dragonhoard models

CLI: python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dragonhoard.py" (called dragonhoard.py below).

1. Read the roles

dragonhoard.py models

It prints the host, then each role's needs, how often it runs and its current model (or unset). Last comes a hint for finding the models this host can use. Requirements stay fixed; only the model per role is a choice.

2. Find the available models

• If the hint names a command, such as opencode models or the Claude Code aliases, use it.
• Otherwise, ask the owner which models or providers this host can use, and wait.
• Never open provider config or auth files to find out. They hold keys.

3. Suggest

Propose one model per role. Give each one a single line of reasoning that ties the model to that role's needs:

• scout: the cheapest fast model that uses tools reliably. Many run per task, and they retrieve rather than reason.
• spiker: a solid coding model at mid cost. Up to 3 run in parallel.
• judge: the strongest reasoning model available. It runs rarely, so cost matters least. On hosts where dragonhoard.py role judge says PAUSE, this is the model the owner will switch the session to.
• builder (hybrid only): the OpenCode model that works each task. It reads skills step by step and runs tests, at the lowest cost that still passes verify; a free model is a fair first choice. In hybrid the judge is external (the orchestrating Claude Code session), so do not suggest a judge model there. Scout and spiker run inside OpenCode too.
• Effort variants (OpenCode): a model can carry a variant, provider/model#variant, such as a reasoning effort (#low, #high, #max). OpenCode rejects a variant the model lacks ("Variant unavailable"), so try one short opencode run -m <model>#<variant> "hi" before setting it. On a free model, the highest effort costs nothing but time: suggest it for builder and spiker.

Also use the journey's evidence, when there is any, from <root>/DIGEST.md and the journals:

• spikes ending INCONCLUSIVE or out of timebox: the spiker may be too weak;
• scouts reporting not found for things that existed: the scout may be too weak;
• judge verdicts that later proved wrong: the judge needs a stronger model;
• a role that never struggles on a costly model: try a cheaper one.

Say plainly that your knowledge of newer models may be out of date, and that the host's own list wins. If a judge is already configured and can run, ask it to review your suggestion (see dragonhoard.py role judge).

4. Owner decides

Show the suggestion as a table (role, model, reason). Ask the owner to confirm or edit it, and wait (in Claude Code, use AskUserQuestion). Never set a model the owner has not confirmed.

5. Record

dragonhoard.py models --set scout=<model> spiker=<model> judge=<model> --why "<one-line reason>"
# hybrid: dragonhoard.py models --set builder=<model> scout=<model> spiker=<model> --why "<reason>"

This saves dragonhoard.json, regenerates the host's role-agent files (OpenCode and Codex), and logs a models.changed event. Show the final dragonhoard.py models output.