dragonhoard next

CLI: python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dragonhoard.py" (called dragonhoard.py below). <root> is the root in dragonhoard.json; the standard checks are its checks.

In hybrid mode, the Claude Code session uses /dragonhoard:conduct instead. It runs this skill itself only when conduct escalates a task. The OpenCode builder runs this skill normally.

0. Pick

dragonhoard.py next

• none ready: report the blockers from <root>/DIGEST.md and stop.
• The task has no pack: run /dragonhoard:shape M<n> first.
• Otherwise set the task's row to IN_PROGRESS in both STATUS.md and the milestone's PROGRESS.md. Then run dragonhoard.py log <task> task.started "<title>".

Load exactly the context in <root>/README.md step 2: the global constitution, DIGEST.md, LESSONS.md, the milestone constitution, the pack, and the files the pack names. Nothing else. Apply every approved lesson that fits this task.

1. Plan

Restate the goal, scope, files and acceptance in five lines or fewer.

If the pack is wrong or stale (it names an interface that does not exist, or contradicts an option.chosen event or a lesson): record BLOCKED (section 7) and stop. Never quietly rewrite a pack's scope.

2. Explore (only if needed)

Skip this step when the pack's Open options says "none" and you have no unknowns.

Otherwise:

• A fact is missing: run scouts, one question each, as dragonhoard.py role scout says. Scouts are cheap and read-only.
• Two or more viable options remain: run /dragonhoard:explore. It spikes options in parallel, gets the judge's comparison and asks the owner to pick.

Never re-open a choice that already has an option.chosen event unless there is new evidence.

3. Verify (before building)

• grep -n every interface the pack's Consumes names. Paste the hits into the journal. If one is missing or different, record BLOCKED.
• Run every standard check and record the counts. If a check is red before your change, record BLOCKED; do not fix unrelated breakage.
• Write the focused test first. Run it and confirm it fails for the expected reason.

4. Build

Make the smallest change that satisfies acceptance, inside the pack's file list only. If you need a file outside the list, stop and report it; that is a pack change.

5. Test

Run the focused test, then every standard check. Tick each acceptance item with evidence.

If a run fails, write a journal attempt (Outcome FAIL, with the root cause and evidence) and try again. Two failed bounded attempts at the same issue means BLOCKED.

6. Record

1. Append the attempt block to tasks/<task>.journal.md. Use the format in <root>/README.md, with the real commands and the shortest decisive output line.
2. Update the milestone PROGRESS.md row: status, the changed files in backticks, the verification command and its real result, and the next task.
3. Update the STATUS.md row with the same status.

4. Log and refresh:

   dragonhoard.py log <task> task.done "<one line: what now works>" --ref tasks/<task>.journal.md
   dragonhoard.py digest
   dragonhoard.py check   # must PASS

5. If this was the milestone's last task, run the GATE.md commands. Append the real output to its results log, set the gate row in STATUS.md, and run dragonhoard.py log - gate.run "G<n> PASSED|PARTIAL: <why>".

7. BLOCKED

• Set BLOCKED in both tables, and put the reason and the owner action in the Blocker column.
• Write a journal attempt with Outcome BLOCKED.
• Log dragonhoard.py log <task> task.blocked "<reason; owner action>", then run digest and check.

8. Stop

Report what was done, the evidence, and the next task (dragonhoard.py next). Commits are an owner action; do not commit unless asked.

If the journal shows a FAIL→PASS story, a surprising root cause or an owner correction, offer /dragonhoard:learn.