Planner Validate Task Alignment

planner-validate-task-alignment

Post-refinement catch-net. Compares the plan's phases, assignments, and WP descriptions against the original task description. Emits warning-severity findings for misalignment. This is a safety net — if upstream task injection works correctly, this step should produce zero findings.

Arguments

• $1 — Absolute path to refined_wps.json (PlanDocument with task, work_packages[])
• $2 — Absolute path to refined_plan.json (PlanDocument with task, phases[])
• $3 — Absolute path to output directory for findings

Critical Constraints

NEVER:

• Block the pipeline — all findings are warning severity
• Write output outside $3/
• Read files not passed as arguments
• Modify input files
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Read the task field from $1 or $2
• Compare every phase goal and every WP description against the task
• Write $3/task_alignment.json with findings array
• Emit: alignment_findings_path = <absolute path to task_alignment.json>; also emit alignment_finding_count

Workflow

Step 1: Read inputs

Read refined_wps.json from $1 and refined_plan.json from $2. Extract:

• The task field (the user's original task description)
• All phase goal and scope fields from $2
• All WP name, deliverables, and acceptance_criteria fields from $1

Step 2: Spawn alignment-check subagents

Spawn 1-2 subagents (model: "sonnet"):

Subagent A — Phase alignment: Provide the task description and all phase goals/scopes. Ask: "For each phase, does its goal directly serve the stated task? Rate each phase as 'aligned', 'tangential', or 'unrelated'. A phase is 'aligned' if the task explicitly or implicitly requires the work. A phase is 'tangential' if it supports the task but was not requested. A phase is 'unrelated' if the task does not mention or imply this work."

Subagent B — WP alignment: Provide the task description and all WP names/deliverables. Ask: "For each WP, do its deliverables serve the stated task? Flag any WP whose deliverables address concerns not mentioned in the task."

Step 3: Collect and write findings

Parse subagent responses. For each phase or WP rated 'tangential' or 'unrelated', emit a finding:

{
  "message": "Phase P2 ('Protocol Sharding') appears unrelated to the stated task",
  "severity": "warning",
  "check": "task_alignment",
  "entity_id": "P2",
  "rating": "unrelated"
}

Write all findings to $3/task_alignment.json:

{
  "schema_version": 1,
  "task": "<original task>",
  "findings": [],
  "summary": {
    "aligned": 4,
    "tangential": 1,
    "unrelated": 0
  }
}

Step 4: Emit output tokens

alignment_findings_path = <absolute path to $3/task_alignment.json>
alignment_finding_count = <N>