Rewrite core runtime for SDK support

[gregpr07]

Summary

• Rewrites live agent execution around a runtime-owned core.
• Routes SDK, CLI, TUI, subagent, cancellation, mailbox, and completion flows through runtime state.
• Adds Python SDK-facing agent/browser APIs and JSON-RPC SDK server support.
• Keeps SDK live state memory-backed; SDK runs do not create state.db.
• Enables real SDK Agent.run() through the live runtime path instead of rejecting non-fake providers.
• Preserves local Chrome profile preflight behavior from the browser profile PR.

Verification

• cargo fmt --check
• cargo test
• uv run --with pytest python -m pytest -q
• cargo build -p browser-use-cli
• uv run --with pytest python -m pytest -q python/tests/test_browser_use_sdk.py::test_runtime_client_round_trips_against_sdk_server_binary
• No-key real-provider SDK smoke reached provider credential resolution instead of SDK rejection: JsonRpcError: no provider credentials found...

Notes

• The SDK server is memory-backed and does not write a SQLite state.db.
• The current agent turn driver still receives an ephemeral in-memory Store compatibility bridge until the remaining legacy RuntimeTurnDriver store dependency is fully removed.

[cubic-dev-ai]

12 issues found

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="python/browser_use/browser.py">

<violation number="1" location="python/browser_use/browser.py:100">
P1: `Browser.close()` clears `browser_id` even when the runtime close call fails, which can orphan the remote browser and prevent retrying cleanup.</violation>
</file>

<file name="crates/browser-use-tui/src/main.rs">

<violation number="1" location="crates/browser-use-tui/src/main.rs:3508">
P2: Agent startup errors are now propagated with `?`, which can terminate the command flow instead of recording a session failure and continuing UI operation.</violation>
</file>

<file name="crates/browser-use-agent/src/tools/handlers/subagent.rs">

<violation number="1" location="crates/browser-use-agent/src/tools/handlers/subagent.rs:1585">
P2: Runtime-backed legacy wait_agent regressed: multi-target requests now fail despite the tool contract allowing multiple target IDs.</violation>
</file>

<file name="python/browser_use/runtime.py">

<violation number="1" location="python/browser_use/runtime.py:58">
P2: Write/drain failures leak pending RPC futures because `_pending` is not cleaned up on I/O errors.</violation>

<violation number="2" location="python/browser_use/runtime.py:75">
P1: Runtime startup is race-prone and can spawn multiple SDK server subprocesses under concurrent first use.</violation>

<violation number="3" location="python/browser_use/runtime.py:119">
P1: Pending RPC calls can hang forever when server stdout closes without a response.</violation>
</file>

<file name="crates/browser-use-agent/src/live_executor.rs">

<violation number="1" location="crates/browser-use-agent/src/live_executor.rs:472">
P2: Cancellation errors can be misclassified as failures because only `session.failed` is checked before appending a new failure event.</violation>
</file>

<file name="crates/browser-use-agent/src/turn/sampling.rs">

<violation number="1" location="crates/browser-use-agent/src/turn/sampling.rs:604">
P1: Retryable streamed provider errors are now treated as terminal failures, bypassing the existing retry policy.</violation>
</file>

<file name="crates/browser-use-tui/src/runtime.rs">

<violation number="1" location="crates/browser-use-tui/src/runtime.rs:215">
P2: Non-atomic runtime lazy initialization can race and start duplicate runtime servers for the same state dir.</violation>
</file>

<file name="crates/browser-use-tui/src/transcript.rs">

<violation number="1" location="crates/browser-use-tui/src/transcript.rs:635">
P2: Active-work checks are evaluated eagerly, causing unnecessary runtime count lookups on every model rebuild.</violation>

<violation number="2" location="crates/browser-use-tui/src/transcript.rs:1663">
P2: UI transcript rendering now performs runtime initialization work, which can spawn a local runtime server and degrade responsiveness.</violation>
</file>

<sub>Tip: instead of fixing issues one by one fix them all with cubic

Re-trigger cubic</sub>

[cubic-dev-ai]

P1: Browser.close() clears browser_id even when the runtime close call fails, which can orphan the remote browser and prevent retrying cleanup.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At python/browser_use/browser.py, line 100:

<comment>`Browser.close()` clears `browser_id` even when the runtime close call fails, which can orphan the remote browser and prevent retrying cleanup.</comment>

<file context>
@@ -0,0 +1,156 @@
+            pass
+        self.browser_id = None
+
+    async def close(self) -> None:
+        if self.browser_id is None:
+            return
</file context>

Suggested change

	async def close(self) -> None:
	if self.browser_id is None:
	return
	try:
	await self._runtime.call("browser.close", {"browser_id": self.browser_id})
	except Exception:
	pass
	self.browser_id = None
	async def close(self) -> None:
	if self.browser_id is None:
	return
	try:
	await self._runtime.call("browser.close", {"browser_id": self.browser_id})
	except Exception:
	return
	self.browser_id = None

[cubic-dev-ai]

P1: Retryable streamed provider errors are now treated as terminal failures, bypassing the existing retry policy.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At crates/browser-use-agent/src/turn/sampling.rs, line 604:

<comment>Retryable streamed provider errors are now treated as terminal failures, bypassing the existing retry policy.</comment>

<file context>
@@ -590,11 +599,12 @@ impl<T: SamplingTransport, R: CallRunner + 'static> ModelSamplingDriver<T, R> {
+                Ok(StreamProgress::Done)
             }
-            // Reasoning, lifecycle markers, provider-side notices, step finishes:
+            LlmEvent::ProviderError { message, .. } => Err(provider_error_event_to_agent(message)),
+            // Reasoning, lifecycle markers, and step finishes:
             // no accumulation; their UI mapping (if any) already happened above.
</file context>

[cubic-dev-ai]

P1: Runtime startup is race-prone and can spawn multiple SDK server subprocesses under concurrent first use.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At python/browser_use/runtime.py, line 75:

<comment>Runtime startup is race-prone and can spawn multiple SDK server subprocesses under concurrent first use.</comment>

<file context>
@@ -0,0 +1,218 @@
+            await self._process.stdin.drain()
+
+    async def start(self) -> None:
+        if self._process is not None and self._process.returncode is None:
+            return
+        command = self.command or _default_sdk_server_command(self.state_dir)
</file context>

[cubic-dev-ai]

P1: Pending RPC calls can hang forever when server stdout closes without a response.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At python/browser_use/runtime.py, line 119:

<comment>Pending RPC calls can hang forever when server stdout closes without a response.</comment>

<file context>
@@ -0,0 +1,218 @@
+    async def _read_stdout(self) -> None:
+        assert self._process is not None
+        assert self._process.stdout is not None
+        async for raw_line in self._process.stdout:
+            line = raw_line.decode("utf-8", errors="replace").strip()
+            if not line:
</file context>

[cubic-dev-ai]

P2: Runtime-backed legacy wait_agent regressed: multi-target requests now fail despite the tool contract allowing multiple target IDs.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At crates/browser-use-agent/src/tools/handlers/subagent.rs, line 1585:

<comment>Runtime-backed legacy wait_agent regressed: multi-target requests now fail despite the tool contract allowing multiple target IDs.</comment>

<file context>
@@ -1483,140 +1536,107 @@ fn store_list_agents(
+        ),
+        _ => {
+            return Err(ToolError::Other(anyhow::anyhow!(
+                "runtime-backed wait_agent accepts at most one target; omit targets to wait for any child"
+            )));
         }
</file context>

[cubic-dev-ai]

P2: Write/drain failures leak pending RPC futures because _pending is not cleaned up on I/O errors.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At python/browser_use/runtime.py, line 58:

<comment>Write/drain failures leak pending RPC futures because `_pending` is not cleaned up on I/O errors.</comment>

<file context>
@@ -0,0 +1,218 @@
+                "params": params or {},
+            }
+            self._process.stdin.write((json.dumps(request) + "\n").encode("utf-8"))
+            await self._process.stdin.drain()
+        try:
+            return await future
</file context>

Suggested change

	await self._process.stdin.drain()
	try:
	try:
	self._process.stdin.write((json.dumps(request) + "\n").encode("utf-8"))
	await self._process.stdin.drain()
	except (BrokenPipeError, ConnectionResetError, ProcessLookupError) as error:
	self._pending.pop(request_id, None)
	raise BrowserUseProtocolError("Rust SDK server stdin is unavailable") from error

[cubic-dev-ai]

P2: Cancellation errors can be misclassified as failures because only session.failed is checked before appending a new failure event.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At crates/browser-use-agent/src/live_executor.rs, line 472:

<comment>Cancellation errors can be misclassified as failures because only `session.failed` is checked before appending a new failure event.</comment>

<file context>
@@ -0,0 +1,789 @@
+        .map(|events| {
+            events
+                .iter()
+                .any(|event| event.event_type == "session.failed")
+        })
+        .unwrap_or(false);
</file context>

[cubic-dev-ai]

P2: Agent startup errors are now propagated with ?, which can terminate the command flow instead of recording a session failure and continuing UI operation.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At crates/browser-use-tui/src/main.rs, line 3508:

<comment>Agent startup errors are now propagated with `?`, which can terminate the command flow instead of recording a session failure and continuing UI operation.</comment>

<file context>
@@ -3262,43 +3505,18 @@ impl App {
-                }
-            })
-            .context("spawn agent thread")?;
+        spawn_tui_agent_run(
+            state_dir,
+            session_id,
</file context>

[cubic-dev-ai]

2 issues found across 2 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="python/browser_use/browser.py">

<violation number="1" location="python/browser_use/browser.py:100">
P1: `Browser.close()` clears `browser_id` even when the runtime close call fails, which can orphan the remote browser and prevent retrying cleanup.</violation>
</file>

<file name="crates/browser-use-tui/src/main.rs">

<violation number="1" location="crates/browser-use-tui/src/main.rs:3508">
P2: Agent startup errors are now propagated with `?`, which can terminate the command flow instead of recording a session failure and continuing UI operation.</violation>
</file>

<file name="crates/browser-use-agent/src/tools/handlers/subagent.rs">

<violation number="1" location="crates/browser-use-agent/src/tools/handlers/subagent.rs:1585">
P2: Runtime-backed legacy wait_agent regressed: multi-target requests now fail despite the tool contract allowing multiple target IDs.</violation>
</file>

<file name="python/browser_use/runtime.py">

<violation number="1" location="python/browser_use/runtime.py:58">
P2: Write/drain failures leak pending RPC futures because `_pending` is not cleaned up on I/O errors.</violation>

<violation number="2" location="python/browser_use/runtime.py:75">
P1: Runtime startup is race-prone and can spawn multiple SDK server subprocesses under concurrent first use.</violation>

<violation number="3" location="python/browser_use/runtime.py:119">
P1: Pending RPC calls can hang forever when server stdout closes without a response.</violation>
</file>

<file name="crates/browser-use-agent/src/live_executor.rs">

<violation number="1" location="crates/browser-use-agent/src/live_executor.rs:472">
P2: Cancellation errors can be misclassified as failures because only `session.failed` is checked before appending a new failure event.</violation>
</file>

<file name="crates/browser-use-agent/src/turn/sampling.rs">

<violation number="1" location="crates/browser-use-agent/src/turn/sampling.rs:604">
P1: Retryable streamed provider errors are now treated as terminal failures, bypassing the existing retry policy.</violation>
</file>

<file name="crates/browser-use-tui/src/runtime.rs">

<violation number="1" location="crates/browser-use-tui/src/runtime.rs:215">
P2: Non-atomic runtime lazy initialization can race and start duplicate runtime servers for the same state dir.</violation>
</file>

<file name="crates/browser-use-tui/src/transcript.rs">

<violation number="1" location="crates/browser-use-tui/src/transcript.rs:635">
P2: Active-work checks are evaluated eagerly, causing unnecessary runtime count lookups on every model rebuild.</violation>
</file>

<file name="crates/browser-use-browser/src/lib.rs">

<violation number="1" location="crates/browser-use-browser/src/lib.rs:551">
P1: Session busy-check and session creation are racy, so a default session can be created while the real session is temporarily checked out.</violation>

<violation number="2" location="crates/browser-use-browser/src/lib.rs:3271">
P2: Timed-out CDP connect leaves detached worker threads running, which can leak threads across retries.</violation>
</file>

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Fix all with cubic | Re-trigger cubic</sub>

[cubic-dev-ai]

P1: Session busy-check and session creation are racy, so a default session can be created while the real session is temporarily checked out.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At crates/browser-use-browser/src/lib.rs, line 551:

<comment>Session busy-check and session creation are racy, so a default session can be created while the real session is temporarily checked out.</comment>

<file context>
@@ -508,29 +540,52 @@ pub fn run_browser_command_with_options_and_registries(
+                .sessions
+                .lock()
+                .expect("browser session registry poisoned");
+            let session = sessions.entry(session_id.to_string()).or_default();
+            session.session_id = Some(session_id.to_string());
             session.log(format!("browser {}", argv.join(" ")));
</file context>

[cubic-dev-ai]

P2: Timed-out CDP connect leaves detached worker threads running, which can leak threads across retries.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At crates/browser-use-browser/src/lib.rs, line 3271:

<comment>Timed-out CDP connect leaves detached worker threads running, which can leak threads across retries.</comment>

<file context>
@@ -3176,17 +3264,25 @@ impl CdpDispatcher {
-        Self::from_socket(socket)
+        }
+
+        let ws_url = ws_url.to_string();
+        let (tx, rx) = std::sync::mpsc::channel();
+        thread::spawn(move || {
</file context>