feat(server): release idle agent runtimes with provider safety checks

Require provider permission before releasing resumable runtimes so foreground idleness does not discard tracked background work. Preserve the saved conversation and serialize prompt admission with runtime closure; allow hosts to disable collection.

docs/agent-lifecycle.md

@@ -25,8 +25,41 @@ Reload releases the old runtime before resuming its durable session: an idle pro
 still own an exclusive writer. A close failure retains that runtime for cleanup and blocks the
 replacement. Once closure succeeds, a failed resume leaves the durable agent closed and retryable.
 
-Idle agents remain resident indefinitely. Runtime closure happens only through an explicit lifecycle
-action such as archive, replacement, reload, workspace teardown, or daemon shutdown.
+Idle resumable runtimes are collected after 30 minutes without activity, checked once a minute.
+Collection releases the process and preserves the agent record, conversation, relationships, and
+provider session. The next prompt resumes that session. Set `agents.idleRuntimeCleanup: false` in
+host config to disable automatic collection; this setting takes effect at daemon startup.
+
+Foreground runs, permissions and responses, queued deliveries, lifecycle operations, and working
+managed or native descendants prevent collection. An idle managed child can still own background
+work. Each provider must positively permit shutdown through `canStopRuntime()`; missing methods,
+failed probes, and probe timeouts retain ownership. Once protected work finishes, a fresh idle
+period starts. There is no deadline that overrides a provider veto.
+
+The collector rechecks daemon state after asynchronous probes and event draining, in the existing
+close/load lifecycle lane. A prompt admitted before closure prevents collection. A prompt arriving
+after closure starts waits for release and resumes the saved session.
+
+Provider checks follow the signals their runtimes expose:
+
+- Claude uses its process-local background roster and pending completion notifications. A fresh
+  process starts empty. CLI 2.1.292 or newer is known to support the roster; older binaries must
+  demonstrate support by emitting one. Unknown versions stay resident until then.
+- Codex queries background terminals for loaded threads, including native children, and protects
+  pending provider requests. App servers without the terminal-list RPC stay resident.
+- OpenCode checks active executions and descendants. V1 checks PTYs; v2 also checks shell and
+  inbox inventories. The shared server remains leased while a session vetoes collection.
+- Pi checks streaming, compaction, pending replies, and adapter-owned operations. Known native
+  subagent signals also protect their parent through the daemon's subagent store.
+- OMP additionally requires `get_state.isSettled: true` and `hasPendingAsyncWork: false`.
+  These protect jobs with unsuppressed completion delivery, pending deliveries, queued result
+  injection, and admitted or scheduled continuations that can re-wake the session. Older OMP
+  runtimes without either field stay resident. These signals are not an inventory of arbitrary
+  processes.
+- ACP has no complete signal for agent-owned background work, so its adapters stay resident.
+
+These checks do not discover arbitrary detached processes or work hidden by third-party
+extensions. A provider plugin opts in through the [runtime safety contract](../public-docs/plugins/providers.md#permit-idle-runtime-cleanup).
 
 A provider runtime can still die on its own — crash, OOM kill, host suspend. Work the agent parked
 inside that process dies with it: Claude Code's background Bash shells, `Monitor` watches, and

docs/plugins.md

@@ -498,6 +498,11 @@ Validate and apply it inside the provider; core does not know your option shape.
 See [provider options](custom-providers.md#provider-options) for configuration and
 merge semantics.
 
+Providers opt in to idle runtime collection through the per-session
+[runtime safety query](../public-docs/plugins/providers.md#permit-idle-runtime-cleanup).
+The daemon calls the same `AgentSession.canStopRuntime()` interface for built-ins and plugins;
+the plugin bridge transports that query through the existing provider request/event protocol.
+
 Agent refresh closes the current provider session and opens it again with current configuration and
 persistence. Re-read credentials and provider-owned configuration on `session.open`; consume the
 daemon launch from `connect` and the per-session env and MCP servers from `session.open`. There is