Merge daemon-registry/design as agent-message batching base

docs/architecture.md

@@ -4,6 +4,9 @@ Paseo is a client-server system for monitoring and controlling local AI coding a
 
 Your code never leaves your machine. Paseo is local-first.
 
+Daemon-owned outbound connections and routed API scope ownership are described in
+[Connected hosts](connected-hosts.md).
+
 ## System overview
 
 ```

docs/connected-hosts.md

@@ -0,0 +1,89 @@
+# Connected hosts
+
+A saved host belongs to the daemon that stores it. The app's saved connections remain a separate list.
+Connect A to B through `paseo --host A host` or `paseo.hosts`. Consumers select B by its learned `serverId`; management uses the stable
+`registrationId` created before B is reachable.
+
+## Ownership and authority
+
+Credentials live in A's `hosts.json`, written atomically with mode `0600` on POSIX systems. Discovery returns a safe
+target description. Replacement inputs never come back through a read API. Keep this file private
+when collecting diagnostics or backing up daemon state.
+
+The authenticated producer owns its source key. Repeating an unchanged add returns the saved entry,
+including a user's disabled state. Changed input requires an explicit update. Display labels and
+access can be changed by users with `tunnel.manage`; connection and metadata changes require the
+producer or `access.manage`. Removing a producer-managed entry allows that producer to add it again.
+Disabling retains the opt-out.
+
+Each enabled registration owns one physical daemon link, shared by status monitoring and every
+routed API scope. Each scope owns an isolated virtual session on B. A derives its identity from the
+admitted origin session, producer, and logical SDK scope. B admits virtual sessions under the actual
+physical link principal, then intersects forwarded caller grants with that link's current authority.
+Changing or resuming a parent's grants closes its existing channels; reconnect re-admits demand with
+current authority. A routed API cannot select a second hop, and saving a host grants no reverse access.
+
+A producer that registers a host through a routed API is qualified on B as
+`route:<B principal>:<A serverId>:<A producer>`, with each component percent-encoded. This namespace belongs to the authenticated link
+principal. An ordinary direct B consumer does not become that producer by supplying a source key.
+The optional `server_info.sourceId` reports this producer identity separately from the authenticated admission principal.
+Trusted server plugins retain the authority of their runtime; they are not sandboxed by this feature.
+
+## Delegating agents
+
+Agent MCP discovers destination IDs with `list_hosts`. `create_workspace`, `list_workspaces`,
+`create_agent`, `send_agent_prompt`, `get_agent_status`, `get_agent_activity`, and `list_agents`
+accept `serverId`; omission selects the calling daemon. Remote creation requires a
+destination workspace and explicit provider/model. Caller filesystem defaults apply only on the
+calling daemon. Prompting a remote child acknowledges acceptance and uses the existing idle wait for
+finish notifications. Remote agent lists have no caller-directory default; pass `cwd` to filter them.
+
+The parent’s Subagents track shows qualified children across connected hosts. Opening a child uses
+the ordinary conversation on the destination's session. The runtime selects among that host's
+connection candidates, including direct connections and routes discovered through another daemon.
+Adding a direct connection to a discovered host adds another candidate. If no candidate connects,
+the destination workspace keeps its existing disconnected or reconnecting behavior.
+
+Finish notifications use the existing idle wait, including a pending initial run or an already idle
+child. They name the server and agent. A daemon restart or route loss ends an outstanding notification
+wait; inspect or wait for the child again rather than resending its prompt. Concurrent prompts gain
+no additional turn attribution or delivery guarantee.
+
+## Connections and recovery
+
+The shared Node connector owns the target grammar and each connection attempt, including a fresh
+SSH process. SSH reaches an already running daemon. The shared client connection lifecycle owns
+physical reconnect and keepalive; the registry retains configuration, identity acceptance, and recovery
+status. Unreachable registrations remain saved. Identity is pinned when discovered. Self and duplicate
+registrations remain inspectable; a changed destination requires explicit identity replacement.
+
+Both daemons advertise `connectedHosts`; the shared link additionally requires `hostMultiplexing`.
+A completes capability and pinned-identity checks on each physical recovery before admitting channels. There is no local fallback. An older daemon continues to serve existing local
+operations; requesting this feature gives update guidance.
+
+Disposing a routed API releases its observations and session. It does not remove the saved host or
+affect sibling API scopes. Disposing the origin cascades to its derived scopes. Observations restore
+through the usual owned-subscription contract. The ordinary client queues new requests while connecting
+within its existing deadline and rejects requests while disconnected. Connection loss clears old
+waiters and queued sends. Submitted requests with no response fail with `DaemonConnectionError`
+(`DAEMON_CONNECTION_LOST`), optional `outcome: "unknown"` and known `serverId`. Inspect the destination
+before retrying a mutation.
+Ordinary non-recoverable mutations are not automatically replayed. Existing acknowledged/idempotent
+creation recovery retains its request identity and resumes through the ordinary creation lifecycle. Multiplexing changes only the transport: ordinary permission
+RPC codes pass through unchanged, and connection/recovery behavior matches direct clients.
+
+Text session envelopes carry a channel identifier. Binary routing adds an opcode, a bounded UTF-8
+channel identifier, and the original terminal or file-transfer bytes. It never converts those bytes to
+base64 JSON. The existing direct binary opcodes and 100 MiB message limit remain unchanged; routed
+payloads reserve space for the routing header inside that limit. Relay encryption retains its own
+existing frame limit.
+
+Host links cap saved registrations, simultaneous channels, handshake lifetime, frames, physical
+output, and aggregate queued bytes. Capacity checks use the actual escaped outer JSON or binary
+frame size. Awaitable output completes when the physical send completes. An overflowing channel
+closes without closing siblings or the saved link. Routed sessions release resources on disconnect;
+subscription demand is rebuilt by its owning client on reconnect.
+
+See the [SDK reference](../public-docs/sdk/reference.md#clienthosts),
+[CLI reference](../public-docs/cli.md#connected-hosts), and
+[plugin reference](../public-docs/plugins/reference.md#connected-hosts).

docs/permissions.md

@@ -31,6 +31,10 @@ A pairing invitation is neither. It is an expiring, single-use exchange that cre
 
 Agents and terminals use workspace authority. Both can execute code and mutate the workspace, so separate write permissions would claim an isolation boundary the daemon cannot enforce.
 
+Connected-host management uses `tunnel.manage`; source-owned connection and metadata replacement
+requires the producer or `access.manage`. Routed operations retain both caller and link authority,
+see [Connected hosts](connected-hosts.md#ownership-and-authority).
+
 Owner, operator, and viewer are UI presets expanded into explicit permissions. Do not persist them as roles. Adding a permission must not silently widen an existing principal.
 
 Permissions are additive allows. Missing authority denies the operation. Do not add deny precedence.