luum-home/luum-cognitive-os/browser-task

Browser Task

Drive a real browser for navigation, extraction, and form-fill workflows.

Trigger

Use this skill when the work needs a live browser - not a plain HTTP fetch. Typical cases:

• "Navigate to a URL and extract the price table."
• "Fill out a form and record the confirmation page."
• "Scrape a dashboard after logging in."

If a plain HTTP fetch (no JS, no auth, no interaction) suffices, prefer WebFetch - it is cheaper and faster.

Prerequisites

1. The web-automation optional dependency must be installed via:

uv sync --extra web-automation.

1. One of the supported LLM credentials must be available in env, OR the

hosted ChatBrowserUse endpoint will be used (pricing: $0.20 / $2.00 per million input / output tokens).

1. The kill switch COS_DISABLE_WEB_AUTOMATION=1 MUST NOT be set. If it is,

the router will refuse and you must escalate to the operator.

How to invoke

From Python (a skill or sub-agent)

from cos_lib.web_automation_router import route, WebAutomationUnavailable
from cos_lib.dispatch_cost_predictor import predict_call_cost
try:
    adapter = route(
        "Navigate to the target URL and extract the headline.",
        llm_provider="browser_use",
        cost_predictor=predict_call_cost,
    )
    result = await adapter.run_task(
        "Navigate to the target URL and extract the headline.",
        max_steps=20,
        headless=True,
    )
except WebAutomationUnavailable as exc:
    # Surface to operator; do NOT silently fall back to a non-browser tool.
    raise

From a synchronous context

result = adapter.run_task_sync(task, max_steps=20)

Result shape

WebAutomationResult has:

• success: bool
• final_url: str
• screenshots_paths: list[str]
• extracted_data: dict
• error: str (empty on success)
• token_usage: {"input": int, "output": int}
• cost_usd: float
• steps: int

Kill switch

Set the env var COS_DISABLE_WEB_AUTOMATION=1.

When set, both cos_lib.web_automation_router.route() and cos_lib.browser_use_adapter.BrowserUseAdapter(...) raise WebAutomationUnavailable. This is the operator-level off switch - there is no silent fallback.

Cost model

• Owned providers (anthropic, openai): cost flows through

cos_lib.dispatch_cost_predictor.predict_call_cost(provider, ...) exactly like any other LLM dispatch. The ADR-228 session budget gate sees web automation as a normal cost source.

• `ChatBrowserUse` (hosted): $0.20 / $2.00 per million input / output

tokens, computed directly by the adapter.

A failed run still records the tokens it consumed. Partial cost is honest cost.

Defaults and limits

• headless=True by default. Override only when an operator is watching.
• max_steps=50 by default. Raise only after a cost prediction shows it is

worth it; a runaway loop is the most likely way to overspend.

• Untrusted URLs: set COS_BROWSER_USE_TRUSTED_ONLY=1 and gate the URL at

the call site until the future dispatch-integration ADR lands sandbox preflight (ADR-232) for browser runs.

Troubleshooting

References

• ADR-288 - Web-Automation Adapter for Dispatch.
• Upstream: browser-use (MIT).
• Cost flow: lib/dispatch_cost_predictor.py.
• Event emission: packages/agent-coordination/lib/agent_bus.py::AgentPublisher.