/refine

command: refine
disable-model-invocation: true
description: Improve an artifact through rubric scoring, cross-model critique, edits, and bounded rescoring

Refine

MANDATORY COMPLIANCE

Execute the host-managed `skill-refine` workflow with real independent judge calls through `scripts/refinement-judge.sh`. PROHIBITED: role-playing other models, inventing scores, bypassing the read-only adapter, or declaring success without the score validator's passing result for the delivered artifact.

Read and follow `skills/skill-refine/SKILL.md` from the installed Octopus root. Use `${CLAUDE_PLUGIN_ROOT}` when available, otherwise `${CLAUDE_PLUGIN_ROOT:-${CODEX_PLUGIN_ROOT:-${HOME}/.claude-octopus/plugin}}`. Do not invoke this command recursively.

Input: $ARGUMENTS

Examples:

/octo:refine plans/migration.md until 95+, using Claude Opus and Codex
/octo:refine this proposal for a technical buyer --max-rounds 3
/octo:refine src/parser.py --rubric review-rubric.json --target 90
/octo:refine draft.md --review-only

These options are instructions to the host workflow, not flags for `orchestrate.sh`: `--rubric <file>`, `--target <0..100>`, `--max-rounds <1..10>`, `--max-minutes <positive integer>`, `--max-calls <positive integer>`, and `--review-only`.

Defaults: target 95/100, five improvement rounds, 30 minutes, 14 external model calls. Stop when any limit is reached or two rounds make no verified progress. An explicit user budget overrides the defaults. Validate options before edits or provider calls. `/octo:ladder` is an alias for this workflow.