Fix benchmark regression

[gregpr07]

Summary by cubic

Restore 88-baseline behavior and add quick wins: faster observe polling, full script output visibility, safer/clearer completion, and smarter prompts/audits to cut wasted turns and finish runs with correct results. Also adds a k=2 validation report showing K1=82, K2=90 (mean 86).

• Bug Fixes

  • Observe: default back to 1s (agent and browser lib), removed the hidden 30s eval coupling, and when a run is already finished, observe now says so and stops suggesting more polls.
  • Output visibility: raised inline browser_script stdout cap to 120KB (matches script limit) and updated truncation guidance so the model recovers missing data instead of guessing.
  • Completion: successful done is terminal; if a bounded run ends on plain text, the loop gives one “call done” nudge; if the model ends without done, emit the best result.* artifact from cwd as the final result.
  • Compaction: fixed empty-summary handling so it falls back to “(no summary available)” before prefixing, preventing post-compaction amnesia.
  • Audits and prompts: null and "" no longer count as placeholders; wording-based rejections removed; prompts drop the one-pass ceiling, add a visual fallback, forbid fabricated/proxy values, require basic reality checks, and add efficiency rules (chunking, checkpointing, limited retries, concurrent fetch, and observe discipline, including reading checkpoint files).
  • Unified exec: added timeout_ms with a safe default and kill the entire process group on timeout to prevent stray processes.
  • Added BROWSER_USE_DISABLE_LOCAL_SEARCH to optionally disable local search while keeping hosted web_search.
  • Docs: added k=2 quick-pack validation report (K1=82, K2=90, mean 86).

• Dependencies

  • Added libc for process-group management on Unix.

^{Written for commit da14329. Summary will update on new commits.}

[cubic-dev-ai]

3 issues found across 22 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:311">
P2: `read_result_file_preview` reads the full file despite the preview cap, causing avoidable large I/O and memory usage.</violation>
</file>

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

<violation number="1" location="crates/browser-use-browser/src/lib.rs:279">
P2: Completed browser_script cache limits were increased enough to risk significant memory growth in long-running sessions.</violation>
</file>

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

<violation number="1" location="crates/browser-use-agent/src/entrypoint/mod.rs:2700">
P1: `session.done` is emitted for aborted turns because `Ok(...)` from `TurnLoop` includes interruption/max-turns paths.</violation>
</file>

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

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

[cubic-dev-ai]

P1: session.done is emitted for aborted turns because Ok(...) from TurnLoop includes interruption/max-turns paths.

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 2700:

<comment>`session.done` is emitted for aborted turns because `Ok(...)` from `TurnLoop` includes interruption/max-turns paths.</comment>

<file context>
@@ -2611,10 +2674,83 @@ impl<Sd: SamplingDriver> RuntimeTurnLoopDriver<Sd> {
+                        result_file = Some(fallback.file);
+                    }
+                }
+                if let Some(text) = final_message.as_deref() {
+                    runtime_handle.append_observed_session_event(
+                        runtime_session_id,
</file context>

[cubic-dev-ai]

P2: read_result_file_preview reads the full file despite the preview cap, causing avoidable large I/O and memory usage.

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 311:

<comment>`read_result_file_preview` reads the full file despite the preview cap, causing avoidable large I/O and memory usage.</comment>

<file context>
@@ -224,3 +229,240 @@ impl ToolRuntime<DoneRequest, ExecOutput> for DoneTool {
+        return Err(format!("result_file `{}` is empty", path));
+    }
+
+    let bytes = fs::read(&resolved).map_err(|error| {
+        format!(
+            "result_file `{}` could not be read at {} ({error})",
</file context>

[cubic-dev-ai]

P2: Completed browser_script cache limits were increased enough to risk significant memory growth in long-running sessions.

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 279:

<comment>Completed browser_script cache limits were increased enough to risk significant memory growth in long-running sessions.</comment>

<file context>
@@ -276,8 +276,8 @@ static MANAGED_BROWSER_PIDS: OnceLock<Mutex<HashSet<u32>>> = OnceLock::new();
 static BROWSER_SCRIPT_RUN_COUNTER: AtomicU64 = AtomicU64::new(1);
-const BROWSER_SCRIPT_COMPLETED_CACHE_TTL_MS: u128 = 10 * 60 * 1_000;
-const BROWSER_SCRIPT_COMPLETED_CACHE_MAX: usize = 128;
+const BROWSER_SCRIPT_COMPLETED_CACHE_TTL_MS: u128 = 60 * 60 * 1_000;
+const BROWSER_SCRIPT_COMPLETED_CACHE_MAX: usize = 2048;
 const MANAGED_BROWSER_PROFILE_PREFIX: &str = "but-managed-browser.";
</file context>

Suggested change

	const BROWSER_SCRIPT_COMPLETED_CACHE_TTL_MS: u128 = 60 * 60 * 1_000;
	const BROWSER_SCRIPT_COMPLETED_CACHE_MAX: usize = 2048;
	const BROWSER_SCRIPT_COMPLETED_CACHE_TTL_MS: u128 = 10 * 60 * 1_000;
	const BROWSER_SCRIPT_COMPLETED_CACHE_MAX: usize = 128;

[cubic-dev-ai]

1 issue found across 10 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="crates/browser-use-agent/src/turn/loop_driver.rs">

<violation number="1" location="crates/browser-use-agent/src/turn/loop_driver.rs:253">
P2: The new done-nudge path can exceed `max_turns` by one sampling round, violating the bounded-run cap.</violation>
</file>

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

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

[cubic-dev-ai]

P2: The new done-nudge path can exceed max_turns by one sampling round, violating the bounded-run cap.

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/loop_driver.rs, line 253:

<comment>The new done-nudge path can exceed `max_turns` by one sampling round, violating the bounded-run cap.</comment>

<file context>
@@ -229,6 +244,18 @@ impl<St: TurnState, Sd: SamplingDriver, Ob: TurnObserver> TurnLoop<St, Sd, Ob> {
+                    // to call done before we accept the text as final.
+                    let ended_without_done =
+                        !matches!(outcome.finish_reason, Some(FinishReason::ToolUse));
+                    if max_turns.is_some() && ended_without_done && !done_nudge_sent {
+                        done_nudge_sent = true;
+                        pending_done_nudge = true;
</file context>

Suggested change

	if max_turns.is_some() && ended_without_done && !done_nudge_sent {
	if max_turns.is_some_and(|limit| turns_run < limit) && ended_without_done && !done_nudge_sent {