rehyra

name	rehyra
description	Run, resume, inspect, or audit a fresh-workspace ReHyra recursive optimization campaign over a code workspace. Use when the user asks Codex to optimize an implementation against tests, benchmarks, scores, latency, cost, accuracy, or another executable objective; compare recursively generated candidates; continue a prior ReHyra run; or explain why candidates were promoted or rejected.

ReHyra

Turn the user's goal into a frozen, evaluator-controlled search. Keep the source workspace unchanged until the user explicitly asks to apply the winner.

Locate the bundled runtime

Resolve the Codex Adapter directory two levels above the directory containing this SKILL.md. Use its rehyra_plugin.py for every deterministic operation. Do not assume the rehyra console script is globally installed.

Run the preflight first:

python <adapter-dir>/rehyra_plugin.py doctor

Start a campaign

1. Identify the workspace root and the requested outcome.
2. Find an executable evaluator before spending model calls. Prefer, in order:
   • a benchmark command named by the user;
   • an existing project benchmark or test command with a scalar metric;
   • a small deterministic evaluator added only to the experiment workspace;
   • ReHyra's generated-evaluator mode when no fixed oracle exists.
3. Define one primary objective, its direction, and hard qualification constraints. Never treat test pass/fail alone as an optimization score.
4. Run the evaluator on the untouched workspace at least twice. If its result is flaky, leaks future/private data, scores the source workspace instead of the candidate, or cannot distinguish obvious good and bad artifacts, stop and explain the blocker.
5. Select the candidate submission path. The Proposal Agent may edit the fresh proposal workspace, so qualification checks must protect behavior outside the primary metric. Plugin-created tasks freeze every other workspace file before evaluation. Use editable-region markers when the task contract requires an even narrower mutation surface inside the submission.
6. Choose a budget. Default to four proposals and concurrency two when the user did not specify either. Treat larger or costly campaigns as material work and state the budget before starting.
7. Create an experiment outside the source workspace:

python <adapter-dir>/rehyra_plugin.py prepare \
  --workspace <workspace> \
  --output <absolute-experiment-directory> \
  --name <short-name> \
  --goal <goal> \
  --submission <relative-submission-path> \
  --objective-key <metric-key> \
  --direction <minimize-or-maximize> \
  --evaluator-program <program> \
  --evaluator-arg <arg-with-{workspace}-or-{submission}> \
  --max-proposals <budget> \
  --max-concurrency <concurrency>

Use --evaluator-mode generated instead of evaluator program/arguments only when a fixed evaluator cannot be established. Label conclusions from a generated evaluator as advisory unless independent calibration exists.

8. Inspect the returned JSON, then validate the task before running:

python <adapter-dir>/rehyra_plugin.py cli describe-task <task.json>
python <adapter-dir>/rehyra_plugin.py cli run <run.toml> \
  --run-dir <run-directory>

Send short progress updates while the campaign runs. A failed proposal remains evidence and still consumes its frozen slot.

Resume or inspect

Re-run the same cli run command with the same run directory to recover an interrupted campaign. Never delete or rewrite completed attempts.

Use:

python <adapter-dir>/rehyra_plugin.py cli status <run-directory>
python <adapter-dir>/rehyra_plugin.py cli validate <run-directory>
python <adapter-dir>/rehyra_plugin.py best <run-directory>

For a larger continuation budget, create a new compatible experiment and use ReHyra's --seed-from option. Preserve the earlier run as immutable evidence.

Deliver the result

After completion:

1. Run validate.
2. Resolve the winner with best.
3. Compare the winning workspace against the original source workspace.
4. Report baseline, winner, absolute and relative change, qualification results, proposal count, failures, model usage, run directory, and the winning attempt. Include the sealed evidence root returned by best.
5. Show a reviewable patch. Apply it to the source workspace only when the user explicitly requests application.

Distinguish artifact discovery from reliable search performance. One winning campaign does not establish low variance or general superiority.