Route a task to a script, an agent or yourself
Gate: none. Routing applies to every lifecycle gate (L0-understand to L10-lessons).
- Print the role table:
python "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" route show - Pick who does the task, in this order:
- a script that already does it: run it;
- deterministic and recurring work: write or extend one parameterized, tested script;
- bounded reading or running:
batuta:searcher(lookups) orbatuta:runner(runs), both on the cheap model; - implementation, tests or docs:
batuta:implementer,batuta:test-author,batuta:doc-writer, on the routine model; - the independent check of a gate:
batuta:adversary;batuta:revieweronly for a critical gate or a critical review; - reasoning, plans, specs and integration stay in the main session.
- Use the model the table names for the role and set it explicitly on the launch; an escalation to a stronger model is a decision you state, never a default.
- Stay inside the caps of
capsin batuta.json (concurrent agents, agents per task, reviews per task, spawn depth, cost). - After editing a role or an agent file, check them against each other:
python "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" route check - Next step for a task that goes to an agent: the
briefskill, then thedelegateskill.
On a non-zero exit: exit 1 means the configuration or an agent file is invalid or its model differs from its role: fix the file or the role, never launch on a table that fails. Exit 2 is a warning (for example no configuration file, so the defaults apply): read the WARN line and say it in your report.