Cut eval cost −19.5%, score held (93 → 94)

[gregpr07]

real_v8, gpt-5.5, 100 tasks, locked judge. $66.84 → $53.82/run.

A token meter over the baseline showed 71% of cost is uncached input (page-HTML + wasted turns replayed every model call) — the prompt is ~99% cached. So: cut turns and page-garbage, not the prompt.

The 3 changes

1. Sync browser_script (lib.rs) — scripts were async; the model burned whole turns polling observe ("done yet?"). Task 6 spent 57/80 turns polling and failed. Start-wait 15s→30s (most scripts return in one call), observe hint 1s→15s. Floor stays 1s — avoids the reverted "observe30" stacking regression.
2. Un-blockable fetch (browser_script_helpers.py) — http_get fell back to raw urllib (bot-blocked) when fetch_use wasn't installed. Vendored the proxy client inline → always routes via Browser-Use Fetch (Chrome TLS + rotating IPs). Task 26: 50 → 684 stores.
3. Prompt trim (prompts/*) — ~900 tokens cut. Honest: saved only ~$0.46 (it's cached) and likely caused 2 nav regressions (72, 98). Revertable as its own commit (64d301c) if they don't recover.

Fails

• Fixed (the targeted family): 1, 6, 21, 26,
• New: 33, 52 (variance) · 72, 98 (nav, suspect the trim)
• Calibration anchors 68, 74 still fail as required.

Runs: real-v8-everything-20260611-211259 vs real-v8-phase1merged-20260610-234722. Token deltas exact; $ priced at labeled GPT-5 rates.

---

Summary by cubic

Cut eval run cost by 19.5% ($66.84 → $53.82/run) on real_v8 while holding score (93 → 94) by reducing observe churn and routing public http_get calls through an unblockable fetch proxy with safe private-host bypass. Restores the browser tool prompt to main to avoid navigation regressions; other prompt trims remain.

• Performance

  • browser_script: raise initial wait to 30s (was 15s) so fast scripts finish in one call; raise observe hint to 15s (was 1s). The 1s observe floor remains.
  • Prompts: keep compressed browser_script/Python/update_goal descriptions; revert the browser tool description to main (cached, minimal cost impact).

• Bug Fixes

  • http_get: route via the Browser‑Use Fetch proxy when BROWSER_USE_API_KEY is set, but never proxy loopback/private/intranet hosts by default; add use_proxy (None=auto/public-only, True=force, False=direct); prefer fetch_use if installed, else a vendored client; on proxy failure fall back to direct and surface the proxy error.
  • Tests: add coverage for vendored-proxy path, private-host bypass, forced proxy, and fallback error surfacing.

^{Written for commit a55c03b. Summary will update on new commits.}