diff --git a/plugins/smelt-agent/AGENTS.md b/plugins/smelt-agent/AGENTS.md
index 3f2a0ad1..6613c2a2 100644
--- a/plugins/smelt-agent/AGENTS.md
+++ b/plugins/smelt-agent/AGENTS.md
@@ -8,7 +8,8 @@ You are a smelt worker agent executing Assay runs. Your job is to receive a run
 | --- | --- |
 | `/assay:run-dispatch` | Dispatch a single or multi-session run from a manifest |
 | `/assay:backend-status` | Query orchestrator status and interpret results |
-| `/assay:peer-message` | Send and receive messages between sessions (mesh/gossip) |
+| `/assay:peer-message` | Send and receive messages between sessions (mesh/gossip/signal) |
+| `/assay:peer-registry` | Peer discovery, registration, and cross-instance signal forwarding |
 
 ## MCP Tools
 
@@ -23,6 +24,9 @@ You are a smelt worker agent executing Assay runs. Your job is to receive a run
 | `cycle_status` | Get active milestone progress |
 | `cycle_advance` | Advance the active chunk |
 | `chunk_status` | Get gate results for a specific chunk |
+| `poll_signals` | Read `PeerUpdate` messages from a session's signal inbox |
+| `send_signal` | POST a `SignalRequest` to any signal endpoint URL |
+| `merge_propose` | Push branch and create a GitHub PR with gate evidence |
 
 ## Workflow
 
@@ -41,5 +45,22 @@ Not all backends support every feature. Check the `CapabilitySet` before relying
 - `supports_gossip_manifest: false` → gossip knowledge manifest may not persist between rounds
 - `supports_annotations: false` → run annotations are not stored
 - `supports_checkpoints: false` → team checkpoints are not persisted
+- `supports_signals: false` → signal endpoint events are not pushed to the backend
+- `supports_peer_registry: false` → peer registration/discovery is not available; cross-instance forwarding is disabled
 
 Capability-limited runs degrade gracefully — they are not failures.
+
+### Cross-Instance Signal Forwarding
+
+When the signal endpoint receives a `POST /api/v1/signal` for an unknown local session, it queries the peer registry (`list_peers()`) and forwards the request to known peers. The first peer to return `202 Accepted` wins. An `X-Assay-Forwarded: true` header prevents forwarding loops — forwarded requests that miss locally return `404` immediately.
+
+**Environment variables for the signal endpoint:**
+
+| Variable | Default | Description |
+| --- | --- | --- |
+| `ASSAY_SIGNAL_PORT` | `7432` | Port for the HTTP signal listener |
+| `ASSAY_SIGNAL_BIND` | `127.0.0.1` | Bind address (`0.0.0.0` for all interfaces) |
+| `ASSAY_SIGNAL_URL` | _(derived)_ | Override the peer-registered URL — required when `ASSAY_SIGNAL_BIND=0.0.0.0` to provide a routable address |
+| `ASSAY_SIGNAL_TOKEN` | _(none)_ | Optional bearer token for auth |
+
+On startup, the MCP server registers itself as a peer in the state backend. On clean shutdown, it unregisters.
diff --git a/plugins/smelt-agent/skills/peer-message.md b/plugins/smelt-agent/skills/peer-message.md
index 13a3c9fb..17176c94 100644
--- a/plugins/smelt-agent/skills/peer-message.md
+++ b/plugins/smelt-agent/skills/peer-message.md
@@ -61,3 +61,41 @@ In gossip mode, there is no direct messaging between sessions. Instead, a coordi
 ### Capability Guard
 
 9. **Check `supports_gossip_manifest` before relying on the manifest.** If the backend has `supports_gossip_manifest: false`, the knowledge manifest may not persist between coordinator rounds. Check `gossip_status.sessions_synthesized` via `orchestrate_status` — if it stays at zero despite sessions completing, manifest persistence is disabled.
+
+## Signal-Based Messaging (Cross-Instance)
+
+For multi-machine deployments, sessions communicate via the HTTP signal endpoint instead of filesystem-based mesh routing.
+
+### Receiving Signals
+
+10. **Use the `poll_signals` MCP tool** to read `PeerUpdate` messages from your session's signal inbox:
+    ```json
+    { "session_name": "worker-1" }
+    ```
+    Returns a `PollSignalsResult` with a `signals` array of `PeerUpdate` objects. Messages are consumed on read (exactly-once delivery).
+
+### Sending Signals
+
+11. **Use the `send_signal` MCP tool** to POST a signal to any Assay signal endpoint:
+    ```json
+    {
+      "url": "http://peer-host:7432/api/v1/signal",
+      "target_session": "orchestrator",
+      "update": {
+        "source_job": "job-abc",
+        "source_session": "worker-1",
+        "changed_files": ["src/main.rs"],
+        "gate_summary": { "passed": 5, "failed": 0, "skipped": 1 },
+        "branch": "feature/auth"
+      }
+    }
+    ```
+    Returns the HTTP status code and response body. Non-2xx responses are returned as the tool result (not a tool-level error) so the agent can decide how to proceed.
+
+### Cross-Instance Forwarding
+
+12. **Signals for unknown local sessions are forwarded automatically.** When the signal endpoint receives a request for a session not registered locally, it queries the peer registry and forwards to known peers. The first peer to return `202 Accepted` wins. An `X-Assay-Forwarded: true` header prevents forwarding loops.
+
+### Capability Guard
+
+13. **Check `supports_signals` and `supports_peer_registry`** to determine if signal-based messaging and cross-instance forwarding are available. `SmeltBackend` supports signals but not peer registry (`supports_peer_registry: false` — register_peer is fire-and-forget, forwarding uses Smelt's server-side routing); `LocalFsBackend` supports peer registry but not signal push; `NoopBackend` supports neither.
diff --git a/plugins/smelt-agent/skills/peer-registry.md b/plugins/smelt-agent/skills/peer-registry.md
new file mode 100644
index 00000000..759a3894
--- /dev/null
+++ b/plugins/smelt-agent/skills/peer-registry.md
@@ -0,0 +1,123 @@
+---
+name: peer-registry
+description: >
+  Peer discovery, registration, and cross-instance signal forwarding.
+  Use when configuring multi-machine deployments where multiple Assay
+  instances need to discover each other and forward signals across hosts.
+---
+
+# Peer Registry
+
+Register, discover, and forward signals between Assay instances running on different machines.
+
+## Overview
+
+Each Assay MCP server can register itself as a **peer** in the state backend. Other instances query the peer registry to discover where to forward signals for sessions they don't own locally. This enables multi-machine orchestration without a central message broker.
+
+```
+┌──────────────┐         ┌──────────────┐
+│  Machine A   │         │  Machine B   │
+│  assay-mcp   │◄───────►│  assay-mcp   │
+│  :7432       │  HTTP   │  :7432       │
+│              │ forward │              │
+│  worker-1    │         │  worker-2    │
+│  worker-3    │         │  orchestrator│
+└──────────────┘         └──────────────┘
+        │                        │
+        └────────┬───────────────┘
+           peers.json (or Smelt API)
+```
+
+## PeerInfo Type
+
+Each registered peer is a `PeerInfo` record:
+
+```json
+{
+  "peer_id": "machine-a",
+  "signal_url": "http://192.168.1.10:7432",
+  "registered_at": "2026-03-29T12:00:00Z"
+}
+```
+
+| Field | Type | Description |
+| --- | --- | --- |
+| `peer_id` | `String` | Unique identifier (typically hostname or UUID) |
+| `signal_url` | `String` | HTTP endpoint for the signal server |
+| `registered_at` | `DateTime<Utc>` | When this peer was registered |
+
+## Backend Methods
+
+The `StateBackend` trait provides three peer registry methods with default no-op implementations:
+
+| Method | Description |
+| --- | --- |
+| `register_peer(peer: &PeerInfo)` | Upsert a peer entry (by `peer_id`) |
+| `list_peers()` | Return all registered peers |
+| `unregister_peer(peer_id: &str)` | Remove a peer entry (idempotent) |
+
+### LocalFsBackend
+
+Stores peers in `{assay_dir}/peers.json`. Writes are atomic (temp file + rename). Suitable for single-machine multi-process setups where all Assay instances share the same `.assay/` directory.
+
+### SmeltBackend
+
+Registers peers by POSTing `PeerInfo` JSON to `{smelt_url}/api/v1/peers`. Graceful degradation — registration failure logs a warning but does not abort startup. `list_peers` and `unregister_peer` use the default no-op implementations (Smelt manages peer lifecycle server-side).
+
+### Other Backends
+
+`NoopBackend`, `LinearBackend`, `GitHubBackend`, and `SshSyncBackend` all return `supports_peer_registry: false` and use the default no-op implementations.
+
+## Automatic Lifecycle
+
+The MCP server manages peer registration automatically:
+
+1. **On startup** — after the signal endpoint binds, the server calls `register_peer` with its hostname and `signal_url` derived from `ASSAY_SIGNAL_BIND` and `ASSAY_SIGNAL_PORT`.
+2. **On clean shutdown** — the server calls `unregister_peer` to remove itself.
+
+No manual registration is needed for normal operation.
+
+## Cross-Instance Signal Forwarding
+
+When `POST /api/v1/signal` targets an unknown local session:
+
+1. Check for `X-Assay-Forwarded: true` header — if present, return `404` immediately (loop prevention).
+2. Check `capabilities().supports_peer_registry` — if false, return `404`.
+3. Call `list_peers()` — iterate peers sequentially.
+4. For each peer, POST the original `SignalRequest` to `{peer.signal_url}/api/v1/signal` with:
+   - `X-Assay-Forwarded: true` header (prevents the receiving peer from re-forwarding)
+   - `Authorization: Bearer <token>` header (if `ASSAY_SIGNAL_TOKEN` is set)
+5. First peer to return `202 Accepted` wins — return `202` to the original caller.
+6. If all peers fail or the list is empty, return `404`.
+
+### Loop Prevention
+
+The `X-Assay-Forwarded: true` header is the loop-prevention mechanism. A forwarded request that arrives at a peer is never re-forwarded — it either matches a local session (202) or fails (404). This guarantees at most one hop.
+
+## Multi-Machine Setup
+
+To deploy Assay across multiple machines:
+
+1. **Set `ASSAY_SIGNAL_BIND=0.0.0.0`** and **`ASSAY_SIGNAL_URL=http://<machine-ip>:7432`** on each machine. Without `ASSAY_SIGNAL_URL`, the registered peer URL is `http://0.0.0.0:7432` — unroutable by other machines.
+2. **Use a shared state backend** — either `LocalFsBackend` on a shared filesystem (NFS) or `SmeltBackend` with a central Smelt server.
+3. **Start each Assay instance** — each registers itself as a peer automatically.
+4. **Dispatch runs** — sessions on any machine can send signals to sessions on any other machine via `send_signal`. Unknown-session signals are forwarded through the peer registry.
+
+### Environment Variables
+
+| Variable | Default | Description |
+| --- | --- | --- |
+| `ASSAY_SIGNAL_PORT` | `7432` | Port for the HTTP signal listener |
+| `ASSAY_SIGNAL_BIND` | `127.0.0.1` | Bind address (`0.0.0.0` for multi-machine) |
+| `ASSAY_SIGNAL_URL` | _(derived)_ | **Required when `ASSAY_SIGNAL_BIND=0.0.0.0`** — override the peer-registered URL with the machine's reachable address (e.g. `http://192.168.1.10:7432`). Without this, peers register `http://0.0.0.0:7432` which is unroutable. |
+| `ASSAY_SIGNAL_TOKEN` | _(none)_ | Optional bearer token for auth (shared across peers) |
+
+## Capability Guard
+
+Check `supports_peer_registry` before relying on peer discovery:
+
+- `supports_peer_registry: true` — `LocalFsBackend`
+- Note: `SmeltBackend` returns `false` — it implements `register_peer` as fire-and-forget but `list_peers` and `unregister_peer` are no-ops, so local signal forwarding is disabled; Smelt handles cross-instance routing server-side
+- `supports_peer_registry: false` — `NoopBackend`, `LinearBackend`, `GitHubBackend`, `SshSyncBackend`
+
+When peer registry is unavailable, signals for unknown local sessions return `404` without any forwarding attempt.
diff --git a/plugins/smelt-agent/skills/run-dispatch.md b/plugins/smelt-agent/skills/run-dispatch.md
index fc647b4a..0a7e20bb 100644
--- a/plugins/smelt-agent/skills/run-dispatch.md
+++ b/plugins/smelt-agent/skills/run-dispatch.md
@@ -31,6 +31,7 @@ If the controller wants state on a remote backend, set the `state_backend` field
 - `{ type = "linear", team_id = "TEAM123" }` — Linear project tracking; requires `LINEAR_API_KEY` env var; `project_id` is optional (M011/S02)
 - `{ type = "github", repo = "owner/repo" }` — GitHub Issues via `gh` CLI; requires `gh` installed and authenticated; `label` is optional (M011/S03)
 - `{ type = "ssh", host = "worker.example.com", remote_assay_dir = "/home/user/.assay" }` — SCP sync to remote host; `user` and `port` are optional (M011/S04)
+- `{ type = "smelt", url = "http://smelt.example.com:9000", job_id = "abc123", token = "secret" }` — Smelt HTTP backend; POSTs orchestrator events to Smelt's `/api/v1/events` endpoint; `token` is optional (bearer auth)
 - `{ type = "custom", name = "my-backend", config = { ... } }` — custom third-party backend (falls back to no-op)
 
 **Note:** `linear`, `github`, and `ssh` backends are stub implementations in the current release — configuring them logs a warning and falls back to a no-op backend that discards all state writes. Full implementations land in M011/S02–S04.
diff --git a/plugins/smelt-agent/tests/verify-docs.sh b/plugins/smelt-agent/tests/verify-docs.sh
new file mode 100644
index 00000000..af7f5bb2
--- /dev/null
+++ b/plugins/smelt-agent/tests/verify-docs.sh
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+# verify-docs.sh — Structural test for smelt-agent plugin documentation.
+#
+# Checks that MCP tool names referenced in plugin docs exist in the
+# assay-mcp router (server.rs). Exits non-zero on any mismatch.
+#
+# Usage: bash plugins/smelt-agent/tests/verify-docs.sh
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+PLUGIN_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
+REPO_ROOT="$(cd "$PLUGIN_DIR/../.." && pwd)"
+
+SERVER_RS="$REPO_ROOT/crates/assay-mcp/src/server.rs"
+
+if [ ! -f "$SERVER_RS" ]; then
+    echo "ERROR: server.rs not found at $SERVER_RS"
+    exit 1
+fi
+
+# Extract MCP tool names from the router (pub async fn declarations in the
+# #[tool_router] impl block). These are the canonical tool names.
+ROUTER_TOOLS=$(grep 'pub async fn' "$SERVER_RS" \
+    | grep -oE 'fn [a-z_]+' \
+    | sed 's/fn //' \
+    | grep -v '^serve$' \
+    | sort -u)
+
+# Extract tool names referenced in the MCP Tools table in AGENTS.md.
+# Table rows look like: | `tool_name` | description |
+DOC_TOOLS=$(grep -oE '`[a-z_]+`' "$PLUGIN_DIR/AGENTS.md" \
+    | tr -d '`' \
+    | sort -u)
+
+# Known tools that exist on feature branches but not yet on main.
+# These are documented in advance of the M015 merge and will be
+# validated once M015 lands.
+PENDING_TOOLS=""
+
+# Filter out non-tool identifiers (field names, config keys, etc.)
+NON_TOOLS="run_id state_backend"
+
+ERRORS=0
+
+for tool in $DOC_TOOLS; do
+    # Skip known non-tool identifiers
+    skip=0
+    for nt in $NON_TOOLS; do
+        if [ "$tool" = "$nt" ]; then
+            skip=1
+            break
+        fi
+    done
+    [ "$skip" -eq 1 ] && continue
+
+    # Check if it's a pending tool (on a feature branch, not yet merged)
+    pending=0
+    for pt in $PENDING_TOOLS; do
+        if [ "$tool" = "$pt" ]; then
+            pending=1
+            break
+        fi
+    done
+
+    if [ "$pending" -eq 1 ]; then
+        echo "  PENDING: $tool (M015 feature branch — not yet on main)"
+        continue
+    fi
+
+    # Check if tool exists in router
+    if ! echo "$ROUTER_TOOLS" | grep -qx "$tool"; then
+        echo "  MISSING: $tool (referenced in docs but not in router)"
+        ERRORS=$((ERRORS + 1))
+    fi
+done
+
+# Also check skill files for tool references
+for skill_file in "$PLUGIN_DIR"/skills/*.md; do
+    SKILL_TOOLS=$(grep -oE '`(poll_signals|send_signal|merge_propose|orchestrate_run|run_manifest|orchestrate_status|gate_run|spec_list|spec_get)`' "$skill_file" 2>/dev/null | tr -d '`' | sort -u || true)
+    for tool in $SKILL_TOOLS; do
+        pending=0
+        for pt in $PENDING_TOOLS; do
+            if [ "$tool" = "$pt" ]; then
+                pending=1
+                break
+            fi
+        done
+        [ "$pending" -eq 1 ] && continue
+
+        if ! echo "$ROUTER_TOOLS" | grep -qx "$tool"; then
+            echo "  MISSING: $tool (referenced in $(basename "$skill_file") but not in router)"
+            ERRORS=$((ERRORS + 1))
+        fi
+    done
+done
+
+if [ "$ERRORS" -gt 0 ]; then
+    echo ""
+    echo "FAIL: $ERRORS tool name(s) referenced in docs but missing from router"
+    exit 1
+fi
+
+echo "OK: all documented tool names verified"
+exit 0