute params into prompts yourself, since the engine's rule check is the boundary.

Always invoke through the ./bin/workflow wrapper. Never use .venv/bin/python bin/workflow. The wrapper re-execs under the repo .venv when one exists and falls back to system Python otherwise. A hardcoded .venv/bin/python fails outright in a git worktree, which never has its own .venv — only the main checkout does, and make targets build it on demand.