configure-models

Set up or adjust local model overrides for review-chain agents. The override hook (hooks/model-override.sh) reads the resulting config; changes take effect on the next agent spawn — no restart.

Engine

The deterministic template generator is bundled at references/gen-model-config.sh (relative to this SKILL.md). Always produce the template via that script — never hand-write it. Resolve its absolute path from this skill's directory and run bash <that-path> [args].

Script args: --project (write ./.claude/review-chain-models.conf), --user (write ~/.claude/review-chain-models.conf), PATH (write there), none (print to stdout). Prefix -f to overwrite an existing file.

Steps

1. Pick the target. Default to --project unless the user asked for user-global (--user) or a specific path.

2. Generate. Run the script for that target.

   • If it refuses because the file already exists, do NOT blindly pass -f. Read the existing file, show the user its active (uncommented) lines, and only regenerate with -f if they want a fresh template — warn that it drops their current edits.

3. Apply any specific overrides the user named. If the user asked to set particular agents (e.g. "implementer to opus", "all the deep reviewers to opus", "everything on opus"), edit the written file: uncomment the matching #agent = line and set the model. Use exact agent names as they appear in the file; a blanket request maps to the * = <model> line. If the user only wants the template to edit themselves, skip this.

4. Report briefly. Give the absolute path of the config file and list which overrides are now active. One-line reminders, only as relevant:

   • Commented lines are inactive; uncomment agent = model to activate.
   • Env vars win over the file: REVIEW_CHAIN_MODEL_<AGENT> and REVIEW_CHAIN_MODEL_ALL (e.g. REVIEW_CHAIN_MODEL_IMPLEMENTER=opus). The project file beats the user file. An explicit model the orchestrator already chose for a spawn is respected over the file.
   • Takes effect on the next agent spawn.
   • The built-in Explore agent is configured separately via REVIEW_CHAIN_EXPLORE_HOOK.