Combo Test: Forked Parallel Exploration with Task Coordination

Forked agents can coordinate by reading and updating shared task state, enabling divide-and-conquer parallel research without explicit message passing.

How It Works

Multiple agents spawned in parallel can operate on the same task directory (~/.claude/tasks/{session}/). Each agent creates its own task, performs independent work, and updates task metadata with findings. All agents see the full task list, including tasks created by other agents in the same session. This creates a coordination layer without requiring direct inter-process communication.

Test Setup and Results

The test spawned two agents in parallel: FORK-EXPLORER-A (investigating task system architecture) and FORK-EXPLORER-B (investigating session ID discovery). Both agents created tasks, performed research, and updated task metadata with findings. Status: VALIDATED. Agent A created task #21, agent B created task #22. Both agents observed the complete shared task list immediately. Findings were persisted in task metadata: architecture details, session ID discovery methods, and available environment variables. The test confirmed session ID is available via hooks (JSON stdin), sessions-index.json, and skill string substitution, but NOT via environment variables.

Why It Matters

This pattern enables divide-and-conquer exploration where parallel agents investigate separate hypotheses, then the orchestrator synthesizes findings from all task metadata. Findings persist even if an agent terminates, since updates are written to shared disk. No custom IPC layer needed.

Caveats

This is a runtime test, not a formal specification. The coordination mechanism relies on agents independently discovering the shared task state; there is no explicit synchronization primitive. Concurrent writes to the same task could create race conditions (not tested). Agents must poll or be notified externally of new tasks from peers.