dragonhoard explore

CLI: python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dragonhoard.py" (called dragonhoard.py below). Input: $ARGUMENTS. That is the open choice, or empty (then take it from the current pack's Open options).

1. Frame the choice

Write down three things:

• the question;
• 2 or 3 candidate options;
• one success criterion per option, as a command whose output decides pass or fail.

You may also add a cost ceiling, such as "no new runtime dependency". Check <root>/DIGEST.md first: if an option.chosen event already answers this question, stop and cite it.

2. Retrieve (cheap, parallel)

Run dragonhoard.py role scout and follow what it prints, one scout per factual question. Where the host has scout subagents, launch them together so they run in parallel. Typical questions: does the library support X in our version; how the codebase already does Y; what an API returns. Drop options the facts rule out.

If only one option survives, skip to step 5 and present it, with the facts that eliminated the others.

3. Spike (parallel, isolated, bounded)

Run dragonhoard.py role spiker and follow what it prints: one spiker per surviving option, at most 3, each in its own isolated worktree. They run in parallel where the host has spiker subagents, and one at a time inline otherwise. Give each spiker:

• the option;
• its success criterion command;
• the files it may touch;
• a timebox of about 15 tool calls;
• the relevant pack excerpt.

Spikes are evidence, not delivery. Their code is never merged.

4. Judge

Run dragonhoard.py role judge and follow what it prints. If it says PAUSE, do exactly what it says: ask the owner to switch models before judging, or (external judge) stop and hand the judge request to the orchestrator. Give the judge the question, the criteria and every spiker report. It verifies the reports, then returns a comparison table, a recommendation and warnings.

5. Owner picks

Ask the owner and wait for the answer (in Claude Code, use AskUserQuestion with the evidence in each option's preview). Put the judge's recommendation first, labelled "(Recommended)". For each option, give its verdict, its criterion result and its cost. The owner may also answer "none — rethink".

6. Record

dragonhoard.py log <task> option.chosen "<question>: <choice> — <decisive evidence>" --ref tasks/<task>.journal.md

Add an Explored: line to the task's journal attempt, listing each option, its verdict and its criterion output.

List the spike worktrees that still exist (git worktree list). Ask the owner before removing them; do not delete them yourself.

Return to the Verify step of /dragonhoard:next. Build the chosen option properly there, test first, inside the pack's file list. Do not copy the spike.