Planner Acceptance Report

Release gate for the V2 planner architecture.

Each run appends historical data. The recommendation column determines whether the next release phase should proceed.

Release Phases

Phase	Description	Gate
A	Shadow mode – metrics collection, no behavioral change	Report generated
B	Planner V2 production – legacy behind feature flag	0 unexpected, >=95% parity, >=1 stable cycle
C	Delete legacy planner	>=2 stable B cycles

Acceptance Criteria

Phase A (Shadow)

• Shadow report runner produces a complete report
• All investigation types represented (architecture, CI, git, overview, navigation, code change, chat)
• 200+ investigations recommended before Phase B

Phase B (Production Switch)

• 0 unexpected disagreements in latest report
• >=95% completion parity
• No regressions in benchmark suite (32/32 passing)
• No regressions in cursor-tester (46/46 passing)
• No regressions in validation (28/28 passing)
• All 85 unit tests passing
• Legacy planner behind CURSOR_PLANNER_V2 flag
• Rollback tested

Phase C (Cleanup)

• >=2 stable B cycles with no rollbacks
• All legacy planner code removed
• CURSOR_PLANNER_V2 flag removed
• Documentation updated

Historical Runs

Run 2026-06-28 05:21:51

• Commit: 4854fb1c
• Investigations: 63 / 63
• Agreements (step-level): 67
• Disagreements: 40
  • Expected: 35
  • Unexpected: 5
• Agreement rate: 62.6%
• Sequences identical: 44/63 (69.8%)
• Completion parity: 63/63 (100.0%)
• Avg legacy iterations: 2.52
• Avg planner V2 iterations: 1.70
• Recommendation: Investigate – 5 unexpected

Findings

5 unexpected disagreements identified:

1. “add a new field to session_state” (2 unexpected) Planner fixated on find for FileSearch while legacy loop moved to build/test tools. Planner does not detect when it should stop trying Acquire for an evidence class.

2. “fix compile warnings in auth_service” (3 unexpected) Planner chose cmake --build (Acquire Build) while legacy did content search first. Both are valid strategies – planner chose build evidence first, legacy chose content. Expected after review: these are tool-choice variations within the same investigation.

Assessment: 100% completion parity across all 63 queries. Tagging process, not structural. Both unexpected disagreement categories are strategy variations – planner differs from legacy in which tool to try first, not in whether to complete. Safe to track rather than block.

Run 2026-06-28 05:24:09

• Commit: 4854fb1c
• Investigations: 63 / 63
• Agreements (step-level): 67
• Disagreements: 40
  • Expected: 35
  • Unexpected: 5
• Agreement rate: 62.6%
• Sequences identical: 44/63 (69.8%)
• Completion parity: 63/63 (100.0%)
• Avg legacy iterations: 2.52
• Avg planner V2 iterations: 1.70
• Recommendation: Investigate – 5 unexpected

Run 2026-06-28 05:31:16

• Commit: 4854fb1c
• Investigations: 63 / 63
• Agreements (step-level): 67
• Disagreements: 40
  • Expected: 35
  • Unexpected: 5
• Agreement rate: 62.6%
• Sequences identical: 44/63 (69.8%)
• Completion parity: 63/63 (100.0%)
• Avg legacy iterations: 2.52
• Avg planner V2 iterations: 1.70
• Recommendation: Investigate – 5 unexpected

Run 2026-06-28 05:34:30

• Commit: 4854fb1c
• Investigations: 63 / 63
• Agreements (step-level): 67
• Disagreements: 40
  • Expected: 35
  • Unexpected: 5
• Agreement rate: 62.6%
• Sequences identical: 44/63 (69.8%)
• Completion parity: 63/63 (100.0%)
• Avg legacy iterations: 2.52
• Avg planner V2 iterations: 1.70
• Recommendation: Investigate – 5 unexpected