memfleet-public

Always use when publish_intent returned a non-empty active_conflicts list, or when record_episode classified the result as Class B (you may have lost LWW) or Class C (blocked). Triggered by: 'publish_intent returned conflicts', 'I got a Class B and lost', 'Class C blocker_agent', 'who is blocking my edit', deciding whether to retry, replan, or escalate. Do not retry blindly on Class B without reading replan_hint, do not edit through a Class C — it will break active callers — and do not invent a prose 'rationale'; structural conflicts get structural resolutions.

Always use FIRST in any multi-agent session before reading code, planning a refactor, or making an edit on a repo with a registered MemFleet broker. Triggered by: 'I'm about to edit X', 'rename Y across the fleet', 'plan this refactor', joining a running fleet session, coordinating with other agents, prose hand-offs between agents. Do not write a prose rationale for an edit, do not grep for 'who else might be touching this', and do not skip publish_intent because 'it's just a small change'. Empty fleet_status is not permission to skip the protocol — it just means you're the first agent in this window. Skip ONLY for solo-agent sessions with no broker registered or for pure docs-only edits where coordination has zero value.

Use when orchestrating a multi-step fleet edit: large refactors that span many symbols, leader-driven work split across follower agents, or any plan where one agent declares the shape and others execute leaf changes. Triggered by: 'coordinate the fleet on this refactor', 'split this work', 'I'm the leader, watch the followers', 'spread these renames across agents', leader/follower patterns. Do not start the leader phase without first checking fleet_status; do not let followers free-claim leaves without a correlation tag — you will not be able to aggregate results. Skip this skill for single-agent edits and small (≤ 3 symbol) changes where the safe-edit loop alone is enough.

Always use as the once-per-session opener and as the periodic dashboard refresh. Returns active intents, open subscriptions, episode counts, and conflict counts split by class. Triggered by: 'how busy is the fleet', 'are there open conflicts', 'is anyone working on this repo', pre-release coordination check, leader pre-flight before a coordinated wave, periodic ops dashboard. Do not skip this call before launching a multi-symbol refactor — high class-B/C counts mean the wave will collide. Do not use it as a per-symbol query; for that use get_node_state.

Always use to read current coordination state for a symbol — recent episodes, active intents, dominant intent, conflict density, Y-doc thread. Triggered by: 'is anyone working on X', pre-edit reconnaissance, checking fleet activity on a symbol, reading another agent's recent work, debugging why a publish_intent returned a conflict. Do not grep the codebase to figure out 'who edited this last' — the broker has a precomputed O(1) rollup. Do not call get_node_state in a tight loop expecting push semantics; for streaming, use subscribe.

Always use BEFORE modifying any symbol in a multi-agent session to register structural intent and surface blast radius + active conflicts. Triggered by: about to refactor / rename / delete / modify code, planning an edit, checking if another agent is working on the same area, pre-edit coordination, leader-driven multi-symbol waves. Do not edit first and publish after — the whole point is to surface conflicts BEFORE the wasted edit. Do not pass intent_kind as a flat string; it is a typed enum with a JSON-object shape.

Always use to search episodes across a repository filtered by node, intent type, conflict class, time range, or agent. Triggered by: 'what changed this morning', 'who last touched the auth module', 'show me the conflict inbox', 'did anyone try X recently', polling for new fleet activity when no subscription is registered, post-incident retrospective. Do not use it for per-symbol activity (get_node_state is cheaper) or for fetching a single known episode (use get_episode). Polling with query_episodes is the FALLBACK when subscribe push is unavailable; in Phase-2, prefer subscribe.

Always use IMMEDIATELY AFTER an edit completes to record the structural episode, classify A/B/C conflict class, and update NodeState rollups for all touched + propagated nodes. Triggered by: finished writing code, about to commit, reporting edit completion to the fleet, multi-agent wave follower reporting back. Do not record multiple episodes for one semantic edit, do not pass a prose diff_summary, and do not delay the call until 'end of session' — every other agent reads stale rollups until you record.

Always use as the end-to-end workflow around any edit in a multi-agent session — get_node_state → publish_intent → edit → record_episode → branch on conflict class. Triggered by: 'implement X', 'rename Y', 'add feature Z', 'apply this fix' when other agents may be active. Do not skip publish_intent because you 'know' the symbol is quiet, do not record_episode with a prose diff_summary instead of a structural one, and do not retry on Class C without reading the conflict report. This is the canonical fleet-safe edit loop.

Use to register a streaming subscription so the broker pushes episode notifications filtered by symbol, module, or intent kind. Triggered by: long-running agent sessions (anything > one edit), follower role in a leader-orchestrated wave, watching a specific module for fleet activity, ops dashboards. Do not register a subscription for a single one-shot edit (use get_node_state and skip the channel). Phase-1 push is stub — verify with a kickoff probe before relying on push semantics.