planner-refine-assignments

planner-refine-assignments

L1 session that refines the assignments for a single phase by spawning one L0 subagent per assignment in parallel (3–5 assignments per phase). Receives a per-phase context file produced by merge_tier_results → _write_refine_contexts, which contains only this phase's assignments and peer_summaries (id/name/goal stubs) for all other phases. Each L0 reviews its assignment in the context of peer_summaries and refined_plan.json, returning structured suggestions. L1 applies suggestions, resolves intra-phase WP ownership conflicts, and writes the phase result file to $3/refine_contexts/{phase_id}_result.json.

When to Use

• Dispatched by the L2 planner recipe in parallel — one session per phase
• Accepts a per-phase context file from $3/refine_contexts/context_{phase_id}.json and refined_plan.json for phase context
• Produces {phase_id}_result.json as input for the downstream merge_refined_assignments step

Arguments

• $1 — Absolute path to the per-phase context file ($3/refine_contexts/context_{phase_id}.json). The file contains:
  • phase_id — identifier for the phase this session processes
  • task_file_path — path to the task description file (read from disk, not inline)
  • assignments — full AssignmentElaborated objects for this phase only (3–5 entries)
  • peer_summaries — {id, name, goal} stubs for all assignments in other phases
• $2 — Absolute path to refined_plan.json (PlanDocument with phases as PhaseElaborated, for phase context)
• $3 — Absolute path to the run-scoped planner directory (e.g., {{AUTOSKILLIT_TEMP}}/planner/run-YYYYMMDD-HHMMSS). Output is written to $3/refine_contexts/{phase_id}_result.json.

Critical Constraints

NEVER:

• Write any file outside $3/
• Directly modify the context file ($1) — always write a new result file
• Allow an L0 subagent to write files directly (L0s return structured text only)
• Emit phase_refined_path before writing the result file
• Skip emitting phase_refined_path even if all L0s fail (write unchanged assignments, still emit)
• Spawn more than 6 L0s in a single parallel batch
• Read {{AUTOSKILLIT_TEMP}} artifacts not passed as positional arguments
• Run subagents in the background (run_in_background: true is prohibited)
• Write L0 prompts to intermediate l0_prompts/ files and read them back into the L1 context — spawn L0 subagents directly from in-memory context packets
• Read source code files, test files, or recipe YAML files directly — codebase exploration is the L0 subagents' responsibility
• Run Bash, Grep, or Glob commands for codebase exploration between L0 spawns