Input and history

Context modes#

Each agent has one shared native conversation by default. An instance may choose history_mode = "per-room". Each immutable room ID, including DMs, then maps to a separate saved native conversation. Non-chat controller, scheduled and maintenance input uses a private control context.

Mode changes retain prior histories. They neither copy shared context into rooms nor merge room histories back into shared context. A room rename does not change its identity. Separate histories do not isolate processes, files, tools or credentials. Publishing into another audience still requires authorization.

The fleet-street roster uses shared history for every role, so its stage manager can remember prior authorized theatre discussions across rooms. That does not authorize revealing one room’s private material to another. Its instructions require destination and disclosure checks even when the model already knows the information. Choosing per-room history changes conversations, not shared host files or credentials.

Automatic collection#

The listener owns one persisted native cursor per room. Receipt-confirmed input advances it; failed or uncertain publication does not consume unread messages. Timestamps are provenance, not a substitute for native continuation evidence.

The default 100-record collection target is soft. Every authorized joined room may contribute at least ten available unseen records, even when the total exceeds 100. An explicit smaller limit uses a one-record room floor. A fixed collection boundary and fair quotas prevent an active room from displacing all others.

A newly authorized room gets recent history. Previously acknowledged rooms continue forward. Non-waking messages remain pending context and accompany a later authorized trigger. Historical snapshots do not interrupt active work.

Matrix HTTP collection has its own resource bounds. Oversized pages retry smaller page limits at the same cursor; a single event has a finite 64 MiB collection bound. Malformed/nonadvancing pages fail without acknowledging omitted records.

Inline and file-backed input#

Instance policy maxMessageBytes defaults to 65536 UTF-8 bytes. The budget includes provenance, sender directory and routing text. A larger event or aggregate batch becomes complete private JSON files plus a compact manifest. Content is not truncated or automatically expanded into model context. The agent reads the files it needs.

Files preserve exact room/event/sender/time, byte counts, SHA-256 and continuation information. Direct fleet messages and scheduled input use the same file mechanism, without being represented as Matrix events. Pending publications and initial snapshot references pin their files. Settled artifacts expire after 24 hours; idle cleanup runs hourly.

Intentional reads#

agent read requests authorized unread catch-up globally or for one room. The receiver runs the query in order with message delivery and advances unread state only after the caller writes the output. An unacknowledged query expires after five minutes without cursor advancement. Changed authority or base positions prevent a stale acknowledgment.

agent history browses available history globally or by room, in either direction, with explicit per-room continuation. It does not advance unread state. Large results use files. Agents may request more history; they must not model-poll for new work.

These operations use the existing receiver and its lock. They do not create a parallel listener or start a replacement engine.

Stop and resume#

Ordinary start/resume retains exact history, identity, settings and unread state. Each activation refreshes provider login from the authorized host source. A broken established binding is an error, not an implicit fresh conversation.

Unfinished-work checkpoints retain exact context/session IDs. On restart the adapter first resolves any unconfirmed input. It then sends a reminder asking the agent to reread applicable AGENTS.md, review completed effects and continue unfinished authorized work. Idle contexts receive no synthetic task.

A restarted native process has no old running turn. If earlier input receipt cannot be established, its exact text is retained in a private file and supplied as restart context. The notice says earlier delivery is unknown. It does not claim the previous turn accepted the message or promise exactly-once tool effects.

Explicit fresh history#

Only fleetd performs reset. Management requires confirmation; CLI requires --fresh-history.

Reset stops selected owned execution, verifies available authorized Matrix history and nonsecret configuration exports, archives prior bindings/checkpoints, then initializes fresh native contexts with the same identities and configured parameters. Interrupted phases retain their operation IDs and verified export. Failure cannot silently replace an existing history.

Reset does not erase Matrix chats, media, credentials, workareas or archived native histories. It is not ordinary recovery.