feat(agent): orchestration v3 — surface MCP model-agnostic (M0→M4) — §14.3

Expose l'orchestration IdeA comme serveur MCP par-dessus le même
OrchestratorService::dispatch, avec repli fichier .ideai/requests pour les
CLI sans MCP. v3 réduite à la surface MCP : la messagerie inter-agents et la
corrélation requête↔réponse étaient déjà résolues par §17 (send_blocking).

- M0 capacité MCP sur le profil (McpCapability/McpConfigStrategy/McpTransport)
- M1 injection conf MCP au LaunchAgent + prose adaptée selon la surface
- M2 serveur/adapter MCP (JSON-RPC 2.0 maison ; outils idea_*) + ListAgents
- M3 câblage par projet (registre mcp_servers jumeau du watcher)
- M4 observabilité UI : OrchestratorRequestProcessed.source = file|mcp + badge

Trois portes d'entrée (fichier, MCP, UI) → un seul dispatch ; aucun nouveau
port applicatif ; MCP confiné à l'adapter infra. Tous lots verts (cycle §3).
Cadrage : .ideai/briefs/orchestration-v3-cadrage.md ; ARCHITECTURE.md §14.3.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

crates/infrastructure/src/lib.rs

@@ -32,6 +32,7 @@ pub use fs::LocalFileSystem;
 pub use git::Git2Repository;
 pub use id::UuidGenerator;
 pub use inspector::ClaudeTranscriptInspector;
+pub use orchestrator::mcp::{McpServer, MemoryTransport, StdioTransport};
 pub use orchestrator::{
     process_request_file, FsOrchestratorWatcher, OrchestratorResponse, OrchestratorWatchHandle,
     REQUESTS_SUBDIR,

crates/infrastructure/src/orchestrator/mcp/jsonrpc.rs

@@ -0,0 +1,181 @@
+//! Minimal **JSON-RPC 2.0** message model + transport seam for the IdeA MCP
+//! adapter (spike **S-MCP**).
+//!
+//! ## Why hand-rolled JSON-RPC instead of an MCP crate
+//!
+//! MCP (Model Context Protocol) rides on JSON-RPC 2.0: a server advertises its
+//! tools via `tools/list` and runs them via `tools/call`, after an `initialize`
+//! handshake. The surface IdeA needs is exactly those three methods — no
+//! resources, no prompts, no sampling. The Rust MCP crate ecosystem
+//! (`rmcp`, `mcp-sdk`, …) is young, churny, and drags in a concrete async
+//! transport/runtime stack that is awkward to drive **without a socket or child
+//! process** in a unit test. The cadrage (S-MCP) explicitly permits implementing
+//! "the strict JSON-RPC 2.0 minimum (`initialize`, `tools/list`, `tools/call`)"
+//! ourselves when that buys **testability and zero-network**. We take that path:
+//! the whole protocol lives in this adapter, behind a tiny [`Transport`] trait, so
+//! tests speak it over an in-memory transport with no I/O at all.
+//!
+//! Everything here is JSON-RPC plumbing — it never knows what an
+//! `OrchestratorCommand` is. The mapping to IdeA semantics lives in
+//! [`super::server`].
+
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+
+/// The JSON-RPC protocol version string every message must carry.
+pub const JSONRPC_VERSION: &str = "2.0";
+
+/// A request id as defined by JSON-RPC 2.0: a string, a number, or null. We keep
+/// it as a raw [`Value`] and echo it back verbatim on the matching response so the
+/// transport-level correlation (Décision 2: corrélation portée par le transport)
+/// is preserved exactly, whatever the client chose.
+pub type RequestId = Value;
+
+/// An incoming JSON-RPC request (or notification, when `id` is absent).
+#[derive(Debug, Clone, Deserialize, Serialize, PartialEq)]
+pub struct JsonRpcRequest {
+    /// Must equal [`JSONRPC_VERSION`].
+    pub jsonrpc: String,
+    /// Correlation id. Absent ⇒ the message is a *notification* (no reply owed).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub id: Option<RequestId>,
+    /// The method name (`initialize`, `tools/list`, `tools/call`, …).
+    pub method: String,
+    /// Method parameters; shape depends on `method`.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub params: Option<Value>,
+}
+
+/// An outgoing JSON-RPC response: exactly one of `result` / `error` is set.
+#[derive(Debug, Clone, Deserialize, Serialize, PartialEq)]
+pub struct JsonRpcResponse {
+    /// Must equal [`JSONRPC_VERSION`].
+    pub jsonrpc: String,
+    /// Echoes the request id (null for errors that could not be correlated).
+    pub id: RequestId,
+    /// Success payload (mutually exclusive with [`Self::error`]).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub result: Option<Value>,
+    /// Failure payload (mutually exclusive with [`Self::result`]).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub error: Option<JsonRpcError>,
+}
+
+impl JsonRpcResponse {
+    /// Builds a success response echoing `id`.
+    #[must_use]
+    pub fn success(id: RequestId, result: Value) -> Self {
+        Self {
+            jsonrpc: JSONRPC_VERSION.to_owned(),
+            id,
+            result: Some(result),
+            error: None,
+        }
+    }
+
+    /// Builds an error response echoing `id`.
+    #[must_use]
+    pub fn error(id: RequestId, error: JsonRpcError) -> Self {
+        Self {
+            jsonrpc: JSONRPC_VERSION.to_owned(),
+            id,
+            result: None,
+            error: Some(error),
+        }
+    }
+}
+
+/// A JSON-RPC error object.
+#[derive(Debug, Clone, Deserialize, Serialize, PartialEq, Eq)]
+pub struct JsonRpcError {
+    /// One of the standard JSON-RPC codes (see [`error_codes`]).
+    pub code: i32,
+    /// Short human-readable description.
+    pub message: String,
+    /// Optional structured detail.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub data: Option<Value>,
+}
+
+impl JsonRpcError {
+    /// Builds an error with no `data`.
+    #[must_use]
+    pub fn new(code: i32, message: impl Into<String>) -> Self {
+        Self {
+            code,
+            message: message.into(),
+            data: None,
+        }
+    }
+}
+
+/// The subset of standard JSON-RPC 2.0 error codes this adapter raises.
+pub mod error_codes {
+    /// Invalid JSON was received by the server.
+    pub const PARSE_ERROR: i32 = -32_700;
+    /// The JSON sent is not a valid Request object.
+    pub const INVALID_REQUEST: i32 = -32_600;
+    /// The method does not exist / is not supported.
+    pub const METHOD_NOT_FOUND: i32 = -32_601;
+    /// Invalid method parameters.
+    pub const INVALID_PARAMS: i32 = -32_602;
+    /// Internal server error (a dispatched IdeA command failed).
+    pub const INTERNAL_ERROR: i32 = -32_603;
+}
+
+/// Errors a [`Transport`] may surface.
+#[derive(Debug, thiserror::Error)]
+pub enum TransportError {
+    /// The peer closed the stream (clean EOF). The serve loop stops on this.
+    #[error("transport closed")]
+    Closed,
+    /// An underlying I/O failure.
+    #[error("transport io error: {0}")]
+    Io(String),
+}
+
+/// Abstract message transport for the MCP server (the **S-MCP transport seam**).
+///
+/// One framed JSON-RPC message per `recv`/`send`. Concrete transports (stdio, and
+/// later a socket) live behind this trait so the server's protocol logic is
+/// **transport-agnostic** and, crucially, **unit-testable over an in-memory
+/// transport** with no socket and no child process. The `stdio` transport is the
+/// default concrete one; `socket` is a documented TODO (cadrage S-MCP).
+#[async_trait::async_trait]
+pub trait Transport: Send {
+    /// Receives the next framed message as raw bytes. Returns
+    /// [`TransportError::Closed`] on clean EOF (the serve loop exits).
+    async fn recv(&mut self) -> Result<Vec<u8>, TransportError>;
+
+    /// Sends one framed message (raw bytes, a complete JSON value).
+    async fn send(&mut self, message: &[u8]) -> Result<(), TransportError>;
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn request_without_id_is_a_notification() {
+        let raw = r#"{"jsonrpc":"2.0","method":"notifications/initialized"}"#;
+        let req: JsonRpcRequest = serde_json::from_str(raw).unwrap();
+        assert!(req.id.is_none());
+        assert_eq!(req.method, "notifications/initialized");
+    }
+
+    #[test]
+    fn success_and_error_are_mutually_exclusive_on_the_wire() {
+        let ok = JsonRpcResponse::success(serde_json::json!(1), serde_json::json!({"k":"v"}));
+        let text = serde_json::to_string(&ok).unwrap();
+        assert!(text.contains("\"result\""));
+        assert!(!text.contains("\"error\""));
+
+        let err = JsonRpcResponse::error(
+            serde_json::json!(1),
+            JsonRpcError::new(error_codes::METHOD_NOT_FOUND, "nope"),
+        );
+        let text = serde_json::to_string(&err).unwrap();
+        assert!(text.contains("\"error\""));
+        assert!(!text.contains("\"result\""));
+    }
+}

crates/infrastructure/src/orchestrator/mcp/mod.rs

@@ -0,0 +1,41 @@
+//! IdeA **MCP server** — a driving adapter that exposes the orchestration of IdeA
+//! as Model-Context-Protocol tools (cadrage `orchestration-v3`, lot **M2**).
+//!
+//! This adapter is the **pair of the filesystem watcher**
+//! ([`super::FsOrchestratorWatcher`]): a second, substitutable entry door onto the
+//! exact same [`application::OrchestratorService::dispatch`]. An MCP-capable CLI
+//! (Claude, Codex, Gemini…) calls a typed `idea_*` tool instead of dropping a JSON
+//! file under `.ideai/requests`; the resulting [`domain::OrchestratorCommand`] and
+//! its [`application::OrchestratorOutcome`] are identical. The server **invents no
+//! semantics** and **never re-routes** a reply (Décision 2 + principe directeur).
+//!
+//! ## Spike S-MCP — what is confined here
+//!
+//! Everything MCP/JSON-RPC/transport lives in this module and **nowhere else**;
+//! the domain and application layers never see it:
+//! - [`jsonrpc`] — a hand-rolled JSON-RPC 2.0 message model + a [`jsonrpc::Transport`]
+//!   seam (no MCP crate; chosen for zero-network testability — see its docs),
+//! - [`transport`] — the `stdio` default transport + an in-memory scriptable
+//!   transport for tests (a `socket` transport is a documented TODO),
+//! - [`tools`] — the tool catalogue and the tool → `OrchestratorCommand` mapping
+//!   (reusing [`domain::OrchestratorRequest::validate`] as the one validator),
+//! - [`server`] — [`McpServer`], the request loop over the transport.
+//!
+//! ## Testability
+//!
+//! [`McpServer::handle_raw`] turns one raw JSON-RPC message into its response with
+//! no I/O, and [`transport::MemoryTransport`] scripts a whole session in memory, so
+//! the agent of Test can cover the adapter **entirely off-network** (no socket, no
+//! child process).
+
+pub mod jsonrpc;
+pub mod server;
+pub mod tools;
+pub mod transport;
+
+pub use jsonrpc::{
+    JsonRpcError, JsonRpcRequest, JsonRpcResponse, Transport, TransportError, JSONRPC_VERSION,
+};
+pub use server::McpServer;
+pub use tools::{catalogue, map_tool_call, tool_returns_reply, ToolDef, ToolMapError};
+pub use transport::{MemoryTransport, StdioTransport};

crates/infrastructure/src/orchestrator/mcp/server.rs

@@ -0,0 +1,246 @@
+//! [`McpServer`] — the IdeA MCP **driving adapter** (cadrage Décision 4).
+//!
+//! This is the **pair of [`FsOrchestratorWatcher`](super::super::FsOrchestratorWatcher)**:
+//! another entry door onto the *same* [`OrchestratorService::dispatch`]. Where the
+//! watcher reads a JSON file, this server reads a JSON-RPC `tools/call`; both build
+//! an [`OrchestratorCommand`] and return its [`OrchestratorOutcome`] unchanged. It
+//! invents **no semantics** and **never re-routes** the reply — for `idea_ask_agent`
+//! it returns `outcome.reply` inline, exactly what `dispatch` produced.
+//!
+//! It implements the strict MCP minimum over JSON-RPC 2.0:
+//! - `initialize` → advertise protocol version + tool capability,
+//! - `tools/list` → the [`tools::catalogue`],
+//! - `tools/call` → map → `dispatch` → MCP tool result,
+//! - notifications (e.g. `notifications/initialized`) → accepted, no reply.
+//!
+//! It receives `Arc<OrchestratorService>` and the target [`Project`] by injection
+//! at the composition root (M3), exactly like the watcher — no application logic is
+//! duplicated here.
+
+use std::sync::Arc;
+
+use application::OrchestratorService;
+use domain::{DomainEvent, OrchestrationSource, Project};
+use serde_json::{json, Value};
+
+use super::jsonrpc::{
+    error_codes, JsonRpcError, JsonRpcRequest, JsonRpcResponse, Transport, TransportError,
+    JSONRPC_VERSION,
+};
+use super::tools::{self, ToolMapError};
+
+/// The MCP protocol version this server speaks (advertised on `initialize`).
+const MCP_PROTOCOL_VERSION: &str = "2024-11-05";
+
+/// The IdeA MCP server: an entry adapter over [`OrchestratorService::dispatch`].
+///
+/// Cheap to clone the dependencies it holds; one instance serves one project's
+/// agents over one transport (M3 owns one server per open project, beside the
+/// watcher).
+pub struct McpServer {
+    service: Arc<OrchestratorService>,
+    project: Project,
+    /// Optional event sink (a domain [`EventBus`](domain::ports::EventBus) publish
+    /// closure), the twin of the file watcher's. When set, a processed `tools/call`
+    /// publishes [`DomainEvent::OrchestratorRequestProcessed`] tagged
+    /// [`OrchestrationSource::Mcp`] so the presentation layer can surface that this
+    /// delegation arrived through the MCP door. `None` ⇒ no publication (the server
+    /// still works), so existing call sites stay valid.
+    events: Option<Arc<dyn Fn(DomainEvent) + Send + Sync>>,
+}
+
+impl McpServer {
+    /// Builds the server from the injected application service and target project.
+    /// No event sink — use [`with_events`](Self::with_events) to surface processed
+    /// requests on the bus.
+    #[must_use]
+    pub fn new(service: Arc<OrchestratorService>, project: Project) -> Self {
+        Self {
+            service,
+            project,
+            events: None,
+        }
+    }
+
+    /// Attaches an event sink so each processed `tools/call` is republished on the
+    /// bus tagged [`OrchestrationSource::Mcp`] — the MCP twin of the file watcher's
+    /// `events` closure. Additive: callers that do not need observability keep using
+    /// [`new`](Self::new).
+    #[must_use]
+    pub fn with_events(mut self, events: Arc<dyn Fn(DomainEvent) + Send + Sync>) -> Self {
+        self.events = Some(events);
+        self
+    }
+
+    /// Serves JSON-RPC messages from `transport` until it closes (clean EOF).
+    ///
+    /// Every inbound line is handled in isolation: a malformed line or an unknown
+    /// method yields a JSON-RPC error response, **never a panic** and never a
+    /// dropped connection. Notifications (no `id`) are processed without a reply.
+    pub async fn serve<T: Transport>(&self, transport: &mut T) {
+        loop {
+            let raw = match transport.recv().await {
+                Ok(bytes) => bytes,
+                Err(TransportError::Closed) => break,
+                Err(TransportError::Io(_)) => break,
+            };
+            if let Some(response) = self.handle_raw(&raw).await {
+                let Ok(bytes) = serde_json::to_vec(&response) else {
+                    continue;
+                };
+                if transport.send(&bytes).await.is_err() {
+                    break;
+                }
+            }
+        }
+    }
+
+    /// Parses one raw message and produces the response to send back, or `None`
+    /// for a notification (no `id`) that owes no reply.
+    ///
+    /// Kept standalone (no transport) so the whole request→response behaviour is
+    /// unit-testable over plain bytes, with no I/O.
+    pub async fn handle_raw(&self, raw: &[u8]) -> Option<JsonRpcResponse> {
+        let request: JsonRpcRequest = match serde_json::from_slice(raw) {
+            Ok(r) => r,
+            Err(e) => {
+                // Could not correlate (no id parsed) ⇒ null id per JSON-RPC.
+                return Some(JsonRpcResponse::error(
+                    Value::Null,
+                    JsonRpcError::new(error_codes::PARSE_ERROR, format!("invalid json: {e}")),
+                ));
+            }
+        };
+
+        if request.jsonrpc != JSONRPC_VERSION {
+            let id = request.id.unwrap_or(Value::Null);
+            return Some(JsonRpcResponse::error(
+                id,
+                JsonRpcError::new(
+                    error_codes::INVALID_REQUEST,
+                    format!("unsupported jsonrpc version: {}", request.jsonrpc),
+                ),
+            ));
+        }
+
+        // Notification (no id): never reply (the `?` short-circuits to `None`).
+        let id = request.id.clone()?;
+
+        let result = self.dispatch_method(&request.method, request.params).await;
+        Some(match result {
+            Ok(value) => JsonRpcResponse::success(id, value),
+            Err(error) => JsonRpcResponse::error(id, error),
+        })
+    }
+
+    /// Routes a method name to its handler.
+    async fn dispatch_method(
+        &self,
+        method: &str,
+        params: Option<Value>,
+    ) -> Result<Value, JsonRpcError> {
+        match method {
+            "initialize" => Ok(self.initialize_result()),
+            "tools/list" => Ok(self.tools_list_result()),
+            "tools/call" => self.tools_call(params.unwrap_or(Value::Null)).await,
+            other => Err(JsonRpcError::new(
+                error_codes::METHOD_NOT_FOUND,
+                format!("method not found: {other}"),
+            )),
+        }
+    }
+
+    /// The `initialize` result: protocol version, server identity, and the fact
+    /// that we expose tools.
+    fn initialize_result(&self) -> Value {
+        json!({
+            "protocolVersion": MCP_PROTOCOL_VERSION,
+            "capabilities": { "tools": {} },
+            "serverInfo": { "name": "idea-orchestrator", "version": env!("CARGO_PKG_VERSION") }
+        })
+    }
+
+    /// The `tools/list` result: the catalogue as MCP tool descriptors.
+    fn tools_list_result(&self) -> Value {
+        let tools: Vec<Value> = tools::catalogue()
+            .into_iter()
+            .map(|t| {
+                json!({
+                    "name": t.name,
+                    "description": t.description,
+                    "inputSchema": t.input_schema,
+                })
+            })
+            .collect();
+        json!({ "tools": tools })
+    }
+
+    /// `tools/call`: map the tool to an [`OrchestratorCommand`], `dispatch` it, and
+    /// fold the [`OrchestratorOutcome`](application::OrchestratorOutcome) into an MCP
+    /// tool result. The `idea_ask_agent` reply is returned **inline** — never
+    /// re-routed.
+    async fn tools_call(&self, params: Value) -> Result<Value, JsonRpcError> {
+        let name = params
+            .get("name")
+            .and_then(Value::as_str)
+            .ok_or_else(|| {
+                JsonRpcError::new(error_codes::INVALID_PARAMS, "missing tool `name`")
+            })?
+            .to_owned();
+        let arguments = params.get("arguments").cloned().unwrap_or(json!({}));
+
+        let command = tools::map_tool_call(&name, &arguments).map_err(map_err_to_jsonrpc)?;
+
+        let result = self.service.dispatch(&self.project, command).await;
+        // Surface the processed delegation on the bus, tagged as the MCP door — the
+        // twin of the file watcher's publish. `ok` mirrors the dispatch outcome; the
+        // action is the tool name. No-op when no sink is attached.
+        self.publish_processed(&name, result.is_ok());
+        match result {
+            Ok(outcome) => {
+                // The text the agent sees: the reply for an `ask`, else the detail.
+                let text = outcome.reply.clone().unwrap_or_else(|| outcome.detail.clone());
+                Ok(tool_result_text(&text, false))
+            }
+            // A failed IdeA command is reported as a tool execution error
+            // (`isError: true`) rather than a protocol error: the agent can read
+            // it and decide, and the connection stays healthy.
+            Err(e) => Ok(tool_result_text(&e.to_string(), true)),
+        }
+    }
+
+    /// Publishes [`DomainEvent::OrchestratorRequestProcessed`] for a handled
+    /// `tools/call`, tagged [`OrchestrationSource::Mcp`]. The requester id is the
+    /// stable `"mcp"` label until the per-session transport bind (open point S-MCP)
+    /// carries the connected agent's identity. No-op without an event sink.
+    fn publish_processed(&self, action: &str, ok: bool) {
+        if let Some(publish) = &self.events {
+            publish(DomainEvent::OrchestratorRequestProcessed {
+                requester_id: "mcp".to_owned(),
+                action: action.to_owned(),
+                ok,
+                source: OrchestrationSource::Mcp,
+            });
+        }
+    }
+}
+
+/// Builds an MCP `tools/call` result with a single text content block.
+fn tool_result_text(text: &str, is_error: bool) -> Value {
+    json!({
+        "content": [ { "type": "text", "text": text } ],
+        "isError": is_error
+    })
+}
+
+/// Maps a [`ToolMapError`] to the right JSON-RPC error code.
+fn map_err_to_jsonrpc(err: ToolMapError) -> JsonRpcError {
+    match err {
+        ToolMapError::UnknownTool(_) => {
+            JsonRpcError::new(error_codes::METHOD_NOT_FOUND, err.to_string())
+        }
+        ToolMapError::BadArguments(_) | ToolMapError::Invalid(_) => {
+            JsonRpcError::new(error_codes::INVALID_PARAMS, err.to_string())
+        }
+    }
+}

crates/infrastructure/src/orchestrator/mcp/tools.rs

@@ -0,0 +1,334 @@
+//! The IdeA MCP **tool catalogue** and the tool → [`OrchestratorRequest`] mapping.
+//!
+//! Each tool here is a thin, typed face over an **already-existing**
+//! [`OrchestratorCommand`] (cadrage Décision 4, principe directeur : « le serveur
+//! MCP n'invente aucune sémantique »). The mapping builds an
+//! [`OrchestratorRequest`] from the tool arguments and lets
+//! [`OrchestratorRequest::validate`] be the **single source of truth** for
+//! validation — the very same path the filesystem watcher takes. No tool
+//! validates anything itself; no tool reaches a use case directly.
+//!
+//! ## Agent discovery (`idea_list_agents`)
+//!
+//! Discovery maps to the [`OrchestratorCommand::ListAgents`] domain variant and is
+//! served through the **same** `OrchestratorService::dispatch` as every other tool
+//! (no use-case is reached directly). Its success carries the agent list inline as
+//! a JSON array in the outcome's reply — see [`tool_returns_reply`].
+
+use domain::{OrchestratorCommand, OrchestratorError, OrchestratorRequest};
+use serde_json::{json, Value};
+
+/// A tool the IdeA MCP server advertises, with the JSON Schema MCP clients read
+/// from `tools/list`.
+pub struct ToolDef {
+    /// Tool name as seen by the agent (`idea_ask_agent`, …).
+    pub name: &'static str,
+    /// One-line human description shown in the client.
+    pub description: &'static str,
+    /// JSON Schema (object) for the tool's arguments.
+    pub input_schema: Value,
+}
+
+/// Errors mapping a tool call into a validated [`OrchestratorCommand`].
+#[derive(Debug, thiserror::Error, PartialEq, Eq)]
+pub enum ToolMapError {
+    /// The tool name is not one this server exposes.
+    #[error("unknown tool: {0}")]
+    UnknownTool(String),
+    /// The `arguments` object was absent or not a JSON object.
+    #[error("tool `{0}` arguments must be a JSON object")]
+    BadArguments(String),
+    /// The arguments did not satisfy the underlying command's invariants.
+    #[error(transparent)]
+    Invalid(#[from] OrchestratorError),
+}
+
+/// Whether a successful call of `tool` carries an inline reply payload back to the
+/// caller: `idea_ask_agent` (the target's `Final`) and `idea_list_agents` (the
+/// agent list as a JSON array).
+#[must_use]
+pub fn tool_returns_reply(tool: &str) -> bool {
+    matches!(tool, "idea_ask_agent" | "idea_list_agents")
+}
+
+/// The full catalogue advertised on `tools/list`.
+///
+/// Exactly the tools whose mapping target already exists as an
+/// [`OrchestratorCommand`].
+#[must_use]
+pub fn catalogue() -> Vec<ToolDef> {
+    vec![
+        ToolDef {
+            name: "idea_list_agents",
+            description: "List the IdeA agents declared in the project's manifest. Returns the \
+                          agents inline as a JSON array (id, name, profile, origin, …).",
+            input_schema: json!({
+                "type": "object",
+                "properties": {},
+                "additionalProperties": false
+            }),
+        },
+        ToolDef {
+            name: "idea_ask_agent",
+            description: "Ask another IdeA agent a task and wait for its reply (synchronous \
+                          inter-agent rendezvous). Returns the target agent's answer inline.",
+            input_schema: json!({
+                "type": "object",
+                "properties": {
+                    "target": { "type": "string", "description": "Target agent display name." },
+                    "task": { "type": "string", "description": "The task/message to send." }
+                },
+                "required": ["target", "task"],
+                "additionalProperties": false
+            }),
+        },
+        ToolDef {
+            name: "idea_launch_agent",
+            description: "Launch (or attach) an IdeA agent. Fire-and-forget: returns once IdeA \
+                          has started the agent, without waiting for any work.",
+            input_schema: json!({
+                "type": "object",
+                "properties": {
+                    "target": { "type": "string", "description": "Target agent display name." },
+                    "profile": { "type": "string", "description": "Optional profile reference for a not-yet-created agent." },
+                    "task": { "type": "string", "description": "Optional initial task/context for the launched agent." },
+                    "visibility": { "type": "string", "enum": ["background", "visible"], "description": "Where to place the session (default background)." },
+                    "nodeId": { "type": "string", "description": "Target layout cell id, required when visibility is \"visible\"." }
+                },
+                "required": ["target"],
+                "additionalProperties": false
+            }),
+        },
+        ToolDef {
+            name: "idea_stop_agent",
+            description: "Stop a running IdeA agent by killing its terminal session.",
+            input_schema: json!({
+                "type": "object",
+                "properties": {
+                    "target": { "type": "string", "description": "Target agent display name." }
+                },
+                "required": ["target"],
+                "additionalProperties": false
+            }),
+        },
+        ToolDef {
+            name: "idea_update_context",
+            description: "Overwrite an IdeA agent's `.md` context with a new Markdown body.",
+            input_schema: json!({
+                "type": "object",
+                "properties": {
+                    "target": { "type": "string", "description": "Target agent display name." },
+                    "context": { "type": "string", "description": "New Markdown body." }
+                },
+                "required": ["target", "context"],
+                "additionalProperties": false
+            }),
+        },
+        ToolDef {
+            name: "idea_create_skill",
+            description: "Create a reusable IdeA skill (Markdown) in the project or global scope.",
+            input_schema: json!({
+                "type": "object",
+                "properties": {
+                    "name": { "type": "string", "description": "Skill name (also the .md stem)." },
+                    "context": { "type": "string", "description": "Markdown body of the skill." },
+                    "scope": { "type": "string", "enum": ["project", "global"], "description": "Scope (default project)." }
+                },
+                "required": ["name", "context"],
+                "additionalProperties": false
+            }),
+        },
+    ]
+}
+
+/// Maps an MCP `tools/call` (`name` + `arguments`) to a validated
+/// [`OrchestratorCommand`], reusing [`OrchestratorRequest::validate`] as the one
+/// validation authority.
+///
+/// `arguments` is the raw object an MCP client passes under `params.arguments`.
+///
+/// # Errors
+/// - [`ToolMapError::UnknownTool`] for a name outside [`catalogue`],
+/// - [`ToolMapError::BadArguments`] when `arguments` is not an object,
+/// - [`ToolMapError::Invalid`] when the underlying command's invariants fail.
+pub fn map_tool_call(
+    name: &str,
+    arguments: &Value,
+) -> Result<OrchestratorCommand, ToolMapError> {
+    let args = arguments
+        .as_object()
+        .ok_or_else(|| ToolMapError::BadArguments(name.to_owned()))?;
+
+    // Small helpers to pull optional string fields out of the arguments object.
+    let s = |key: &str| -> Option<String> {
+        args.get(key).and_then(Value::as_str).map(str::to_owned)
+    };
+
+    // Translate the tool into the wire-level request shape the watcher already
+    // uses (v2 `type`), then let `validate` enforce the invariants once.
+    let request = match name {
+        "idea_list_agents" => OrchestratorRequest {
+            request_type: Some("agent.list".to_owned()),
+            ..base()
+        },
+        "idea_ask_agent" => OrchestratorRequest {
+            request_type: Some("agent.message".to_owned()),
+            target_agent: s("target"),
+            task: s("task"),
+            ..base()
+        },
+        "idea_launch_agent" => OrchestratorRequest {
+            request_type: Some("agent.run".to_owned()),
+            target_agent: s("target"),
+            profile: s("profile"),
+            task: s("task"),
+            visibility: s("visibility"),
+            node_id: parse_node_id(args.get("nodeId")),
+            ..base()
+        },
+        "idea_stop_agent" => OrchestratorRequest {
+            request_type: Some("agent.stop".to_owned()),
+            target_agent: s("target"),
+            ..base()
+        },
+        "idea_update_context" => OrchestratorRequest {
+            request_type: Some("agent.update_context".to_owned()),
+            target_agent: s("target"),
+            context: s("context"),
+            ..base()
+        },
+        "idea_create_skill" => OrchestratorRequest {
+            request_type: Some("skill.create".to_owned()),
+            name: s("name"),
+            context: s("context"),
+            scope: s("scope"),
+            ..base()
+        },
+        other => return Err(ToolMapError::UnknownTool(other.to_owned())),
+    };
+
+    Ok(request.validate()?)
+}
+
+/// An all-`None` request to spread over; keeps each arm above to just its fields.
+fn base() -> OrchestratorRequest {
+    OrchestratorRequest {
+        action: None,
+        request_type: None,
+        requested_by: None,
+        name: None,
+        target_agent: None,
+        profile: None,
+        context: None,
+        task: None,
+        visibility: None,
+        node_id: None,
+        attach_to_cell: None,
+        scope: None,
+    }
+}
+
+/// Parses an optional `nodeId` JSON value into a [`domain::NodeId`], silently
+/// dropping a malformed/absent one (validation then rejects a `visible` launch
+/// missing its node, with a precise field error).
+fn parse_node_id(value: Option<&Value>) -> Option<domain::NodeId> {
+    let raw = value?.as_str()?;
+    uuid::Uuid::parse_str(raw).ok().map(domain::NodeId::from_uuid)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use domain::OrchestratorVisibility;
+
+    #[test]
+    fn ask_agent_maps_to_ask_command() {
+        let cmd = map_tool_call(
+            "idea_ask_agent",
+            &json!({ "target": "Architect", "task": "Analyse §17" }),
+        )
+        .unwrap();
+        assert_eq!(
+            cmd,
+            OrchestratorCommand::AskAgent {
+                target: "Architect".to_owned(),
+                task: "Analyse §17".to_owned(),
+            }
+        );
+    }
+
+    #[test]
+    fn launch_agent_maps_to_spawn_background_by_default() {
+        let cmd = map_tool_call("idea_launch_agent", &json!({ "target": "Dev" })).unwrap();
+        assert_eq!(
+            cmd,
+            OrchestratorCommand::SpawnAgent {
+                name: "Dev".to_owned(),
+                profile: None,
+                context: None,
+                visibility: OrchestratorVisibility::Background,
+            }
+        );
+    }
+
+    #[test]
+    fn stop_update_and_skill_map_to_their_commands() {
+        assert_eq!(
+            map_tool_call("idea_stop_agent", &json!({ "target": "Dev" })).unwrap(),
+            OrchestratorCommand::StopAgent { name: "Dev".to_owned() }
+        );
+        assert_eq!(
+            map_tool_call(
+                "idea_update_context",
+                &json!({ "target": "Dev", "context": "# body" })
+            )
+            .unwrap(),
+            OrchestratorCommand::UpdateAgentContext {
+                name: "Dev".to_owned(),
+                context: "# body".to_owned(),
+            }
+        );
+        assert_eq!(
+            map_tool_call(
+                "idea_create_skill",
+                &json!({ "name": "deploy", "context": "# steps" })
+            )
+            .unwrap(),
+            OrchestratorCommand::CreateSkill {
+                name: "deploy".to_owned(),
+                content: "# steps".to_owned(),
+                scope: domain::SkillScope::Project,
+            }
+        );
+    }
+
+    #[test]
+    fn missing_required_field_surfaces_validation_error() {
+        let err = map_tool_call("idea_ask_agent", &json!({ "target": "Architect" }));
+        assert!(matches!(err, Err(ToolMapError::Invalid(_))));
+    }
+
+    #[test]
+    fn list_agents_maps_to_list_command() {
+        let cmd = map_tool_call("idea_list_agents", &json!({})).unwrap();
+        assert_eq!(cmd, OrchestratorCommand::ListAgents);
+    }
+
+    #[test]
+    fn unknown_tool_is_rejected() {
+        let err = map_tool_call("idea_made_up_tool", &json!({}));
+        assert_eq!(
+            err,
+            Err(ToolMapError::UnknownTool("idea_made_up_tool".to_owned()))
+        );
+    }
+
+    #[test]
+    fn non_object_arguments_are_rejected() {
+        let err = map_tool_call("idea_ask_agent", &json!("nope"));
+        assert_eq!(
+            err,
+            Err(ToolMapError::BadArguments("idea_ask_agent".to_owned()))
+        );
+    }
+}

crates/infrastructure/src/orchestrator/mcp/transport.rs

@@ -0,0 +1,162 @@
+//! Concrete transports for the MCP adapter.
+//!
+//! - [`StdioTransport`] — the **default** transport (cadrage S-MCP): newline-
+//!   delimited JSON ("JSON Lines") over a pair of async byte streams. A CLI that
+//!   IdeA spawns with the injected MCP config speaks to the server over its
+//!   stdin/stdout; IdeA bridges those streams here.
+//! - [`MemoryTransport`] — an in-memory, **scriptable** transport used by tests to
+//!   drive the server with **no socket and no child process** (S-MCP testability
+//!   requirement). Lives in production code (not `#[cfg(test)]`) so the test crate
+//!   and downstream adapters can construct it.
+//!
+//! A `socket` transport is a documented **TODO** (cadrage S-MCP, point ouvert):
+//! the [`Transport`] seam means it can be added without touching the server.
+
+use std::collections::VecDeque;
+
+use tokio::io::{AsyncBufReadExt, AsyncWrite, AsyncWriteExt, BufReader};
+use tokio::sync::mpsc;
+
+use super::jsonrpc::{Transport, TransportError};
+
+/// Newline-delimited JSON transport over a generic async reader/writer.
+///
+/// Generic over the streams so the default real wiring uses
+/// `tokio::io::Stdin`/`Stdout`, while tests (or a future socket) can plug any
+/// `AsyncRead`/`AsyncWrite`. Each `recv` reads one line; each `send` writes one
+/// JSON value followed by `\n` and flushes.
+pub struct StdioTransport<R, W> {
+    reader: BufReader<R>,
+    writer: W,
+}
+
+impl<R, W> StdioTransport<R, W>
+where
+    R: tokio::io::AsyncRead + Unpin + Send,
+    W: AsyncWrite + Unpin + Send,
+{
+    /// Wraps a reader/writer pair as a line-delimited JSON transport.
+    pub fn new(reader: R, writer: W) -> Self {
+        Self {
+            reader: BufReader::new(reader),
+            writer,
+        }
+    }
+}
+
+#[async_trait::async_trait]
+impl<R, W> Transport for StdioTransport<R, W>
+where
+    R: tokio::io::AsyncRead + Unpin + Send,
+    W: AsyncWrite + Unpin + Send,
+{
+    async fn recv(&mut self) -> Result<Vec<u8>, TransportError> {
+        let mut line = String::new();
+        let n = self
+            .reader
+            .read_line(&mut line)
+            .await
+            .map_err(|e| TransportError::Io(e.to_string()))?;
+        if n == 0 {
+            return Err(TransportError::Closed);
+        }
+        Ok(line.into_bytes())
+    }
+
+    async fn send(&mut self, message: &[u8]) -> Result<(), TransportError> {
+        self.writer
+            .write_all(message)
+            .await
+            .map_err(|e| TransportError::Io(e.to_string()))?;
+        self.writer
+            .write_all(b"\n")
+            .await
+            .map_err(|e| TransportError::Io(e.to_string()))?;
+        self.writer
+            .flush()
+            .await
+            .map_err(|e| TransportError::Io(e.to_string()))?;
+        Ok(())
+    }
+}
+
+/// An in-memory, scriptable transport for tests (zero I/O, zero network).
+///
+/// `recv` drains a pre-loaded queue of inbound messages and then reports
+/// [`TransportError::Closed`] (so the serve loop terminates deterministically);
+/// every `send` is captured on an `mpsc` channel the test can drain to assert the
+/// exact responses the server produced.
+pub struct MemoryTransport {
+    inbound: VecDeque<Vec<u8>>,
+    outbound: mpsc::UnboundedSender<Vec<u8>>,
+}
+
+impl MemoryTransport {
+    /// Builds a transport that will hand `messages` to the server in order, then
+    /// signal EOF. Returns the transport and a receiver of everything the server
+    /// `send`s back.
+    #[must_use]
+    pub fn scripted(
+        messages: Vec<Vec<u8>>,
+    ) -> (Self, mpsc::UnboundedReceiver<Vec<u8>>) {
+        let (tx, rx) = mpsc::unbounded_channel();
+        (
+            Self {
+                inbound: messages.into(),
+                outbound: tx,
+            },
+            rx,
+        )
+    }
+
+    /// Convenience: script from JSON strings.
+    #[must_use]
+    pub fn scripted_str(messages: &[&str]) -> (Self, mpsc::UnboundedReceiver<Vec<u8>>) {
+        Self::scripted(messages.iter().map(|m| m.as_bytes().to_vec()).collect())
+    }
+}
+
+#[async_trait::async_trait]
+impl Transport for MemoryTransport {
+    async fn recv(&mut self) -> Result<Vec<u8>, TransportError> {
+        self.inbound.pop_front().ok_or(TransportError::Closed)
+    }
+
+    async fn send(&mut self, message: &[u8]) -> Result<(), TransportError> {
+        self.outbound
+            .send(message.to_vec())
+            .map_err(|_| TransportError::Closed)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn memory_transport_replays_then_closes() {
+        let (mut t, _rx) = MemoryTransport::scripted_str(&["a", "b"]);
+        assert_eq!(t.recv().await.unwrap(), b"a");
+        assert_eq!(t.recv().await.unwrap(), b"b");
+        assert!(matches!(t.recv().await, Err(TransportError::Closed)));
+    }
+
+    #[tokio::test]
+    async fn memory_transport_captures_sends() {
+        let (mut t, mut rx) = MemoryTransport::scripted(vec![]);
+        t.send(b"hello").await.unwrap();
+        assert_eq!(rx.recv().await.unwrap(), b"hello");
+    }
+
+    #[tokio::test]
+    async fn stdio_transport_round_trips_lines() {
+        let input = b"{\"x\":1}\n{\"y\":2}\n".to_vec();
+        let mut out: Vec<u8> = Vec::new();
+        {
+            let mut t = StdioTransport::new(&input[..], &mut out);
+            assert_eq!(t.recv().await.unwrap(), b"{\"x\":1}\n");
+            t.send(b"{\"ok\":true}").await.unwrap();
+        }
+        assert_eq!(out, b"{\"ok\":true}\n");
+    }
+}

crates/infrastructure/src/orchestrator/mod.rs

@@ -23,11 +23,13 @@
 //! notify just lowers latency. The per-file logic ([`process_request_file`]) is a
 //! standalone async fn, unit-tested against a temp dir independently of the watch.
 
+pub mod mcp;
+
 use std::path::{Path, PathBuf};
 use std::sync::Arc;
 
 use application::OrchestratorService;
-use domain::{DomainEvent, OrchestratorRequest, Project};
+use domain::{DomainEvent, OrchestrationSource, OrchestratorRequest, Project};
 use serde::{Deserialize, Serialize};
 use tokio::sync::mpsc;
 
@@ -205,6 +207,8 @@ async fn scan_once(
                 requester_id: requester_id.clone(),
                 action: outcome.action.clone().unwrap_or_default(),
                 ok: outcome.ok,
+                // This adapter is the filesystem entry door — tag it as such.
+                source: OrchestrationSource::File,
             });
         }
     }