1.5 KiB
1.5 KiB
Sprint 122 Plan: Deterministic Patching and Replay for SLM Agents
Context
Sprint 121 established deterministic failure capture and loop scaffolding. Sprint 122 expands the debugging substrate so low-context agents can request root-cause clusters, assemble minimal fix context, propose constrained patches, and archive/replay repro packets through MCP tools.
Goals
- Add deterministic cluster/context/patch MCP tools
- Add dry-run patch application and invariant enforcement
- Persist and replay repro packets for training/eval loops
- Emit machine-readable debugging metrics for SLM readiness
Steps
Step 1459: Failure fixture catalog for deterministic debug benchmarks (12 tests)
Step 1460: Patch dry-run applier and validation result model (10 tests)
Step 1461: Patch invariant checker (scope + safety rules) (10 tests)
Step 1462: whetstone_cluster_failures MCP tool (8 tests)
Step 1463: whetstone_assemble_fix_context MCP tool (8 tests)
Step 1464: whetstone_propose_patch_for_failure MCP tool (8 tests)
Step 1465: whetstone_save_repro_packet MCP tool (8 tests)
Step 1466: whetstone_replay_repro_packet MCP tool (8 tests)
Step 1467: Debugging metrics packet (time_to_green, iterations, etc.) (8 tests)
Step 1468: Sprint 122 integration summary + regression (8 tests)
Architecture Gate
- Patch generation must be validated before apply
- Repro packets must be deterministic and portable
- MCP outputs must be machine-readable and stable
- Max 600 lines per header