diff --git a/src/workers/continuum-core/src/inference/llm_module_service.rs b/src/workers/continuum-core/src/inference/llm_module_service.rs
new file mode 100644
index 000000000..54542b6f0
--- /dev/null
+++ b/src/workers/continuum-core/src/inference/llm_module_service.rs
@@ -0,0 +1,350 @@
+//! `inference-llm` PR-2: `InferenceLlmModule` ServiceModule impl.
+//!
+//! PR-1 (#1387) shipped the typed event surface. PR-2 wires the
+//! ServiceModule that accepts InferenceRequest commands + emits
+//! the response events. The actual llama.cpp invoke lands in PR-3;
+//! PR-2 ships a STUB inference that returns canned tokens so the
+//! seam is testable end-to-end + downstream consumers
+//! (sentinel-observer, VDD harness) can wire to it today.
+//!
+//! ## What PR-2 ships
+//!
+//! - `InferenceLlmModule` struct implementing `ServiceModule`
+//! - `inference/llm/request` command — accepts InferenceRequest
+//!   JSON, runs the stub inference, returns InferenceComplete +
+//!   FirstTokenEmitted as JSON
+//! - Stub inference returns 3 canned tokens [1, 2, 3] with
+//!   `FinishReason::Stop`. Documented as PR-3 deferral.
+//! - Tests pin the wire contract: request → response correlation
+//!   via `requestId`, finish reason, token count, TTFT field
+//!
+//! ## What PR-2 does NOT ship (PR-3)
+//!
+//! - Real llama.cpp invocation (`LlamaCppAdapter` integration)
+//! - Tokenizer (composition_plan → prompt_tokens)
+//! - Token streaming via channels (PR-2 is request/response)
+//! - Bus-event subscription path (`artifact_subscriptions`)
+//! - ResidencyFault emission on missing-page (needs working-set
+//!   integration)
+//! - Runtime registration (separate wiring PR or registers when
+//!   PR-3 lands the real engine)
+
+use async_trait::async_trait;
+use serde_json::Value;
+use std::any::Any;
+
+use super::llm_module::{
+    FinishReason, FirstTokenEmitted, InferenceComplete, InferenceRequest,
+};
+use crate::runtime::module_context::ModuleContext;
+use crate::runtime::service_module::{
+    CommandResult, ModuleConfig, ModulePriority, ServiceModule,
+};
+
+/// Per-process implementation of `inference-llm`. ServiceModule
+/// trait impl that handles `inference/llm/request` commands.
+///
+/// PR-2 is stub-backed (canned tokens); PR-3 replaces the stub
+/// with the real `LlamaCppAdapter` invoke. The module's external
+/// contract (commands + response shapes) stays identical across
+/// the stub-vs-real transition — downstream consumers don't
+/// need to know which is running.
+pub struct InferenceLlmModule;
+
+impl InferenceLlmModule {
+    pub fn new() -> Self {
+        Self
+    }
+}
+
+impl Default for InferenceLlmModule {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// The command the module accepts. Producers (persona-cognition)
+/// send the InferenceRequest as JSON to this command and receive
+/// an InferenceComplete + FirstTokenEmitted bundle in the
+/// `CommandResult::Json` payload.
+pub const COMMAND_REQUEST: &str = "inference/llm/request";
+
+/// PR-2 stub inference output. Canned 3-token response so tests
+/// can pin the wire contract without requiring a real model load.
+/// PR-3 replaces with real generation.
+const STUB_COMPLETION_TOKENS: &[u32] = &[1, 2, 3];
+
+/// Result of one (stubbed) inference call: the complete event +
+/// the first-token event. The command returns both as a JSON
+/// object so the caller can publish them individually if it
+/// wants, or treat the pair atomically.
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct InferenceResponse {
+    pub complete: InferenceComplete,
+    pub first_token: FirstTokenEmitted,
+}
+
+#[async_trait]
+impl ServiceModule for InferenceLlmModule {
+    fn config(&self) -> ModuleConfig {
+        ModuleConfig {
+            name: "inference-llm",
+            priority: ModulePriority::High,
+            command_prefixes: &["inference/llm/"],
+            event_subscriptions: &[],
+            needs_dedicated_thread: false,
+            // Inference is single-flight per persona; the substrate
+            // serializes per-persona at a higher layer. PR-2's stub
+            // is reentrant + cheap; PR-3 may need a semaphore when
+            // the real backend lands. 0 = unlimited (module manages
+            // own concurrency).
+            max_concurrency: 0,
+            tick_interval: None,
+        }
+    }
+
+    async fn initialize(&self, _ctx: &ModuleContext) -> Result<(), String> {
+        Ok(())
+    }
+
+    async fn handle_command(
+        &self,
+        command: &str,
+        params: Value,
+    ) -> Result<CommandResult, String> {
+        match command {
+            COMMAND_REQUEST => self.handle_request(params).await,
+            other => Err(format!(
+                "inference-llm: unknown command '{other}' (expected '{COMMAND_REQUEST}')"
+            )),
+        }
+    }
+
+    fn as_any(&self) -> &dyn Any {
+        self
+    }
+}
+
+impl InferenceLlmModule {
+    /// Run the (stubbed) inference for one request. PR-3 replaces
+    /// the body with the real llama.cpp invoke path; the outer
+    /// shape (params → request, generate, complete + first-token)
+    /// stays the same.
+    async fn handle_request(&self, params: Value) -> Result<CommandResult, String> {
+        let request: InferenceRequest = serde_json::from_value(params)
+            .map_err(|e| format!("inference-llm: invalid InferenceRequest payload: {e}"))?;
+
+        // PR-2 stub: pretend we ran a model + emit canned tokens.
+        // PR-3 replaces this block with the real LlamaCppAdapter
+        // invoke. The InferenceComplete + FirstTokenEmitted wire
+        // shapes stay identical across the transition.
+        let complete = run_stub_inference(&request);
+        let first_token = first_token_for(&request, &complete);
+
+        let response = InferenceResponse {
+            complete,
+            first_token,
+        };
+        CommandResult::json(&response)
+    }
+}
+
+/// PR-2 stub inference. Returns the canned 3-token response with
+/// FinishReason::Stop. Useful for testing the request/response
+/// wire shape end-to-end without loading a real model.
+///
+/// Visibility: `pub(super)` so PR-3 can call it from a test that
+/// pins "stub vs real produce same wire shape" before swapping
+/// the implementation. Production code calls the trait method, not
+/// this directly.
+pub(super) fn run_stub_inference(request: &InferenceRequest) -> InferenceComplete {
+    InferenceComplete {
+        request_id: request.request_id,
+        persona: request.persona,
+        completion_tokens: STUB_COMPLETION_TOKENS.to_vec(),
+        finish_reason: FinishReason::Stop,
+        elapsed_ms: 1, // stub is fast; real engine fills in real time
+        tokens_generated: STUB_COMPLETION_TOKENS.len() as u32,
+    }
+}
+
+/// Build the FirstTokenEmitted event paired with a completion.
+/// PR-2's stub emits TTFT ≈ 0 (inference was instant). PR-3
+/// will capture the real first-token wall-clock from inside the
+/// streaming generation loop.
+pub(super) fn first_token_for(
+    request: &InferenceRequest,
+    complete: &InferenceComplete,
+) -> FirstTokenEmitted {
+    let _ = complete; // PR-3 will use complete.elapsed_ms for atomic-engine fallback
+    FirstTokenEmitted {
+        request_id: request.request_id,
+        persona: request.persona,
+        elapsed_us: 0, // stub: instant TTFT
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    //! Pin the ServiceModule contract + wire shape. PR-3 will add
+    //! integration tests that exercise the real engine; PR-2's
+    //! tests pin the seam.
+    use super::*;
+    use crate::genome::working_set::{ArtifactId, PersonaId};
+    use crate::inference::llm_module::{
+        CompositionPlan, GenerationBudget, InferenceRequestId, SamplingParams,
+    };
+    use uuid::Uuid;
+
+    fn sample_request() -> InferenceRequest {
+        InferenceRequest {
+            request_id: InferenceRequestId::new(Uuid::from_u128(42)),
+            persona: PersonaId::new(Uuid::from_u128(1)),
+            composition: CompositionPlan(ArtifactId::new(Uuid::from_u128(100))),
+            prompt_tokens: vec![10, 11, 12],
+            budget: GenerationBudget {
+                max_tokens: 100,
+                max_duration_ms: 5000,
+            },
+            sampling: SamplingParams::default(),
+            stop_sequences: vec![],
+        }
+    }
+
+    /// What this catches: module config reports its name +
+    /// command prefix. The registry uses this for routing; if the
+    /// prefix drifts, persona-cognition's request goes to the
+    /// wrong module.
+    #[test]
+    fn config_reports_name_and_command_prefix() {
+        let m = InferenceLlmModule::new();
+        let cfg = m.config();
+        assert_eq!(cfg.name, "inference-llm");
+        assert_eq!(cfg.command_prefixes, &["inference/llm/"]);
+        assert!(!cfg.needs_dedicated_thread);
+    }
+
+    /// What this catches: the module returns High priority. Local
+    /// inference is on the user-perceived critical path; the
+    /// scheduler treats this above Background but below Realtime
+    /// (which is reserved for audio/voice).
+    #[test]
+    fn config_priority_is_high() {
+        let m = InferenceLlmModule::new();
+        assert_eq!(m.config().priority, ModulePriority::High);
+    }
+
+    /// What this catches: COMMAND_REQUEST constant matches the
+    /// canonical wire name. Consumers refer to the constant via
+    /// `inference::llm_module_service::COMMAND_REQUEST` so renames
+    /// propagate; the literal string here is what drift on.
+    #[test]
+    fn command_request_has_canonical_string_value() {
+        assert_eq!(COMMAND_REQUEST, "inference/llm/request");
+    }
+
+    /// What this catches: handle_command routes the canonical
+    /// command to the stub inference; the response carries the
+    /// expected InferenceComplete + FirstTokenEmitted bundle.
+    /// End-to-end test of the seam.
+    #[tokio::test]
+    async fn handle_command_routes_request_to_stub_inference() {
+        let m = InferenceLlmModule::new();
+        let req = sample_request();
+        let params = serde_json::to_value(&req).unwrap();
+
+        let result = m.handle_command(COMMAND_REQUEST, params).await.unwrap();
+        match result {
+            CommandResult::Json(v) => {
+                let response: InferenceResponse = serde_json::from_value(v).unwrap();
+                assert_eq!(response.complete.request_id, req.request_id);
+                assert_eq!(response.complete.persona, req.persona);
+                assert_eq!(response.complete.completion_tokens, vec![1, 2, 3]);
+                assert_eq!(response.complete.finish_reason, FinishReason::Stop);
+                assert_eq!(response.complete.tokens_generated, 3);
+                assert_eq!(response.first_token.request_id, req.request_id);
+            }
+            CommandResult::Binary { .. } => panic!("expected Json response"),
+        }
+    }
+
+    /// What this catches: handle_command for an unknown command
+    /// returns a typed Err with the canonical-name in the message.
+    /// Loud rejection per Joel's never-swallow rule.
+    #[tokio::test]
+    async fn handle_command_unknown_returns_loud_error() {
+        let m = InferenceLlmModule::new();
+        let result = m
+            .handle_command("inference/llm/bogus", Value::Null)
+            .await;
+        match result {
+            Err(msg) => {
+                assert!(msg.contains("unknown command"));
+                assert!(msg.contains(COMMAND_REQUEST));
+                assert!(msg.contains("bogus"));
+            }
+            Ok(_) => panic!("unknown command must return Err"),
+        }
+    }
+
+    /// What this catches: handle_command for a malformed payload
+    /// returns a typed Err with the serde error context. Loud
+    /// rejection again — caller can debug from the message.
+    #[tokio::test]
+    async fn handle_command_invalid_payload_returns_typed_error() {
+        let m = InferenceLlmModule::new();
+        let result = m
+            .handle_command(COMMAND_REQUEST, serde_json::json!({"not": "a request"}))
+            .await;
+        match result {
+            Err(msg) => {
+                assert!(msg.contains("invalid InferenceRequest payload"));
+            }
+            Ok(_) => panic!("invalid payload must return Err"),
+        }
+    }
+
+    /// What this catches: the InferenceResponse bundle round-trips
+    /// through serde. Wire-stable shape for callers that decompose
+    /// the bundle into the two events for separate publishing.
+    #[tokio::test]
+    async fn inference_response_round_trips_through_serde() {
+        let req = sample_request();
+        let complete = run_stub_inference(&req);
+        let first_token = first_token_for(&req, &complete);
+        let response = InferenceResponse {
+            complete,
+            first_token,
+        };
+        let json = serde_json::to_string(&response).unwrap();
+        let back: InferenceResponse = serde_json::from_str(&json).unwrap();
+        assert_eq!(back.complete.request_id, req.request_id);
+        assert_eq!(back.first_token.request_id, req.request_id);
+    }
+
+    /// What this catches: object-safety + dyn dispatch. The
+    /// registry holds `Arc<dyn ServiceModule>`; if a future PR
+    /// adds a generic method, this construction fails.
+    #[tokio::test]
+    async fn module_is_object_safe_for_dyn_service_module() {
+        let module: std::sync::Arc<dyn ServiceModule> =
+            std::sync::Arc::new(InferenceLlmModule::new());
+        let cfg = module.config();
+        assert_eq!(cfg.name, "inference-llm");
+
+        let req = sample_request();
+        let params = serde_json::to_value(&req).unwrap();
+        let result = module
+            .handle_command(COMMAND_REQUEST, params)
+            .await
+            .unwrap();
+        match result {
+            CommandResult::Json(v) => {
+                let response: InferenceResponse = serde_json::from_value(v).unwrap();
+                assert_eq!(response.complete.request_id, req.request_id);
+            }
+            _ => panic!("expected Json"),
+        }
+    }
+}
diff --git a/src/workers/continuum-core/src/inference/mod.rs b/src/workers/continuum-core/src/inference/mod.rs
index 2c3dcd950..8b4502bdb 100644
--- a/src/workers/continuum-core/src/inference/mod.rs
+++ b/src/workers/continuum-core/src/inference/mod.rs
@@ -34,6 +34,7 @@ pub mod footprint_registry;
 pub mod kv_quant;
 pub mod llamacpp_adapter;
 pub mod llm_module;
+pub mod llm_module_service;
 pub mod lora;
 pub mod model;
 pub mod ort_providers;