agent-browser-shield

Install the agent-browser-shield Chromium extension into a claw browser session at startup. Three install paths — local headed Chrome (unpacked load), Browserbase remote (ZIP upload), and CDP-attach to a Chrome where the extension is already loaded. Use at session bootstrap, before navigating to any target page.

Run an experiment to understand how the guarded extension affects an agent on a specific task — for one model or across several — and turn the observations into concrete rule-change proposals (tighten a built-in default, add a site-specific rule, drop a rule, or ship a new defense). Use when the user asks "research how guard does on task X", "is the extension helping on Y", "what should we change to do better on Z", "test this task across models", "should we add a site rule for this", or "see if our rules pay off here". NOT for explaining an existing benchmark row (use `agent-browser-shield-diagnose`), authoring new task CSV rows (use `agent-browser-shield-tasks`), adding selectors to a known site after you've already decided the recipe (use `agent-browser-shield-site-rules`), or extension build/install configuration (use `agent-browser-shield-install`).

How to operate a browser session with the agent-browser-shield Chromium extension active. Lists the DOM markers it leaves behind and the required agent behavior in response.

How to configure the agent-browser-shield Chromium extension — toggling individual rules, pasting JSON to apply a full configuration, and the full list of rule IDs. Use when the user wants to change which agent-browser-shield rules are active.

Install and operate the agent-browser-shield Chromium extension — masks PII/secrets, neutralizes dark patterns, and strips prompt-injection surfaces before the agent sees the page. Use at session bootstrap; covers headed-Chromium install and the agent behavior contract for handling its DOM markers.

Diagnose why a specific benchmark task underperformed on guarded vs baseline — whether that's a judge pass/fail regression or a cost regression (extra tokens, steps, or duration). Use when the user asks "why did task X fail on guarded", "why is the guarded run worse than baseline on Y", "why did guarded cost more on Z", "what went wrong with run_<id>", or wants to investigate a flaky or expensive (scenario, task) cell in a benchmark report.

How to add a new host to `extension/data/sites/*.yaml` using the Playwright MCP — per-rule extraction recipe, selector signal hierarchy, false-positive checks, validation, and one-PR-per-host workflow. Use when the user wants to extend site coverage for `reviews-redact`, `comments-redact`, `footer-redact`, or `search-url-helper`, or when extracting/refreshing selectors for an existing host.

Cut a release of the agent-browser-shield Chromium extension. CalVer YYYY.M.D.<github.run_number> versioning, workflow-driven — .github/workflows/release-extension.yml computes the version, patches the manifest, builds the ZIP, and creates the GitHub Release with the ZIP attached. Use when the user asks to ship, release, publish, or push a new build of the extension. NOT for editing rule code, running benchmarks, or local-only testing.

How to write or edit benchmark task rows in `benchmark/tasks.csv` so the extraction format is unambiguous and consensus comparison across repetitions doesn't trip on cosmetic differences. Use when the user is adding a new task, tightening an existing task's instructions, or diagnosing mismatched/judge-disagreement results in a run.