mirror of
https://github.com/browser-use/browser-use.git
synced 2026-09-14 19:59:47 +08:00
Saurav Panda
64981e41b8
Small local models handle reasoning better as a bounded JSON field than as native reasoning tokens. This adds that as an opt-in flash mode variant and wires it into the eval harness so it can be measured rather than assumed. type_with_custom_actions_flash_mode gains include_thinking and include_memory. Properties are rebuilt rather than deleted piecemeal because schema field order is emission order, and reasoning is only useful before the action it justifies. FLASH_THINKING_DESCRIPTION caps the field at 120 words; uncapped, a chatty model spends its latency budget on prose. thinking was never persisted into message history, so reasoning was write-only and could not inform later steps. HistoryItem now carries it, rendered ahead of the fields it explains. Persistence is gated on include_thinking_in_history: outside flash mode thinking is unbounded and would balloon every prompt. system_prompt_flash_thinking.md describes the matching contract. The four existing flash prompts are untouched, and plain flash mode keeps its exact schema, prompt and history behaviour - covered by explicit regression tests. beta mirrors the new constructor params, which tests/ci/test_beta_agent.py requires to match the main agent's signature and order. The eval harness gains a matching profile. It also drops the wait action via the existing exclude_actions hook: wait defaults to 3s and sleeps seconds-1, so every call the model makes costs ~2s of wall clock for nothing. Fixes max_history_items in the fastest profiles, which was set below the message manager's assert of > 5 and would have failed every task.