Fix/real v8 simple

[gregpr07]

Summary by cubic

Restores pre-regression real_v8 behavior with faster observe polling, full browser_script output visibility, terminal completion capture, and safer process execution. Improves eval reliability and preserves results even without an explicit done().

• New Features

  • Successful done is terminal in the fused loop; if no done, emit a final result from the best result* file in the session cwd.
  • exec_command adds timeout_ms and kills the whole process group on timeout (Unix); default cap DEFAULT_EXEC_COMMAND_TIMEOUT_MS=600_000 using libc.
  • Observe control: defaults set to 1s in browser-use-agent and browser-use-browser; eval caps now only follow BROWSER_USE_EVAL_MAX_OBSERVE_TIMEOUT_MS. Added BROWSER_USE_DISABLE_LOCAL_SEARCH to disable the local search tool.
  • Tool schema/docs: exec_command includes timeout_ms; done requires a complete, verified result with a step-exhaustion fallback.

• Bug Fixes

  • Inline browser_script stdout cap raised to 120 KB (matches SCRIPT_MAX_OUTPUT_CHARS); truncation guidance updated so the model re-reads artifacts when needed.
  • Compaction applies "(no summary available)" to the suffix before the prefix to avoid losing context on empty summaries.
  • Done-audit treats null and empty string values as genuine absences, not placeholders; reduces false rejections.
  • Prompts remove the one-pass repair ceiling, add visual fallbacks, and forbid fabricated values.
  • Reverts the 30s observe-window regression to 1s to prevent long polls from burning the run timebox.

^{Written for commit 84703c1. Summary will update on new commits.}

[cubic-dev-ai]

3 issues found across 23 files

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="crates/browser-use-agent/src/tools/handlers/done.rs">

<violation number="1" location="crates/browser-use-agent/src/tools/handlers/done.rs:261">
P2: Non-UTF8 `result_file` content is incorrectly treated as a valid material answer, allowing weak/empty completions to pass audit.</violation>
</file>

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

<violation number="1" location="crates/browser-use-agent/src/tools/handlers/done_tests.rs:195">
P2: The missing-file test uses a shared temp-dir filename, which can make the assertion flaky when that file already exists.</violation>
</file>

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

<violation number="1" location="crates/browser-use-agent/src/entrypoint/mod.rs:2685">
P1: Successful runs can fail before `session.done` if terminal cleanup errors, causing lost final result emission.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

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

[cubic-dev-ai]

P1: Successful runs can fail before session.done if terminal cleanup errors, causing lost final result emission.

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/entrypoint/mod.rs, line 2685:

<comment>Successful runs can fail before `session.done` if terminal cleanup errors, causing lost final result emission.</comment>

<file context>
@@ -2611,10 +2674,83 @@ impl<Sd: SamplingDriver> RuntimeTurnLoopDriver<Sd> {
+                    runtime_handle.clone(),
+                    runtime_session_id.clone(),
+                )
+                .await?;
+                // Never lose finished work: if the model ended without a final
+                // message (no done() call — e.g. ended on leaked planning text
</file context>

[cubic-dev-ai]

P2: Non-UTF8 result_file content is incorrectly treated as a valid material answer, allowing weak/empty completions to pass audit.

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/done.rs, line 261:

<comment>Non-UTF8 `result_file` content is incorrectly treated as a valid material answer, allowing weak/empty completions to pass audit.</comment>

<file context>
@@ -224,3 +229,244 @@ impl ToolRuntime<DoneRequest, ExecOutput> for DoneTool {
+    {
+        match read_result_file_preview(result_file, ctx) {
+            Ok(preview) => {
+                has_material_answer = true;
+                if let Some(preview) = preview {
+                    reasons.extend(json_audit_reasons(&preview));
</file context>

Suggested change

	has_material_answer = true;
	if let Some(preview) = preview {
	reasons.extend(json_audit_reasons(&preview));
	if !audit_text.is_empty() {
	audit_text.push('\n');
	}
	audit_text.push_str(&preview);
	}
	if let Some(preview) = preview {
	has_material_answer = true;
	reasons.extend(json_audit_reasons(&preview));
	if !audit_text.is_empty() {
	audit_text.push('\n');
	}
	audit_text.push_str(&preview);
	} else {
	reasons.push(format!("result_file `{}` is not valid UTF-8 text", result_file));
	}

[cubic-dev-ai]

P2: The missing-file test uses a shared temp-dir filename, which can make the assertion flaky when that file already exists.

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/done_tests.rs, line 195:

<comment>The missing-file test uses a shared temp-dir filename, which can make the assertion flaky when that file already exists.</comment>

<file context>
@@ -175,3 +175,70 @@ fn done_is_not_parallel_safe() {
+
+#[test]
+fn eval_done_audit_rejects_missing_result_file() {
+    let req = DoneRequest {
+        result_file: Some("missing-result.json".to_string()),
+        ..DoneRequest::default()
</file context>

Suggested change

	let req = DoneRequest {
	result_file: Some("missing-result.json".to_string()),
	..DoneRequest::default()
	};
	let temp = tempfile::tempdir().unwrap();
	let req = DoneRequest {
	result_file: Some(temp.path().join("missing-result.json").to_string_lossy().to_string()),
	..DoneRequest::default()
	};