fix: normalize claude system events and refresh tests

Author: NathanFlurry

CLAUDE.md

@@ -34,6 +34,8 @@ Universal schema guidance:
 - Do not make breaking changes to API endpoints.
 - When changing API routes, ensure the HTTP/SSE test suite has full coverage of every route.
 - When agent schema changes, ensure API tests cover the new schema and event shapes end-to-end.
+- Never use synthetic data or mocked responses in tests.
+- Never manually write agent types; always use generated types in `resources/agent-schemas/`. If types are broken, fix the generated types.
 
 ### CLI ⇄ HTTP endpoint map (keep in sync)
 

README.md

@@ -3,12 +3,18 @@
 Universal API for running Claude Code, Codex, OpenCode, and Amp inside sandboxes.
 
 - **Any coding agent**: Universal API to interact with all agents with full feature coverage
-- **Server Mode**: Run as HTTP server from any sandbox provider or as TypeScript & Python SDK
+- **Server, stdin/stdout, or SDK mode**: Run as an HTTP server, CLI using stdin/stdout, or with the SDK
 - **Universal session schema**: Universal schema to store agent transcripts
 - **Supports your sandbox provider**: Daytona, E2B, Vercel Sandboxes, and more
 - **Lightweight, portable Rust binary**: Install anywhere with 1 curl command
 - **OpenAPI spec**: Versioned API schema tracked in `sdks/openapi/openapi.json`
 
+Coming soon:
+
+- **Vercel AI SDK Compatibility**: Works with existing AI SDK tooling, like `useChat`
+- **Auto-configure MCP & Skills**: Auto-load MCP servers & skills for your agents
+- **Process & logs manager**: Manage processes, logs, and ports for your agents to run background processes
+
 ## Agent Support
 
 | Feature | [Claude Code](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/overview) | [Codex](https://github.com/openai/codex) | [OpenCode](https://github.com/opencode-ai/opencode) | [Amp](https://ampcode.com) |
@@ -62,13 +68,22 @@ Features out of scope:
 ## FAQ
 
 **Why not use PTY?**
-PTY-based approaches require parsing terminal escape sequences and dealing with interactive prompts. The agents we support all have machine-readable output modes (JSONL, HTTP APIs) that provide structured events, making integration more reliable.
+
+PTY-based approaches require parsing terminal escape sequences and dealing with interactive prompts.
+
+The agents we support all have machine-readable output modes (JSONL, HTTP APIs) that provide structured events, making integration more reliable.
 
 **Why not use features that already exist on sandbox provider APIs?**
-Sandbox providers focus on infrastructure (containers, VMs, networking). This project focuses specifically on coding agent orchestration—session management, HITL (human-in-the-loop) flows, and universal event schemas. These concerns are complementary.
+
+Sandbox providers focus on infrastructure (containers, VMs, networking).
+
+This project focuses specifically on coding agent orchestration: session management, HITL (human-in-the-loop) flows, and universal event schemas. These concerns are complementary.
 
 **Does it support [platform]?**
 The server is a single Rust binary that runs anywhere with a curl install. If your platform can run Linux binaries (Docker, VMs, etc.), it works. See the deployment guides for E2B, Daytona, Vercel Sandboxes, and Docker.
 
 **Can I use this with my personal API keys?**
 Yes. Use `sandbox-agent credentials extract-env` to extract API keys from your local agent configs (Claude Code, Codex, OpenCode, Amp) and pass them to the sandbox environment.
+
+**Why rust?**
+TODO

ROADMAP.md

@@ -1,5 +1,7 @@
 ## soon
 
+- implement stdin/stdout
+- switch sdk to use sdtin/stdout for embedded mdoe
 - discuss actor arch in readme + give example
 - skillfile
     - specifically include the release checklist

pnpm-lock.yaml

@@ -112,11 +112,13 @@ impl AgentManager {
 
     pub fn install(&self, agent: AgentId, options: InstallOptions) -> Result<InstallResult, AgentError> {
         let install_path = self.binary_path(agent);
-        if install_path.exists() && !options.reinstall {
-            return Ok(InstallResult {
-                path: install_path,
-                version: self.version(agent).unwrap_or(None),
-            });
+        if !options.reinstall {
+            if let Ok(existing_path) = self.resolve_binary(agent) {
+                return Ok(InstallResult {
+                    path: existing_path,
+                    version: self.version(agent).unwrap_or(None),
+                });
+            }
         }
 
         fs::create_dir_all(&self.install_dir)?;
@@ -135,7 +137,9 @@ impl AgentManager {
     }
 
     pub fn is_installed(&self, agent: AgentId) -> bool {
-        self.binary_path(agent).exists() || find_in_path(agent.binary_name()).is_some()
+        self.binary_path(agent).exists()
+            || find_in_path(agent.binary_name()).is_some()
+            || default_install_dir().join(agent.binary_name()).exists()
     }
 
     pub fn binary_path(&self, agent: AgentId) -> PathBuf {
@@ -368,6 +372,10 @@ impl AgentManager {
         if let Some(path) = find_in_path(agent.binary_name()) {
             return Ok(path);
         }
+        let fallback = default_install_dir().join(agent.binary_name());
+        if fallback.exists() {
+            return Ok(fallback);
+        }
         Err(AgentError::BinaryNotFound { agent })
     }
 }
@@ -780,6 +788,12 @@ fn find_in_path(binary_name: &str) -> Option<PathBuf> {
     None
 }
 
+fn default_install_dir() -> PathBuf {
+    dirs::data_dir()
+        .map(|dir| dir.join("sandbox-agent").join("bin"))
+        .unwrap_or_else(|| PathBuf::from(".").join(".sandbox-agent").join("bin"))
+}
+
 fn download_bytes(url: &Url) -> Result<Vec<u8>, AgentError> {
     let client = Client::builder().build()?;
     let mut response = client.get(url.clone()).send()?;

server/packages/agent-management/src/agents.rs

@@ -19,7 +19,7 @@ use tower_http::cors::CorsLayer;
 const PROMPT: &str = "Reply with exactly the single word OK.";
 const PERMISSION_PROMPT: &str = "List files in the current directory using available tools.";
 const QUESTION_PROMPT: &str =
-    "Ask the user a multiple-choice question with options yes/no using any built-in AskUserQuestion tool, then wait.";
+    "Use the AskUserQuestion tool to ask exactly one yes/no question, then wait for a reply. Do not answer yourself.";
 
 struct TestApp {
     app: Router,
@@ -1022,7 +1022,7 @@ async fn approval_flow_snapshots() {
         }
 
         let question_reply_session = format!("question-reply-{}", config.agent.as_str());
-        create_session(&app.app, config.agent, &question_reply_session, test_permission_mode(config.agent)).await;
+        create_session(&app.app, config.agent, &question_reply_session, "plan").await;
         let status = send_status(
             &app.app,
             Method::POST,
@@ -1083,7 +1083,7 @@ async fn approval_flow_snapshots() {
         }
 
         let question_reject_session = format!("question-reject-{}", config.agent.as_str());
-        create_session(&app.app, config.agent, &question_reject_session, test_permission_mode(config.agent)).await;
+        create_session(&app.app, config.agent, &question_reject_session, "plan").await;
         let status = send_status(
             &app.app,
             Method::POST,

server/packages/sandbox-agent/tests/http_sse_snapshots.rs

@@ -1,6 +1,5 @@
 ---
 source: server/packages/sandbox-agent/tests/http_sse_snapshots.rs
-assertion_line: 978
 expression: normalize_events(&permission_events)
 ---
 - agent: claude
@@ -9,8 +8,10 @@ expression: normalize_events(&permission_events)
   started:
     message: session.created
 - agent: claude
-  kind: unknown
+  kind: started
   seq: 2
+  started:
+    message: system.init
 - agent: claude
   kind: message
   message:

server/packages/sandbox-agent/tests/snapshots/http_sse_snapshots__approval_flow_snapshots@permission_events_claude.snap

@@ -1,6 +1,5 @@
 ---
 source: server/packages/sandbox-agent/tests/http_sse_snapshots.rs
-assertion_line: 1100
 expression: normalize_events(&reject_events)
 ---
 - agent: claude
@@ -9,8 +8,10 @@ expression: normalize_events(&reject_events)
   started:
     message: session.created
 - agent: claude
-  kind: unknown
+  kind: started
   seq: 2
+  started:
+    message: system.init
 - agent: claude
   kind: message
   message:

server/packages/sandbox-agent/tests/snapshots/http_sse_snapshots__approval_flow_snapshots@question_reject_events_claude.snap

@@ -8,8 +8,10 @@ expression: normalize_events(&question_events)
   started:
     message: session.created
 - agent: claude
-  kind: unknown
+  kind: started
   seq: 2
+  started:
+    message: system.init
 - agent: claude
   kind: message
   message:

server/packages/sandbox-agent/tests/snapshots/http_sse_snapshots__approval_flow_snapshots@question_reply_events_claude.snap

@@ -1,6 +1,5 @@
 ---
 source: server/packages/sandbox-agent/tests/http_sse_snapshots.rs
-assertion_line: 1232
 expression: snapshot
 ---
 session_a:
@@ -10,8 +9,10 @@ session_a:
     started:
       message: session.created
   - agent: claude
-    kind: unknown
+    kind: started
     seq: 2
+    started:
+      message: system.init
   - agent: claude
     kind: message
     message:
@@ -27,8 +28,10 @@ session_b:
     started:
       message: session.created
   - agent: claude
-    kind: unknown
+    kind: started
     seq: 2
+    started:
+      message: system.init
   - agent: claude
     kind: message
     message: