agent-optimize

agent-optimize — the free-form loop you own

The one algorithm with no deterministic subprocess and no per-iteration optimizer: you — the agent that ran intake — are the optimizer, the scheduler and the stopping rule. cap-evolve run (with orchestration_mode: agent) does check → baseline, prints a handoff with the run_dir, and returns. From there the search is yours, bounded by the invariants core enforces and the project's free-text stop_condition. Drive the existing primitives, so the run dir and dashboard stay populated as in a deterministic run.

Shell variables used below

R="<run_dir from the agent-mode handoff>"      # e.g. .capevolve/run_20250101_120000
P="<project dir>"                              # the dir holding capevolve.yaml + adapters/
S="${CAPEVOLVE_SKILLS_DIR:?set CAPEVOLVE_SKILLS_DIR to the skills/ dir}"
A="$S/algorithms/agent-optimize/scripts"       # this skill's helpers
mkdir -p "$R/work"                             # working copies live here (RunDir does NOT create it)

Every script imports _bootstrap itself (no PYTHONPATH) and prints JSON on stdout.

Phase 0 — understand before you optimize

Once, before any edit, and ask the user any blocking question here so the loop then runs unattended. Read PROJECT.md, capevolve.yaml, the adapter and every file under capability_path, and understand what one evaluation does: what a task is, what run_target produces, what score() rewards, and what the per-task feedback says — that is your learning signal. Note the val/test sizes, num_trials, gate_mode/gate_k_se and the allowed edit surface.

Then let spend.py parse the free-text stop_condition rather than restating it from memory: it prints constraints.predicates, every concrete check it could extract, with its measured actual. If constraints.ambiguous is non-empty, ASK THE USER before the loop starts — a vague clause ("don't spend too much", a bare number with no unit) is reported, never guessed at, and this is the one moment where asking is cheap.