optimize.
optimize turns “make this number better” into a reproducible hypothesis loop. You define a single scalar metric and a measurement script; the engine — not the LLM — measures every commit and logs a TSV leaderboard, so the result cannot be gamed by a hallucinated number. It pauses for your approval at batch boundaries. At completion it squashes the branch down to baseline..best_commit and ships an MR/PR carrying the leaderboard.
Steps
- 01Setupread-only
Interactive metric elicitation: the metric + direction, termination conditions, noise control, batch size, frozen paths. Writes the validated settings to optimize-state.json (the baseline step later adds the measured baseline to it).
- 02Metricread-only
Author the measurement script — prints exactly one float, exit 0, deterministic. Validated by a double-run stability gate before you can advance.
- 03Baselinemeasure
Capture the baseline metric on HEAD; project total cost from real timing.
- 04Experimentationmany runs
The hypothesis loop: one hypothesis = one focused commit. The engine measures each commit. Checkpoints between batches return to discussion for your approval.
many agents fan out in parallel - 05Synthesissynthesize
Findings document, plus a non-blocking metric-gaming audit.
- 06Code-reviewspec gate
Full two-stage gate on the cumulative baseline..best_commit diff.
OpusAntigravityCodex - 07Completiondocs · push
Squash from best_commit; ship the leaderboard as the MR/PR description.
writes back to the shared LLM wiki
What makes it distinctive
Engine-owned measurement
The engine runs the metric script on every commit and writes the TSV leaderboard. The LLM’s reported numbers are untrusted — the leaderboard is the single source of truth, so it cannot be gamed.
Frozen paths
During experimentation, the metric script and test files are frozen — a hook physically blocks edits to them even in implementation mode. You can’t move the goalposts mid-experiment.
Batch checkpoints
After each batch of hypotheses the protocol returns DAIC to discussion for your approval before continuing — keeping a human in the loop on an otherwise mechanical loop.
branch optimize/ — o-<name>