skills: add attractor skill

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

skills/attractor/SKILL.md

@@ -0,0 +1,94 @@
+---
+name: attractor
+description: Build, extend, and debug Attractor implementations from the strongdm/attractor specs. Use when work involves DOT pipeline DSL parsing, graph execution traversal, node handlers, checkpoint/resume state, human-in-the-loop gates, condition expressions, model stylesheet rules, or integrating coding-agent-loop/unified-llm backends.
+metadata: { "openclaw": { "emoji": "🧲", "homepage": "https://github.com/strongdm/attractor" } }
+---
+
+# Attractor
+
+Implement Attractor as a spec-driven orchestration layer, not as ad-hoc glue code.
+
+## Source of truth
+
+Read these specs before changing architecture:
+
+- `attractor-spec.md`
+- `coding-agent-loop-spec.md`
+- `unified-llm-spec.md`
+
+If local files are unavailable, use:
+
+- <https://github.com/strongdm/attractor/blob/main/attractor-spec.md>
+- <https://github.com/strongdm/attractor/blob/main/coding-agent-loop-spec.md>
+- <https://github.com/strongdm/attractor/blob/main/unified-llm-spec.md>
+
+## Execution order
+
+1. Confirm scope.
+2. Map requested work to specific spec sections.
+3. Implement minimal, testable slices in this order:
+   - DOT parser + schema validation
+   - Graph normalization + static validation
+   - Execution engine (state, traversal, edge selection)
+   - Node handler registry + built-in handlers
+   - Checkpoint/resume persistence
+   - Human-in-the-loop pause/resume flow
+   - Model stylesheet and condition evaluator
+4. Add focused tests for each implemented slice.
+5. Verify behavior against acceptance criteria from the relevant section.
+
+## Scope mapping checklist
+
+Map each task to one or more of these components:
+
+- **DOT DSL schema**: grammar subset, typed attributes, defaults, chained edges, subgraph scoping.
+- **Execution engine**: session/run state, deterministic traversal, retries, timeouts, loop handling.
+- **Node handlers**: `start`, `exit`, `codergen`, `wait.human`, `conditional`, `parallel`, `parallel.fan_in`, `tool`, `stack.manager_loop`.
+- **State and context**: run context shape, per-node outputs, status normalization.
+- **Human-in-the-loop**: blocking prompts, explicit response routing, resumable state.
+- **Model stylesheet**: selector precedence, inheritance, node-level override merge.
+- **Condition language**: parser/evaluator behavior, truthiness, error handling.
+- **Transforms/extensibility**: pre/post transforms, plugin-like handler registration, event hooks.
+
+## Guardrails
+
+- Keep Attractor decoupled from the LLM backend. Use backend interfaces, not provider-specific calls in core traversal logic.
+- Keep graph execution deterministic. If multiple edges are eligible, apply documented priority rules (condition, weight, label preferences) consistently.
+- Reject invalid graph inputs early with clear errors (do not silently coerce malformed data).
+- Keep checkpoint payloads stable and serializable to support crash-safe resume.
+- Preserve separation between:
+  - orchestration (Attractor),
+  - coding agent loop,
+  - unified LLM client.
+- Do not invent undocumented node types, attributes, or condition operators unless explicitly requested.
+
+## Testing requirements
+
+Always add or update tests for changed behavior:
+
+- Parser acceptance tests for valid DOT samples.
+- Parser rejection tests for unsupported syntax and bad attributes.
+- Execution tests for linear, conditional, retry, and parallel paths.
+- Checkpoint/resume tests that simulate interruption and restart.
+- Human-gate tests that pause, accept input, and continue correctly.
+- Condition evaluation tests for happy path and invalid expressions.
+
+## Implementation strategy for large requests
+
+For multi-feature asks, split work into PR-sized increments:
+
+1. Parser + validator foundation.
+2. Core traversal + deterministic routing.
+3. Handler implementations and HITL integration.
+4. Stylesheet/condition language refinements.
+
+Each increment should keep the engine runnable and test-backed.
+
+## Delivery checklist
+
+Before finishing:
+
+- Confirm implementation matches cited spec sections.
+- Confirm tests cover both success and failure paths.
+- Confirm any new config/attributes are documented where users read them.
+- Confirm no coupling leaks between orchestration and provider-specific adapters.