Use this skill for Attractor graph work. Treat DOT as executable workflow policy, not a diagram.
Workflow
- Read the repo's
AGENTS.mdand the local Attractor spec or compact LLM guide if present:attractor-spec.md,cmd/attractor/attractor-spec.md, orcmd/attractor/llms.txt. - Inspect nearby examples before editing syntax. Prefer the target repo's own
Attractor examples, commonly under
examples/,attractor-go/examples/, or another path named by the repo'sAGENTS.md. - Identify the graph's goal, phases of intent, loop placement, routing points, critical goal gate, provider/model assumptions, and proof command.
- Keep graphs small. Put iteration where repeated attempts improve the result; do not add nodes just to mirror a prose checklist.
- Prefer semantic routing and fan-in evaluation where the decision is semantic. Do not confuse execution success with approval.
- Prove the exact DOT or exact failing example. Do not substitute a similar graph unless the user explicitly accepts the substitution.
- For repo fixes, use worktree isolation, record proof, put proof in the PR body, squash merge when allowed, and clean up the worktree/runtime afterward.
When To Read References
- Read patterns.md when choosing routing, parallel, steering, fidelity, or goal-gate patterns.
- Read proof.md before calling an Attractor graph fixed, especially for examples or issue work.
Output
Return or commit:
- graph/pipeline files changed,
- why the loop/routing structure fits the goal,
- exact proof command and result,
- artifact/log paths that prove the run,
- any upstream or runner-neutral boundary issue.
source: agents