Add reproducible GPT-5.5 browser stability experiment

[waple0820]

What changed

• define the paired GPT-5.5 / LexBench-Browser protocol for Lexmount versus local Chrome
• pin browseruse-agent-bench and the 210-task dataset hash
• add secret-free runtime config and fixed smoke/pilot task sets
• add bounded Lexmount session-capacity probing with cleanup reconciliation
• add Linux cgroup/PSS/Chrome/GPU/vLLM resource sampling
• add deterministic task selection, run summarization, and paired bootstrap comparison
• record the current preflight evidence and remaining target-host blocker

Why

This repository is the continuation point for the environment-correctness and resource-efficiency work. The benchmark implementation remains upstream in browseruse-agent-bench; this repo keeps only experiment-specific protocol, tooling, and reviewed summaries.

Current evidence

• both Lexmount projects passed c20 session probes with active=0 after cleanup
• the English project exceeded the old 20-session limit, reaching 27 active sessions during c40, but saturated requests queued for more than 180 seconds
• one real GPT-5.5 Lexmount task completed in 6 steps / 105.242s at $0.23505 agent cost
• gpt-5.4 stepwise Judge scored that task 61/100 and the framework threshold passed
• the 5090 host is currently unreachable over SSH, so local Chrome and Linux resource sampling remain unverified

Validation

• uv run pytest (6 passed)
• uv run ruff check .
• bash -n scripts/run_benchmark.sh
• real Lexmount session create/cleanup probes
• real GPT-5.5 browser-use trajectory plus gpt-5.4 evaluation

[github-actions]

Claude finished @waple0820's task in 1m 36s —— View job

---

PR Review

• Gather context
• Read key files (scripts, config)
• Identify bugs and security issues
• Post review

---

Two bugs found:

1. Unhandled exception in JSONL fallback parser — scripts/summarize_run.py:46

except json.JSONDecodeError:
    records = [json.loads(line) for line in text.splitlines() if line.strip()]

If the outer json.loads(text) fails (i.e., input is JSONL rather than a JSON array), the fallback list comprehension calls json.loads(line) per line. Any malformed line in the JSONL file will raise json.JSONDecodeError here unhandled, crashing the script with a traceback rather than a useful error message. The except block only catches the first call. Fix this →

---

2. Shared client used concurrently in reconciliation — scripts/probe_lexmount_sessions.py:192-196

with ThreadPoolExecutor(max_workers=max(1, min(len(new_records), 32))) as executor:
    for error in executor.map(
        lambda item: _cleanup_session_record(control_client, item), new_records
    ):

control_client is a single Lexmount instance shared across multiple threads here. The earlier cleanup (lines 180–183) avoids this by giving each SessionHandle its own client. If the Lexmount HTTP client is not thread-safe (connection reuse, shared state), concurrent deletes will race. At minimum this is inconsistent with the pattern used elsewhere; at worst it causes silent failures or crashes during reconciliation. Fix this →

[github-actions]

Automated Claude review completed. Review run: https://github.com/lexmount/LexBrowserEnv/actions/runs/29198648439. Please check Claude's inline feedback and summary comment for details.