f24a5dd5ee
Add eval.rs module that calls ollama to independently verify task success/failure. Prefers small fast models (deepseek-r1:7b, qwen2.5:3b, llama3.2:1b) via detect_best_local_model(). Binary PASS/FAIL prompt with 2000-char task description trim. 30s timeout. Wired into record_task_completion() via tokio::task::spawn_blocking after agent self-report — non-blocking, best-effort. Uses INSERT OR REPLACE to overwrite correctness_check from local model. Keeps agent self-report eval_mode: 'agent' row; local eval writes a second row with eval_mode: 'local-llm'. 3 unit tests for prompt building + struct fields. All 102 daemon tests pass. fmt + clippy clean. |
||
|---|---|---|
| .. | ||
| 2026-06-28-eval-phase2-local-llm.md | ||