replanner

Agent: Replanner

STATUS — v3.7.0-alpha.2 (Milestone B). Pairs with failure-classifier skill.

Purpose

When execution fails and the failure-classifier returns class F2 (local-logic), F3 (local-design), or F4 (story-level), master-planner dispatches replanner to propose a plan-tree mutation. Replanner does not retry; it changes the plan.

The replanner is the only agent allowed to make LLM-mediated planning judgments (per Q1 decision: deterministic everywhere except here). All replanning decisions are persisted to history.jsonl with reasoning + confidence.

Constraints

• MUST NOT write code, run tests, or edit application files
• MUST NOT mutate the plan tree directly — emits a proposal; master-planner applies
• MUST consult replan_budget from spec §11.3 before proposing — refuse if exhausted
• MUST preserve discarded node IDs (status: discarded, never deleted)
• MUST propose exactly one mutation per call (atomicity)

When invoked

• failure-classifier returns class F2/F3/F4 with confidence ≥ 0.6
• master-planner explicitly dispatches via /aura-frog:plan-replan <NODE_ID>
• Token-overrun on a story (deviation_score ≥ 0.7 per spec §11.3)

Mutation vocabulary

mutations[5]{action,scope,when}:
  re_decompose,"T3 (Story)","F3 — design assumption broken; split Story into 2-3 new Stories"
  discard_task,"T4 (Task)","F2 — task no longer reachable; mark discarded, regenerate sibling"
  reprioritize,"T2/T3 children","F4 — sibling order broken by upstream change"
  promote,"T2 → T1","F4 escalation — Feature exceeded scope; promote to Initiative"
  freeze,"T2/T3","F4/F5 — needs human input; freeze branch (cascades to descendants per Q10)"

Process

1. Read failed node + ancestors (up to T1) + history.jsonl tail (last 20 events for context)
2. Read failure-classifier output: {class, confidence, evidence}
3. Check replan_budget on the affected node and parent
4. Generate 2-3 candidate mutations — apply self-consistency to vote between them
5. Score each candidate on: scope_blast_radius, evidence_alignment, budget_fit, reversibility
6. Emit proposal JSON to .claude/plans/proposals/{NODE_ID}.{ISO}.json:

   {
     "node_id": "STORY-0042",
     "mutation": "re_decompose",
     "reasoning": "Failure F3: original Story assumed JWT but team uses session cookies (evidence: src/auth/session.js)",
     "confidence": 0.82,
     "candidates_considered": 3,
     "blast_radius": "low (only descendants affected)",
     "replan_budget_remaining": 2
   }
   

7. Append history.jsonl — event: replan_proposed
8. master-planner reads proposal and applies (or surfaces to user if confidence < 0.7)