Merge spike/claude-code-integration: PTY-based Claude Code with multi-agent support

Spike proved: spawning claude -p in a PTY from Rust gets Max subscription
billing. Multi-agent concurrency confirmed with session resumption.
Includes AgentPool REST API, claude-code provider, and spike documentation.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

# Conflicts:
#	.ignore

server/Cargo.toml

@@ -22,6 +22,8 @@ rust-embed = { workspace = true }
 mime_guess = { workspace = true }
 homedir = { workspace = true }
 serde_yaml = "0.9"
+portable-pty = { workspace = true }
+strip-ansi-escapes = { workspace = true }
 
 
 [dev-dependencies]

server/src/agents.rs

@@ -0,0 +1,307 @@
+use portable_pty::{CommandBuilder, PtySize, native_pty_system};
+use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
+use std::io::{BufRead, BufReader};
+use std::sync::Mutex;
+
+/// Manages multiple concurrent Claude Code agent sessions.
+///
+/// Each agent is identified by a string name (e.g., "coder-1", "coder-2").
+/// Agents run `claude -p` in a PTY for Max subscription billing.
+/// Sessions can be resumed for multi-turn conversations.
+pub struct AgentPool {
+    agents: Mutex<HashMap<String, AgentState>>,
+}
+
+#[derive(Clone, Serialize)]
+pub struct AgentInfo {
+    pub name: String,
+    pub role: String,
+    pub cwd: String,
+    pub session_id: Option<String>,
+    pub status: AgentStatus,
+    pub message_count: usize,
+}
+
+#[derive(Clone, Serialize)]
+#[serde(rename_all = "snake_case")]
+pub enum AgentStatus {
+    Idle,
+    Running,
+}
+
+struct AgentState {
+    role: String,
+    cwd: String,
+    session_id: Option<String>,
+    message_count: usize,
+}
+
+#[derive(Deserialize)]
+pub struct CreateAgentRequest {
+    pub name: String,
+    pub role: String,
+    pub cwd: String,
+}
+
+#[derive(Deserialize)]
+pub struct SendMessageRequest {
+    pub message: String,
+}
+
+#[derive(Serialize)]
+pub struct AgentResponse {
+    pub agent: String,
+    pub text: String,
+    pub session_id: Option<String>,
+    pub model: Option<String>,
+    pub api_key_source: Option<String>,
+    pub rate_limit_type: Option<String>,
+    pub cost_usd: Option<f64>,
+    pub input_tokens: Option<u64>,
+    pub output_tokens: Option<u64>,
+    pub duration_ms: Option<u64>,
+}
+
+impl AgentPool {
+    pub fn new() -> Self {
+        Self {
+            agents: Mutex::new(HashMap::new()),
+        }
+    }
+
+    pub fn create_agent(&self, req: CreateAgentRequest) -> Result<AgentInfo, String> {
+        let mut agents = self.agents.lock().map_err(|e| e.to_string())?;
+
+        if agents.contains_key(&req.name) {
+            return Err(format!("Agent '{}' already exists", req.name));
+        }
+
+        let state = AgentState {
+            role: req.role.clone(),
+            cwd: req.cwd.clone(),
+            session_id: None,
+            message_count: 0,
+        };
+
+        let info = AgentInfo {
+            name: req.name.clone(),
+            role: req.role,
+            cwd: req.cwd,
+            session_id: None,
+            status: AgentStatus::Idle,
+            message_count: 0,
+        };
+
+        agents.insert(req.name, state);
+        Ok(info)
+    }
+
+    pub fn list_agents(&self) -> Result<Vec<AgentInfo>, String> {
+        let agents = self.agents.lock().map_err(|e| e.to_string())?;
+        Ok(agents
+            .iter()
+            .map(|(name, state)| AgentInfo {
+                name: name.clone(),
+                role: state.role.clone(),
+                cwd: state.cwd.clone(),
+                session_id: state.session_id.clone(),
+                status: AgentStatus::Idle,
+                message_count: state.message_count,
+            })
+            .collect())
+    }
+
+    /// Send a message to an agent and wait for the complete response.
+    /// This spawns a `claude -p` process in a PTY, optionally resuming
+    /// a previous session for multi-turn conversations.
+    pub async fn send_message(
+        &self,
+        agent_name: &str,
+        message: &str,
+    ) -> Result<AgentResponse, String> {
+        let (cwd, role, session_id) = {
+            let agents = self.agents.lock().map_err(|e| e.to_string())?;
+            let state = agents
+                .get(agent_name)
+                .ok_or_else(|| format!("Agent '{}' not found", agent_name))?;
+            (
+                state.cwd.clone(),
+                state.role.clone(),
+                state.session_id.clone(),
+            )
+        };
+
+        let agent = agent_name.to_string();
+        let msg = message.to_string();
+        let role_clone = role.clone();
+
+        let result = tokio::task::spawn_blocking(move || {
+            run_agent_pty(&agent, &msg, &cwd, &role_clone, session_id.as_deref())
+        })
+        .await
+        .map_err(|e| format!("Agent task panicked: {e}"))??;
+
+        // Update session_id for next message
+        if let Some(ref sid) = result.session_id {
+            let mut agents = self.agents.lock().map_err(|e| e.to_string())?;
+            if let Some(state) = agents.get_mut(agent_name) {
+                state.session_id = Some(sid.clone());
+                state.message_count += 1;
+            }
+        }
+
+        Ok(result)
+    }
+}
+
+fn run_agent_pty(
+    agent_name: &str,
+    message: &str,
+    cwd: &str,
+    role: &str,
+    resume_session: Option<&str>,
+) -> Result<AgentResponse, String> {
+    let pty_system = native_pty_system();
+
+    let pair = pty_system
+        .openpty(PtySize {
+            rows: 50,
+            cols: 200,
+            pixel_width: 0,
+            pixel_height: 0,
+        })
+        .map_err(|e| format!("Failed to open PTY: {e}"))?;
+
+    let mut cmd = CommandBuilder::new("claude");
+    cmd.arg("-p");
+    cmd.arg(message);
+    cmd.arg("--output-format");
+    cmd.arg("stream-json");
+    cmd.arg("--verbose");
+
+    // Append role as system prompt context
+    cmd.arg("--append-system-prompt");
+    cmd.arg(format!(
+        "You are agent '{}' with role: {}. Work autonomously on the task given.",
+        agent_name, role
+    ));
+
+    // Resume previous session if available
+    if let Some(session_id) = resume_session {
+        cmd.arg("--resume");
+        cmd.arg(session_id);
+    }
+
+    cmd.cwd(cwd);
+    cmd.env("NO_COLOR", "1");
+
+    eprintln!(
+        "[agent:{}] Spawning claude -p (session: {:?})",
+        agent_name,
+        resume_session.unwrap_or("new")
+    );
+
+    let mut child = pair
+        .slave
+        .spawn_command(cmd)
+        .map_err(|e| format!("Failed to spawn claude for agent {agent_name}: {e}"))?;
+
+    drop(pair.slave);
+
+    let reader = pair
+        .master
+        .try_clone_reader()
+        .map_err(|e| format!("Failed to clone PTY reader: {e}"))?;
+
+    drop(pair.master);
+
+    let buf_reader = BufReader::new(reader);
+    let mut response = AgentResponse {
+        agent: agent_name.to_string(),
+        text: String::new(),
+        session_id: None,
+        model: None,
+        api_key_source: None,
+        rate_limit_type: None,
+        cost_usd: None,
+        input_tokens: None,
+        output_tokens: None,
+        duration_ms: None,
+    };
+
+    for line in buf_reader.lines() {
+        let line = match line {
+            Ok(l) => l,
+            Err(_) => break,
+        };
+
+        let trimmed = line.trim();
+        if trimmed.is_empty() {
+            continue;
+        }
+
+        let json: serde_json::Value = match serde_json::from_str(trimmed) {
+            Ok(j) => j,
+            Err(_) => continue, // skip non-JSON (terminal escapes)
+        };
+
+        let event_type = json.get("type").and_then(|t| t.as_str()).unwrap_or("");
+
+        match event_type {
+            "system" => {
+                response.session_id = json
+                    .get("session_id")
+                    .and_then(|s| s.as_str())
+                    .map(|s| s.to_string());
+                response.model = json
+                    .get("model")
+                    .and_then(|s| s.as_str())
+                    .map(|s| s.to_string());
+                response.api_key_source = json
+                    .get("apiKeySource")
+                    .and_then(|s| s.as_str())
+                    .map(|s| s.to_string());
+            }
+            "rate_limit_event" => {
+                if let Some(info) = json.get("rate_limit_info") {
+                    response.rate_limit_type = info
+                        .get("rateLimitType")
+                        .and_then(|s| s.as_str())
+                        .map(|s| s.to_string());
+                }
+            }
+            "assistant" => {
+                if let Some(message) = json.get("message") {
+                    if let Some(content) = message.get("content").and_then(|c| c.as_array()) {
+                        for block in content {
+                            if let Some(text) = block.get("text").and_then(|t| t.as_str()) {
+                                response.text.push_str(text);
+                            }
+                        }
+                    }
+                }
+            }
+            "result" => {
+                response.cost_usd = json.get("total_cost_usd").and_then(|c| c.as_f64());
+                response.duration_ms = json.get("duration_ms").and_then(|d| d.as_u64());
+                if let Some(usage) = json.get("usage") {
+                    response.input_tokens =
+                        usage.get("input_tokens").and_then(|t| t.as_u64());
+                    response.output_tokens =
+                        usage.get("output_tokens").and_then(|t| t.as_u64());
+                }
+            }
+            _ => {}
+        }
+    }
+
+    let _ = child.kill();
+
+    eprintln!(
+        "[agent:{}] Done. Session: {:?}, tokens: {:?}/{:?}",
+        agent_name, response.session_id, response.input_tokens, response.output_tokens
+    );
+
+    Ok(response)
+}

server/src/http/agents.rs

@@ -0,0 +1,127 @@
+use crate::http::context::{AppContext, OpenApiResult, bad_request};
+use poem_openapi::{Object, OpenApi, Tags, payload::Json};
+use serde::Serialize;
+use std::sync::Arc;
+
+#[derive(Tags)]
+enum AgentsTags {
+    Agents,
+}
+
+#[derive(Object)]
+struct CreateAgentPayload {
+    name: String,
+    role: String,
+    cwd: String,
+}
+
+#[derive(Object)]
+struct SendMessagePayload {
+    message: String,
+}
+
+#[derive(Object, Serialize)]
+struct AgentInfoResponse {
+    name: String,
+    role: String,
+    cwd: String,
+    session_id: Option<String>,
+    status: String,
+    message_count: usize,
+}
+
+#[derive(Object, Serialize)]
+struct AgentMessageResponse {
+    agent: String,
+    text: String,
+    session_id: Option<String>,
+    model: Option<String>,
+    api_key_source: Option<String>,
+    rate_limit_type: Option<String>,
+    cost_usd: Option<f64>,
+    input_tokens: Option<u64>,
+    output_tokens: Option<u64>,
+    duration_ms: Option<u64>,
+}
+
+pub struct AgentsApi {
+    pub ctx: Arc<AppContext>,
+}
+
+#[OpenApi(tag = "AgentsTags::Agents")]
+impl AgentsApi {
+    /// Create a new agent with a name, role, and working directory.
+    #[oai(path = "/agents", method = "post")]
+    async fn create_agent(
+        &self,
+        payload: Json<CreateAgentPayload>,
+    ) -> OpenApiResult<Json<AgentInfoResponse>> {
+        let req = crate::agents::CreateAgentRequest {
+            name: payload.0.name,
+            role: payload.0.role,
+            cwd: payload.0.cwd,
+        };
+
+        let info = self.ctx.agents.create_agent(req).map_err(bad_request)?;
+
+        Ok(Json(AgentInfoResponse {
+            name: info.name,
+            role: info.role,
+            cwd: info.cwd,
+            session_id: info.session_id,
+            status: "idle".to_string(),
+            message_count: info.message_count,
+        }))
+    }
+
+    /// List all registered agents.
+    #[oai(path = "/agents", method = "get")]
+    async fn list_agents(&self) -> OpenApiResult<Json<Vec<AgentInfoResponse>>> {
+        let agents = self.ctx.agents.list_agents().map_err(bad_request)?;
+
+        Ok(Json(
+            agents
+                .into_iter()
+                .map(|info| AgentInfoResponse {
+                    name: info.name,
+                    role: info.role,
+                    cwd: info.cwd,
+                    session_id: info.session_id,
+                    status: match info.status {
+                        crate::agents::AgentStatus::Idle => "idle".to_string(),
+                        crate::agents::AgentStatus::Running => "running".to_string(),
+                    },
+                    message_count: info.message_count,
+                })
+                .collect(),
+        ))
+    }
+
+    /// Send a message to an agent and wait for its response.
+    #[oai(path = "/agents/:name/message", method = "post")]
+    async fn send_message(
+        &self,
+        name: poem_openapi::param::Path<String>,
+        payload: Json<SendMessagePayload>,
+    ) -> OpenApiResult<Json<AgentMessageResponse>> {
+        let result = self
+            .ctx
+            .agents
+            .send_message(&name.0, &payload.0.message)
+            .await
+            .map_err(bad_request)?;
+
+        Ok(Json(AgentMessageResponse {
+            agent: result.agent,
+            text: result.text,
+            session_id: result.session_id,
+            model: result.model,
+            api_key_source: result.api_key_source,
+            rate_limit_type: result.rate_limit_type,
+            cost_usd: result.cost_usd,
+            input_tokens: result.input_tokens,
+            output_tokens: result.output_tokens,
+            duration_ms: result.duration_ms,
+        }))
+    }
+}

server/src/http/context.rs

@@ -1,3 +1,4 @@
+use crate::agents::AgentPool;
 use crate::state::SessionState;
 use crate::store::JsonFileStore;
 use crate::workflow::WorkflowState;
@@ -9,6 +10,7 @@ pub struct AppContext {
     pub state: Arc<SessionState>,
     pub store: Arc<JsonFileStore>,
     pub workflow: Arc<std::sync::Mutex<WorkflowState>>,
+    pub agents: Arc<AgentPool>,
 }
 
 pub type OpenApiResult<T> = poem::Result<T>;

server/src/http/mod.rs

@@ -1,3 +1,4 @@
+pub mod agents;
 pub mod anthropic;
 pub mod assets;
 pub mod chat;
@@ -10,6 +11,7 @@ pub mod workflow;
 pub mod project;
 pub mod ws;
 
+use agents::AgentsApi;
 use anthropic::AnthropicApi;
 use chat::ChatApi;
 use context::AppContext;
@@ -45,6 +47,7 @@ type ApiTuple = (
     IoApi,
     ChatApi,
     WorkflowApi,
+    AgentsApi,
 );
 
 type ApiService = OpenApiService<ApiTuple, ()>;
@@ -58,6 +61,7 @@ pub fn build_openapi_service(ctx: Arc<AppContext>) -> (ApiService, ApiService) {
         IoApi { ctx: ctx.clone() },
         ChatApi { ctx: ctx.clone() },
         WorkflowApi { ctx: ctx.clone() },
+        AgentsApi { ctx: ctx.clone() },
     );
 
     let api_service =
@@ -69,7 +73,8 @@ pub fn build_openapi_service(ctx: Arc<AppContext>) -> (ApiService, ApiService) {
         AnthropicApi::new(ctx.clone()),
         IoApi { ctx: ctx.clone() },
         ChatApi { ctx: ctx.clone() },
-        WorkflowApi { ctx },
+        WorkflowApi { ctx: ctx.clone() },
+        AgentsApi { ctx },
     );
 
     let docs_service =

server/src/llm/chat.rs

@@ -189,12 +189,55 @@ where
         .clone()
         .unwrap_or_else(|| "http://localhost:11434".to_string());
 
-    let is_claude = config.model.starts_with("claude-");
+    eprintln!("[chat] provider={} model={}", config.provider, config.model);
+    let is_claude_code = config.provider == "claude-code";
+    let is_claude = !is_claude_code && config.model.starts_with("claude-");
 
-    if !is_claude && config.provider.as_str() != "ollama" {
+    if !is_claude_code && !is_claude && config.provider.as_str() != "ollama" {
         return Err(format!("Unsupported provider: {}", config.provider));
     }
 
+    // Claude Code provider: bypasses our tool loop entirely.
+    // Claude Code has its own agent loop, tools, and context management.
+    // We just pipe the user message in and stream raw output back.
+    if is_claude_code {
+        use crate::llm::providers::claude_code::ClaudeCodeProvider;
+
+        let user_message = messages
+            .iter()
+            .rev()
+            .find(|m| m.role == Role::User)
+            .map(|m| m.content.clone())
+            .ok_or_else(|| "No user message found".to_string())?;
+
+        let project_root = state
+            .get_project_root()
+            .unwrap_or_else(|_| std::path::PathBuf::from("."));
+
+        let provider = ClaudeCodeProvider::new();
+        let response = provider
+            .chat_stream(
+                &user_message,
+                &project_root.to_string_lossy(),
+                &mut cancel_rx,
+                |token| on_token(token),
+            )
+            .await
+            .map_err(|e| format!("Claude Code Error: {e}"))?;
+
+        let assistant_msg = Message {
+            role: Role::Assistant,
+            content: response.content.unwrap_or_default(),
+            tool_calls: None,
+            tool_call_id: None,
+        };
+
+        let mut result = messages.clone();
+        result.push(assistant_msg);
+        on_update(&result);
+        return Ok(result);
+    }
+
     let tool_defs = get_tool_definitions();
     let tools = if config.enable_tools.unwrap_or(true) {
         tool_defs.as_slice()

server/src/llm/providers/claude_code.rs

@@ -0,0 +1,299 @@
+use portable_pty::{CommandBuilder, PtySize, native_pty_system};
+use std::io::{BufRead, BufReader};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use tokio::sync::watch;
+
+use crate::llm::types::CompletionResponse;
+
+/// Manages a Claude Code session via a pseudo-terminal.
+///
+/// Spawns `claude -p` in a PTY so isatty() returns true (which may
+/// influence billing), while using `--output-format stream-json` to
+/// get clean, structured NDJSON output instead of TUI escape sequences.
+pub struct ClaudeCodeProvider;
+
+impl ClaudeCodeProvider {
+    pub fn new() -> Self {
+        Self
+    }
+
+    pub async fn chat_stream<F>(
+        &self,
+        user_message: &str,
+        project_root: &str,
+        cancel_rx: &mut watch::Receiver<bool>,
+        mut on_token: F,
+    ) -> Result<CompletionResponse, String>
+    where
+        F: FnMut(&str) + Send,
+    {
+        let message = user_message.to_string();
+        let cwd = project_root.to_string();
+        let cancelled = Arc::new(AtomicBool::new(false));
+        let cancelled_clone = cancelled.clone();
+
+        let mut cancel_watch = cancel_rx.clone();
+        tokio::spawn(async move {
+            while cancel_watch.changed().await.is_ok() {
+                if *cancel_watch.borrow() {
+                    cancelled_clone.store(true, Ordering::Relaxed);
+                    break;
+                }
+            }
+        });
+
+        let (token_tx, mut token_rx) = tokio::sync::mpsc::unbounded_channel::<String>();
+
+        let pty_handle = tokio::task::spawn_blocking(move || {
+            run_pty_session(&message, &cwd, cancelled, token_tx)
+        });
+
+        let mut full_output = String::new();
+        while let Some(token) = token_rx.recv().await {
+            full_output.push_str(&token);
+            on_token(&token);
+        }
+
+        pty_handle
+            .await
+            .map_err(|e| format!("PTY task panicked: {e}"))??;
+
+        Ok(CompletionResponse {
+            content: Some(full_output),
+            tool_calls: None,
+        })
+    }
+}
+
+/// Run `claude -p` with stream-json output inside a PTY.
+///
+/// The PTY makes isatty() return true. The `-p` flag gives us
+/// single-shot non-interactive mode with structured output.
+fn run_pty_session(
+    user_message: &str,
+    cwd: &str,
+    cancelled: Arc<AtomicBool>,
+    token_tx: tokio::sync::mpsc::UnboundedSender<String>,
+) -> Result<(), String> {
+    let pty_system = native_pty_system();
+
+    let pair = pty_system
+        .openpty(PtySize {
+            rows: 50,
+            cols: 200,
+            pixel_width: 0,
+            pixel_height: 0,
+        })
+        .map_err(|e| format!("Failed to open PTY: {e}"))?;
+
+    let mut cmd = CommandBuilder::new("claude");
+    cmd.arg("-p");
+    cmd.arg(user_message);
+    cmd.arg("--output-format");
+    cmd.arg("stream-json");
+    cmd.arg("--verbose");
+    cmd.cwd(cwd);
+    // Keep TERM reasonable but disable color
+    cmd.env("NO_COLOR", "1");
+
+    eprintln!("[pty-debug] Spawning: claude -p \"{}\" --output-format stream-json --verbose", user_message);
+
+    let mut child = pair
+        .slave
+        .spawn_command(cmd)
+        .map_err(|e| format!("Failed to spawn claude: {e}"))?;
+
+    eprintln!("[pty-debug] Process spawned, pid: {:?}", child.process_id());
+    drop(pair.slave);
+
+    let reader = pair
+        .master
+        .try_clone_reader()
+        .map_err(|e| format!("Failed to clone PTY reader: {e}"))?;
+
+    // We don't need to write anything — -p mode takes prompt as arg
+    drop(pair.master);
+
+    // Read NDJSON lines from stdout
+    let (line_tx, line_rx) = std::sync::mpsc::channel::<Option<String>>();
+
+    std::thread::spawn(move || {
+        let buf_reader = BufReader::new(reader);
+        eprintln!("[pty-debug] Reader thread started");
+        for line in buf_reader.lines() {
+            match line {
+                Ok(l) => {
+                    eprintln!("[pty-debug] raw line: {}", l);
+                    if line_tx.send(Some(l)).is_err() {
+                        break;
+                    }
+                }
+                Err(e) => {
+                    eprintln!("[pty-debug] read error: {e}");
+                    let _ = line_tx.send(None);
+                    break;
+                }
+            }
+        }
+        eprintln!("[pty-debug] Reader thread done");
+        let _ = line_tx.send(None);
+    });
+
+    let mut got_result = false;
+
+    loop {
+        if cancelled.load(Ordering::Relaxed) {
+            let _ = child.kill();
+            return Err("Cancelled".to_string());
+        }
+
+        match line_rx.recv_timeout(std::time::Duration::from_millis(500)) {
+            Ok(Some(line)) => {
+                let trimmed = line.trim();
+                if trimmed.is_empty() {
+                    continue;
+                }
+
+                eprintln!("[pty-debug] processing: {}...", &trimmed[..trimmed.len().min(120)]);
+
+                // Try to parse as JSON
+                if let Ok(json) = serde_json::from_str::<serde_json::Value>(trimmed) {
+                    if let Some(event_type) = json.get("type").and_then(|t| t.as_str()) {
+                        match event_type {
+                            // Streaming deltas (when --include-partial-messages is used)
+                            "stream_event" => {
+                                if let Some(event) = json.get("event") {
+                                    handle_stream_event(event, &token_tx);
+                                }
+                            }
+                            // Complete assistant message
+                            "assistant" => {
+                                if let Some(message) = json.get("message") {
+                                    if let Some(content) = message.get("content").and_then(|c| c.as_array()) {
+                                        for block in content {
+                                            if let Some(text) = block.get("text").and_then(|t| t.as_str()) {
+                                                let _ = token_tx.send(text.to_string());
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                            // Final result with usage stats
+                            "result" => {
+                                if let Some(cost) = json.get("total_cost_usd").and_then(|c| c.as_f64()) {
+                                    let _ = token_tx.send(format!("\n\n---\n_Cost: ${cost:.4}_\n"));
+                                }
+                                if let Some(usage) = json.get("usage") {
+                                    let input = usage.get("input_tokens").and_then(|t| t.as_u64()).unwrap_or(0);
+                                    let output = usage.get("output_tokens").and_then(|t| t.as_u64()).unwrap_or(0);
+                                    let cached = usage.get("cache_read_input_tokens").and_then(|t| t.as_u64()).unwrap_or(0);
+                                    let _ = token_tx.send(format!("_Tokens: {input} in / {output} out / {cached} cached_\n"));
+                                }
+                                got_result = true;
+                            }
+                            // System init — log billing info
+                            "system" => {
+                                let api_source = json.get("apiKeySource").and_then(|s| s.as_str()).unwrap_or("unknown");
+                                let model = json.get("model").and_then(|s| s.as_str()).unwrap_or("unknown");
+                                let _ = token_tx.send(format!("_[{model} | apiKey: {api_source}]_\n\n"));
+                            }
+                            // Rate limit info
+                            "rate_limit_event" => {
+                                if let Some(info) = json.get("rate_limit_info") {
+                                    let status = info.get("status").and_then(|s| s.as_str()).unwrap_or("unknown");
+                                    let limit_type = info.get("rateLimitType").and_then(|s| s.as_str()).unwrap_or("unknown");
+                                    let _ = token_tx.send(format!("_[rate limit: {status} ({limit_type})]_\n\n"));
+                                }
+                            }
+                            _ => {}
+                        }
+                    }
+                }
+                // Ignore non-JSON lines (terminal escape sequences)
+
+                if got_result {
+                    break;
+                }
+            }
+            Ok(None) => break, // EOF
+            Err(std::sync::mpsc::RecvTimeoutError::Timeout) => {
+                // Check if child has exited
+                if let Ok(Some(_status)) = child.try_wait() {
+                    // Drain remaining lines
+                    while let Ok(Some(line)) = line_rx.try_recv() {
+                        let trimmed = line.trim();
+                        if let Ok(json) = serde_json::from_str::<serde_json::Value>(trimmed) {
+                            if let Some(event) = json
+                                .get("type")
+                                .filter(|t| t.as_str() == Some("stream_event"))
+                                .and_then(|_| json.get("event"))
+                            {
+                                handle_stream_event(event, &token_tx);
+                            }
+                        }
+                    }
+                    break;
+                }
+            }
+            Err(std::sync::mpsc::RecvTimeoutError::Disconnected) => break,
+        }
+
+        // Don't set got_result here — just let the process finish naturally
+        let _ = got_result;
+    }
+
+    let _ = child.kill();
+    Ok(())
+}
+
+/// Extract text from a stream event and send to the token channel.
+fn handle_stream_event(
+    event: &serde_json::Value,
+    token_tx: &tokio::sync::mpsc::UnboundedSender<String>,
+) {
+    let event_type = event.get("type").and_then(|t| t.as_str()).unwrap_or("");
+
+    match event_type {
+        // Text content streaming
+        "content_block_delta" => {
+            if let Some(delta) = event.get("delta") {
+                let delta_type = delta.get("type").and_then(|t| t.as_str()).unwrap_or("");
+                match delta_type {
+                    "text_delta" => {
+                        if let Some(text) = delta.get("text").and_then(|t| t.as_str()) {
+                            let _ = token_tx.send(text.to_string());
+                        }
+                    }
+                    "thinking_delta" => {
+                        if let Some(thinking) = delta.get("thinking").and_then(|t| t.as_str()) {
+                            let _ = token_tx.send(format!("[thinking] {thinking}"));
+                        }
+                    }
+                    _ => {}
+                }
+            }
+        }
+        // Message complete — log usage info
+        "message_delta" => {
+            if let Some(usage) = event.get("usage") {
+                let output_tokens = usage
+                    .get("output_tokens")
+                    .and_then(|t| t.as_u64())
+                    .unwrap_or(0);
+                let _ = token_tx.send(format!("\n[tokens: {output_tokens} output]\n"));
+            }
+        }
+        // Log errors
+        "error" => {
+            if let Some(error) = event.get("error") {
+                let msg = error
+                    .get("message")
+                    .and_then(|m| m.as_str())
+                    .unwrap_or("unknown error");
+                let _ = token_tx.send(format!("\n[error: {msg}]\n"));
+            }
+        }
+        _ => {}
+    }
+}

server/src/llm/providers/mod.rs

@@ -1,2 +1,3 @@
 pub mod anthropic;
+pub mod claude_code;
 pub mod ollama;

server/src/main.rs

@@ -1,3 +1,4 @@
+mod agents;
 mod http;
 mod io;
 mod llm;
@@ -5,6 +6,7 @@ mod state;
 mod store;
 mod workflow;
 
+use crate::agents::AgentPool;
 use crate::http::build_routes;
 use crate::http::context::AppContext;
 use crate::state::SessionState;
@@ -22,11 +24,13 @@ async fn main() -> Result<(), std::io::Error> {
         JsonFileStore::from_path(PathBuf::from("store.json")).map_err(std::io::Error::other)?,
     );
     let workflow = Arc::new(std::sync::Mutex::new(WorkflowState::default()));
+    let agents = Arc::new(AgentPool::new());
 
     let ctx = AppContext {
         state: app_state,
         store,
         workflow,
+        agents,
     };
 
     let app = build_routes(ctx);