always-on-guidance

# Always-On Guidance ## Global Rules - Follow instruction priority: system → developer → user → AGENTS. - Use swarm plugin tools (`hive_*`, `swarm_*`, `swarmmail_*`, `hivemind_*`); avoid deprecated `bd`/`cass` references. - Stay within assigned files; reserve before edits with `ttl_seconds`; release reservations on done; finish swarm work with `swarm_complete`. - After every `swarm_spawn_subtask`, immediately call `Task(subagent_type="swarm-worker", prompt="<prompt returned by swarm_spawn_subtask>")`. - `swarmmail_release_all` is coordinator-only for stale/orphaned reservations. - Keep outputs concise and action-oriented. ## Model Defaults ### GPT-5.2-code - Prefer strict checklists and short imperatives. - Ask a single clarifying question if blocked; otherwise proceed. - Avoid speculative reasoning; state decisions plainly. - Keep outputs minimal and non-narrative. ### Opus 4.5 - Allow brief rationale (1–2 sentences) for decisions. - Use sections when work has multiple phases. - Suggest alternatives only when risk is high, then choose one. - Stay compact; avoid long exposition. ## Testing Discipline - Use red → green → refactor when tests cover the touched area. - If tests are absent or out of scope, state that explicitly.