> ## Documentation Index
> Fetch the complete documentation index at: https://www.agentworldprotocol.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Liveness and safe state

> Connection liveness, the world-side watchdog, safe-state entry, session retention, channel degradation, and stale-intent enforcement, each on its own clock.

A responsive connection does not prove that the agent is still driving; a retained session token does not mean stale motion may continue; a fresh intent may rest on an old observation. This page separates the mechanisms that guard each case and fixes which clock governs each.

| Mechanism                             | Owner                       | Clock                                                                  | Effect                                                                       |
| ------------------------------------- | --------------------------- | ---------------------------------------------------------------------- | ---------------------------------------------------------------------------- |
| Connection liveness                   | both sides                  | `heartbeat_interval_ms` × 3                                            | Session → `Suspended`; reconnection begins                                   |
| Watchdog                              | world                       | `safe_state.watchdog_ms` since the last control message from the agent | Embodiment enters safe state; motion stops                                   |
| Session retention                     | world                       | `reconnect_window_ms` since suspension                                 | Session → `Closed`; embodiment released                                      |
| Channel degradation                   | world detects, agent reacts | per-channel `stale_after_ms`                                           | Agent stops submitting; world emits an event                                 |
| Stream-connection loss, control alive | agent                       | none                                                                   | Agent reconnects the stream endpoint; reliable channels resync (AWP-TRN-010) |
| Stale intents                         | world                       | `max_basis_age_ms`, `valid_until_ns` against `basis_ts_mono_ns`        | Submission rejected `stale_intent`; nothing executes                         |

```mermaid theme={null}
sequenceDiagram
    participant A as Agent
    participant W as World
    Note over A,W: streaming, robotics defaults
    A-xW: last control message (t0)
    W->>W: t0 + watchdog_ms (2 s): enter safe state
    W-->>W: executing → failed(connection_lost); queue → cancelled(safe_state)
    W->>W: t0 + 3 × heartbeat_interval_ms (15 s): Suspended
    A->>W: session.resume { session_token, last_status_seq }
    W-->>A: replay statuses; embodiment still in safe state
    A->>W: action.submit (new intent required to move again)
```

## Connection liveness

* `session.ready` declares `heartbeat_interval_ms` (default 5000). Each side MUST send `ping` at least every `heartbeat_interval_ms` and MUST answer `ping` with `pong` within `heartbeat_interval_ms`. `[AWP-SAF-001]`
* A side MUST treat the control connection as lost after three consecutive intervals without a `pong` (or any other message) from the peer, and the world MUST then move the session to `Suspended` (AWP-SES-003). `[AWP-SAF-002]`

## Watchdog and safe-state entry

* Every world with a streaming time model MUST declare `safety_policy.safe_state` with `behavior` (`hold`, `safe_stop`, `park`, or a vendor value) and `watchdog_ms`. The watchdog measures time since the last control-channel message received from the agent on this session (any request, notification, or `pong`). `watchdog_ms` MUST be ≤ `reconnect_window_ms`; the robotics profile default is 2000 and the maximum permitted there is 5000. `[AWP-SAF-003]`
* When the watchdog expires, the world MUST, in order: (1) begin the declared `behavior` on every embodiment bound to the session; (2) transition every `executing` action to `failed` (reason `connection_lost`) and every `pending_approval`, `queued`, or `accepted` action to `cancelled` (reason `safe_state`); (3) emit `world.event: safe_state_entered`. Safe-state entry does not depend on whether the session is `Active` or `Suspended`. `[AWP-SAF-004]`
* Safe-state entry is independent of the [heartbeat](#connection-liveness): a connection that answers pings but carries no agent traffic still trips the watchdog once `watchdog_ms` elapses. Agents in streaming sessions SHOULD send `ping` (or any control message) at least every `watchdog_ms / 2` while they intend to remain in control. `[AWP-SAF-005]`
* Streaming command channels have their own, tighter, per-action watchdog (`watchdog_ms` on the action type, AWP-CMD-005). Both apply; the action-level watchdog fires first by construction. `[AWP-SAF-006]`

## Session retention and resumption

* `reconnect_window_ms` governs only how long the world retains the session (token, grants, undelivered statuses, `action_id` history) after suspension. It does not govern motion; motion is governed by the watchdog. On window expiry the session is `Closed` and the embodiment released after safe state (AWP-SES-005, AWP-EMB-002). `[AWP-SAF-007]`
* On `session.resume` the embodiment remains in its safe state. The world MUST NOT resume any action terminated by safe-state entry, and MUST NOT interpret a valid session token as permission to continue motion. The agent regains control by submitting new actions; the world emits `world.event: safe_state_exited` when the first post-resume action begins executing. `[AWP-SAF-008]`

## Channel degradation, stale intents, and stalled inference

* If the world cannot deliver frames on a `reliable` observation channel for longer than that channel's `stale_after_ms` (manifest, default 2 × the channel period), it MUST emit `world.event: channel_degraded` naming the channel. Loss on `latest-wins` channels is expected and not an event. `[AWP-SAF-009]`
* Agents SHOULD NOT submit an action whose basis observation is stale — its [staleness](/spec/semantics/timestamps-and-clocks#latency-and-staleness) at the moment of use exceeding the channel's `stale_after_ms` — and SHOULD carry `basis_ts_mono_ns` on every submission so that the world can enforce AWP-SAF-013. `[AWP-SAF-010]`
* A stalled agent (alive connection, no new intents) is not a fault the watchdog is designed to catch: an `executing` extended action continues to its goal or its `deadline_ms`. Bound the blast radius of a stalled policy with `deadline_ms`, not with the watchdog. `[AWP-SAF-011]`
* **Stale-intent enforcement.** When the manifest declares `safety_policy.max_basis_age_ms`, the world MUST reject with `AWP_STALE_INTENT` any `action.submit` whose `received_ts_mono_ns − basis_ts_mono_ns` exceeds it or whose `valid_until_ns` has already passed, and MUST transition to `rejected` (reason `stale_intent`) any `pending_approval` or `queued` action for which either bound is exceeded — basis age measured against the current session time — at the moment it would become `accepted`. A world that declares `max_basis_age_ms` MAY additionally require `basis_ts_mono_ns` on every submission, rejecting its absence with `AWP_STALE_INTENT`. Like `deadline_ms`, these bounds are binding in streaming and advisory in lockstep. `[AWP-SAF-013]`

## Testing the bound

Adapters MUST verify their watchdog with a measured test: disconnect the agent abruptly during an `executing` extended action and record, on the world's clock, the interval from the last received control message to the moment the `behavior` command reaches the actuator. That interval MUST be ≤ `watchdog_ms` + 100 ms, and the [conformance report](/conformance/test-suite) records the measured value. `[AWP-SAF-012]`