evolve-score

evolve-score

Score a single candidate. This is deterministic — the work is "run the user's evaluator.py and record the number." There is no algorithmic judgment to make. The reason it is a skill (and, in the omnibus, a Haiku subagent) is purely to keep evaluator output — which can be thousands of lines — out of the main conversation.

Resolve the plugin root

This skill runs with $CLAUDE_PLUGIN_ROOT set. Capture it:

echo "PLUGIN_ROOT=$CLAUDE_PLUGIN_ROOT"

If a subagent invoked you without it, fall back to the path the caller passed you in the prompt.

Score the candidate

Run exactly one command. It syntax-checks the candidate, runs validator.py if the workspace has one, runs evaluator.py under the sandbox, parses the score, and writes the result to the CSV:

python3 "$CLAUDE_PLUGIN_ROOT/scripts/score.py" --working-dir "<WORKING_DIR>" <CANDIDATE_ID>

• <WORKING_DIR> is the evolution workspace (the directory containing config.yaml). If the user didn't specify one, omit --working-dir and let it auto-detect evolution/config.yaml or ./config.yaml.
• The script prints one JSON line on stdout (everything else is stderr noise you can ignore):
  • success: {"id":"...","ok":true,"score":1.23,"status":"complete","extra":{...}}
  • failure: {"id":"...","ok":false,"status":"failed|failed-validation","error":"...","stage":"syntax|validator|evaluator"}

The script has already written the status and performance to evolution.csv either way — you do not touch the CSV yourself.

Report

Return a single line to whoever called you:

• success → <id>: score=<n> (complete)
• failure → <id>: FAILED at <stage> — <short error> (status=<status>)

Do not dump the full evaluator output into your reply. If the user explicitly asks why something failed, include the error field; otherwise keep it to the one-liner. That brevity is the whole point of running scoring in an isolated context.

Do not

• Do not edit the candidate's code to make it pass — that is evolve-code's job. A failing score is a real result, not a problem to fix here.
• Do not retry with a different evaluator or fabricate a score. If the evaluator fails, report the failure honestly; the engine has recorded failed.