Module browser 

SOTA browser-bot harness: drive a real Chrome via the chrome-devtools MCP server, with the reliability and safety layer that makes a web agent actually work (per tasks/browser-bot-sota-2026-05-31.md).

The browser-control problem (“how to call Chrome”) is already solved: the MCP stdio client (crate::tool::mcp) turns the server’s tools into Arc<dyn Tool>, and the bundled chrome-devtools preset (crate::connect_preset) spawns it with one call. This module adds the agent-layer capabilities the research found decisive — starting with observation distillation (distill) — built on the existing agent runner, dynamic-workflow combinators, structured output, guardrails, and memory.

Built incrementally (spec §9, cheap reliability wins first):

• B1 (this): distill — prune the a11y snapshot before it reaches the LLM.
• B2+: stale-uid retry, post-action verification, settle, replan-loop, completion judge, browser guardrails. (See the spec.)

Re-exports

pub use bench::BenchResult;

pub use bench::BenchTask;

pub use bench::Oracle;

pub use bench::run_bench;

pub use bench::scorecard;

pub use builder::BROWSER_SYSTEM_PROMPT;

pub use builder::BrowserAgentBuilder;

pub use builder::browser_guardrails;

pub use builder::filter_tools;

pub use builder::wrap_browser_tools;

pub use confirm::ActionRisk;

pub use confirm::ConfirmActionTool;

pub use confirm::ConfirmPolicy;

pub use confirm::classify_label;

pub use confirm::label_for_uid;

pub use distill::DistillConfig;

pub use distill::distill_snapshot;

pub use distill::interactive_uids;

pub use distill_tool::DistillingTool;

pub use guard::DomainAllowlistGuard;

pub use harness::ReliableInteractionTool;

pub use inject::InjectionRisk;

pub use inject::scan_snapshot_for_injection;

pub use judge::CompletionVerdict;

pub use judge::build_completion_prompt;

pub use judge::parse_completion_verdict;

pub use plan::Plan;

pub use plan::PlanStep;

pub use plan::ReplanAction;

pub use plan::ReplanBudget;

pub use plan::StepOutcome;

pub use plan::StepStatus;

pub use plan::replan_decision;

pub use settle::Probe;

pub use settle::SettleConfig;

pub use settle::SettleOutcome;

pub use settle::parse_dom_ready;

pub use settle::settle;

pub use verify::ActionEffect;

pub use verify::SnapshotSignature;

pub use verify::diff;

pub use verify::signature;

Modules

bench

Live browser-agent benchmark — the “go to a real site and perform actions” smoke test, as a repeatable evaluation process.

builder

BrowserAgentBuilder — assembles a SOTA browser agent from the harness capabilities (spec §5.10, the capstone that ties the module together).

confirm

Destructive-action confirmation (capability 21 of the browser-bot spec).

distill

Snapshot distiller (capability 2 of the browser-bot spec).

distill_tool

DistillingTool — an Arc<dyn Tool> decorator that distills snapshot output before it re-enters the agent’s context (token-control wiring).

guard

Browser safety guardrails (spec capabilities 19–21).

harness

Browser reliability decorators over raw MCP tools (spec capability 4).

inject

Prompt-injection heuristic for page content (capability 20 of the spec).

judge

Task-completion judge (capability 15 of the browser-bot spec).

plan

Plan artifact + replan decision logic (capability 8 of the browser-bot spec).

settle

Wait-for-stability / network-idle settle (capability 6 of the browser-bot spec).

verify

Post-action verification via state-diff (spec capability 5).