Skip to content

Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.

Record oversized eval history persistence proof

Magnus Müllerfcf98176067f
diff --git a/RUST_AGENT_INTEGRATION_FEATURES.md b/RUST_AGENT_INTEGRATION_FEATURES.md
index 51e82255a..f87b5fa4b 100644
--- a/RUST_AGENT_INTEGRATION_FEATURES.md
+++ b/RUST_AGENT_INTEGRATION_FEATURES.md
@@ -1473,6 +1473,14 @@ Terminal core branch: `magnus/browser-use-rust-main-integration` at terminal mai
- Observed general risks: long trajectories were the main cost/speed issue (tasks 1 and 5 ran 428s/436s and cost `$7.33`/`$3.49`), Laminar export hit resource/deadline limits on very large traces, and one AUD vendor task reached a Cloudflare captcha despite Browser Use cloud proxy. These are not yet fixed because this gate showed them as evidence to investigate across more tasks rather than single-domain overfitting.
- Proof: browser-use `uv run pytest -q tests/ci/test_rust_agent.py -k 'process_failure_ignores_empty_stream_text or loads_started_session_events_after_terminal_timeout or laminar_run_summary' --disable-warnings`, browser-use `uv run python -m py_compile browser_use/rust/service.py tests/ci/test_rust_agent.py`, evaluations-internal `PYTHONPATH=. uv run pytest -q tests/test_service_cli.py -k 'uses_usage_total_tokens_when_metadata_missing or synthesizes_partial_result_on_timeout or pre_judge_checkpoint'`, evaluations-internal `PYTHONPATH=. uv run python -m py_compile eval/service.py tests/test_service_cli.py`, and dashboard query `/api/getRunResults?runId=kh703qjpt46yfdkyv741pecgq58809h6&format=only_judge`.
+280. Rust eval oversized dashboard history persistence
+ - The eval runner now trims only Convex-facing `actionHistory` and `completeHistory` text fields before save, keeping large extracted-content runs below the Convex 1 MiB field limit while preserving local `saved_trajectories` files, screenshots, final result text, structured usage/cost data, and Laminar traces.
+ - This addresses a general runner failure from real_v8 task 6 where a single `Captured requests` extraction was about 2.29 MiB and was duplicated into both dashboard history fields. Before the fix, the agent and judge ran but final `saveTaskResult` failed with `Value is too large`, making the task appear missing/empty.
+ - Five-task real_v8 local-platform gate `kh71j6cq0cwegjjb6rh4c4840h881wm8` used local browser-use `6c7e4e38`, terminal `9102fb2`, `BU_RUNTIME=brust`, Browser Use cloud CDP with US proxy, `--parallel-runs 5`, `--max-steps 100`, `--task-timebox-seconds 600`, Claude Sonnet 4.6 executor, and Agent SDK judge for tasks 6-10.
+ - Gate outcome before the persistence fix was 4/5 saved results: task 7 `1.0`, task 8 `0.8`, task 9 `0.85`, task 10 `0.5`; task 6 reached judge score `0.0` but failed final dashboard save on the oversized payload. The recorded four-task mean was `0.7875`; treating task 6 as zero gives `0.63`.
+ - The same gate proved the Rust usage dashboard fix: saved tasks now show nonzero `tokensUsed` from `history.usage.total_tokens` and include cache/cost buckets in `usage`.
+ - Proof: evaluations-internal `PYTHONPATH=. uv run python -m py_compile eval/task_types.py tests/test_service_cli.py`, evaluations-internal `PYTHONPATH=. uv run pytest -q tests/test_service_cli.py -k 'server_payload_trims_oversized_history_fields or pre_judge_checkpoint or uses_usage_total_tokens_when_metadata_missing'`, evaluations-internal `git diff --check -- eval/task_types.py tests/test_service_cli.py`, and dashboard query `/api/getRunResults?runId=kh71j6cq0cwegjjb6rh4c4840h881wm8&format=only_judge`.
+
## Current Verification
- terminal `cargo test -p browser-use-agent driver_passes_populated_per_call_request_to_open_stream -- --nocapture`