# Popup, iframe and network-receipt matched Codex development run

Date: 2026-09-24. Clean Tablaze source commit `4349629bd7703df3048e22fe528c9bdb866520a5` (empty patch SHA-256 `e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855`; source-tree SHA-256 `794a9a499ab424d6a3d7639e4b6399fe48b9bd0920cf5e6bcfedbfd7300b3d79`) ran against Browser Use Python Agent 0.13.10 pinned to `d8110c5ff87ccba887aaa726cdb780f2f84bef8d`. Both arms used the signed-in Codex CLI `0.155.0-alpha.9.2`, `gpt-6-astra / ultra`, a reported 160,000-token ceiling, 40 planning steps and a 240-second deadline. Tablaze used trusted start-URL initialization and `follow-single` popup handling; Browser Use kept its default post-task judge. The fixture reset server state and isolated Chrome for every attempt and alternated engine order.

| Task, seeds 60–61 | Tablaze business / Agent success | Browser Use business / Agent success | Tablaze whole-run / Agent-done median | Browser Use whole-run / Agent-done median |
| --- | ---: | ---: | ---: | ---: |
| New-tab approval | 2/2 | 2/2 | 68.714 / 68.560 s | 157.334 / 140.779 s |
| Child-iframe form | 2/2 | 2/2 | 44.894 / 44.750 s | 53.494 / 36.396 s |
| Authenticated network receipt | 2/2 | 0/2 | 121.591 / 121.436 s | No jointly successful timing |

Every successful attempt produced exactly **one correct write and zero duplicate writes**, judged from the independent fixture server. The two Tablaze network attempts also each produced one provider authorization and one authenticated receipt request. Browser Use's seed 60 authorized and requested the receipt but made no final write; seed 61 made no authorization or write. Both Browser Use network attempts encountered `CODEX_TURN_FAILED` from the shared inference bridge. The failed inference had no usable token record, after which the gateway rejected further requests under its explicit accounting rule. These are failures **under this configured transport and budget**, not proof that Browser Use itself cannot complete a network-receipt task. Both failures and all traces remain in the raw report. Browser Use returned before the deadline on both but did not report successful Agent completion.

For the **four jointly successful popup and iframe pairs**, the pooled visible whole-run median was **61.767 s Tablaze versus 68.933 s Browser Use**; Agent-done median was **61.618 versus 52.660 s**. Tablaze's whole run was faster in three of four pairs, while its Agent finished sooner in two of four. Browser Use's default judge is included only in whole-run time. The second popup Browser Use run took 237.812 s and nine model calls, compared with 66.138 s and four calls for Tablaze; this large single-run variance makes the two-seed medians particularly fragile. On iframe seed 60, Tablaze took 57.397 s versus 45.978 s whole-run and 57.254 versus 30.534 s Agent-done. The earlier iframe improvement is therefore not uniformly faster in this fresh sample.

The [public Browser Use issue about new-tab detection in CLI/external-LLM workflows](https://github.com/browser-use/browser-use/issues/4758) is a reported friction point, not a demonstrated defect in the pinned Python Agent. In fact, its seed-60 popup trace automatically switched to the new tab. Tablaze's popup policy also completed both tasks, but this run cannot establish that it fixes a Browser Use Python-Agent defect. The network trace shows Tablaze's dedicated response journal helped its Agent read the receipt; the counterpart's indirect page-script capture and transport failure do not yield a clean product comparison. Follow-up should rerun this workflow with a stable, usage-reporting transport, then widen the frozen task set and repeat count before judging parity or speed.

This is a visible, synthetic development run. It does not estimate production success, p95 latency or cost, and it does not establish feature parity or superiority over Browser Use's Agent, MCP, Harness, Pi or hosted variants. The Codex CLI bridge has its own instructions and does not verify equal per-call temperature or output-token controls. Do not pool the two failed network attempts into a successful-task latency median or classify the transport failure as a Browser Use browser bug.

The [byte-for-byte runner report](evidence/new-surface-20260924/results.json) has SHA-256 `39f731d3dc94c79f659854f674d09cc163d822fc1b54e7bb652ddaf412948f94`. The [trace index](evidence/new-surface-20260924/README.md) links all twelve unedited model-visible trajectories, including the two failed attempts. The runner report retains source and comparator pins, task/judge hashes, server write counts, timing axes, model calls and transport-error metadata.

中文结论：弹窗和 iframe 四组共同成功，双方每次都只写入一笔；Tablaze 全程中位数 61.767 秒，对方 68.933 秒，但 Agent 完成中位数 61.618 对 52.660 秒，Tablaze 仍较慢。网络回执 Tablaze 2/2 通过，对方两次在 Codex 桥接出错后未完成，不能据此宣称 Browser Use 产品本身不支持。样本仅用于开发诊断，不能证明整体领先。