Context

Extracted from #14 (task decomposition and routing engine). The DecompositionStrategy protocol and ManualDecompositionStrategy are implemented. This issue covers the LLM-based strategy where an agent autonomously decides how to break down a parent task into subtasks.

Depends on the agent execution loop being available to make LLM calls.

Acceptance Criteria

• LlmDecompositionStrategy implements DecompositionStrategy protocol
• Uses an LLM provider to analyze the parent task and generate a DecompositionPlan
• Respects DecompositionContext constraints (max_subtasks, max_depth)
• Validates LLM output (well-formed subtask IDs, valid dependencies, no cycles)
• Handles LLM failures gracefully (malformed output, refusal, timeout)
• Unit tests with mocked provider calls
• Integration test with real provider (optional, behind marker)

Design Spec Reference

• §6.9 — Task Decomposability & Coordination Topology
• §6 — Task Workflow

Notes

• The DecompositionStrategy protocol (engine/decomposition/protocol.py) already defines the interface
• DecompositionService accepts any strategy — plug in the new one via dependency injection
• Consider prompt engineering for consistent subtask ID generation and dependency specification