[eric] browser: speed + robustness + learning tiers + UX (session work)

[ciregenz]

What this is

A large body of browser-agent work (191 commits, 107 files, ~13.5k insertions) accumulated on eric/dev. Opening as a draft for review before anything is merged or released. This is NOT validated end-to-end on a clean machine yet (see caveats), so it should not go straight to the stable channel.

Themes

Browser agent speed + robustness (the bulk)

• Converged the speed-lever A/B: shipped the two proven winners (think-shorter, merge-verify) unconditionally, deleted the dead-end flags (cheap-laps, compose-helper, send-prior all proved neutral/negative).
• Completion detection: a confirmed/clean Send-class click ends the run instead of stalling to re-verify (kills the post-send dither); patient target_only confirm so a late-rendering send is not falsely "not confirmed".
• Send-button handoff: poll the interactives list for the real Send button after a composer fill and hand the model its index, so it stops hunting via CSS/JS.
• Message integrity: never leave a tool_use unanswered (mid-turn backfill + two-way validator) so a 30s upstream reset stops 400-killing the run.
• WS resilience: ride out a momentary dashboard WS drop instead of failing the run under load.
• Surrogate sanitizer: strip orphan UTF-16 surrogates from webview text so emoji pages do not crash a turn.

Learning tiers

• Seed playbooks shipped for 30 popular sites (judgment, not mechanical replay) so a fresh install starts warm.
• Tier 3 cross-site meta-playbook (site-agnostic priors) + tier 4 self-audit of the learning machinery (proposes for human review, never self-edits, fires every 25 tasks in a daemon thread).

UX

• Agent writes a plain human final message (no UI jargon / OUTCOME mechanics leaking to users).
• Fast-path emits a synthetic BrowserAgent tool-call/result so the "Browser Agent" bubble always shows, even when the orchestrator is skipped.
• Electron: silenced the per-CDP-node ipc.handle log spam (OPENSWARM_DIAG_IPC=1 restores it).

Test state

• Backend suite: 1178 passing.
• The benchmark went from ~582s first-success to a ~5-6 turn / ~30-47s true-cold floor (with the shipped seed/meta priors), and the cold-with-memory floor at ~55-60s.

Honest caveats (why draft, why not straight to stable)

• This machine was under heavy load (Wallspace + Spotlight) for most of the session, so several behavior/UI fixes (friendly final message, send-handoff firing, the fast-path bubble) are unit-green but not visually confirmed on a clean machine. They need one clean-machine pass.
• Some A/B runs were load-confounded; the holistic bench chart labels session-2 wall times as load-confounded for that reason.
• Two files remain oversized (agent_manager.py 4357 lines, browser_agent.py 2203 lines); a phased split is filed as deferred tech-debt. The import tree itself is a clean downward DAG.

Next

Review the diff, then (once merged to main) cut an experimental/prerelease build first to soak before any stable promotion, rather than shipping all of this to every user at once.