planning-with-files

Persist task memory for long-running agent workflows by writing canonical files like task_plan.md and findings.md to a durable session directory, ensuring state survives context compaction and orchestrator handoffs.

Manage orchestrator and worker state for multi-phase tasks using a single-writer discipline on canonical memory files, while workers stage work in separate scratch directories to avoid touching canonical state.

Agents lose progress and findings during long-running tasks because in-memory state is lost during context compaction, a problem solved by persisting durable canonical memory files under a dedicated session path.

Maintain an audit trail for task lifecycle phases and claims by writing canonical files including progress.md and claims.md under .openscientist/sessions to ensure a coherent record across all participating workers.

Yes, this memory protocol supports scratch-agent directories specifically for workers to stage work without touching canonical files, enforcing a single-writer discipline to keep state coherent across the orchestrator.

Yes, you can use this session memory protocol for multi-phase tasks requiring durable final reports, as it persists report.md alongside task plans and findings to survive worker handoffs.