# Isolated local browser integration — no VM execution `run_pending.py` uses existing test bootstrap helpers to enroll a real mTLS device fixture on the local dev backend, never a forged database agent identity. It sends a ClientHello advertising `run_script` only to exercise plan validation; it does not implement an executor, ask for a permit, authorize or report an execution. The browser sends a durable pending request via the actual UI. The fixture is revoked afterwards; its pending request and enrollment history are retained. An artifact-count guard checks that jobs/outbox/grants did not increase. Requirements: initialized/running dev Docker backend on loopback 18443, built UI served from `web/dist`, public CA `devstate/pki/ca.pem`, private existing browser login `devstate/browser-login.json` (owned, mode 0600). Credential values are never CLI arguments, printed or copied into source. No VM credentials are used. Project-local tools (commands for reproducing setup; actual execution occurred in the implementation session): ```sh cd web npm ci --ignore-scripts PLAYWRIGHT_BROWSERS_PATH="$PWD/node_modules/.cache/playwright-browsers" npx playwright install chromium mkdir -p node_modules/.cache/nss-tools cd node_modules/.cache/nss-tools apt-get download libnss3-tools dpkg-deb -x libnss3-tools*.deb root ``` No package installed globally. The runner uses installed `certutil` if available, otherwise the locally extracted binary. It imports the **public** CA into isolated `web/node_modules/.cache/browser-home/.pki/nssdb`, with browser HOME restricted to that directory. Existing global browser/user trust and hosts files are untouched. DNS overrides map only fixed test names to 127.0.0.1; TLS/SNI still verifies the original DNS name. No certificate bypass flag or ignoreHTTPS option exists. From repository root: ```sh python3 deploy/browser/run_pending.py ``` Real Playwright tests verify normal trusted HTTPS plus rejection of wrong DNS name and an untrusted CA. The positive UI case only plans/submits and never clicks approval. CSRF is tested by deleting the token from an actual logout UI request; 403 and continued session visibility are required, followed by valid logout. No direct API/DB approval path is used. Reporter suppresses raw errors, action logs, stdout/stderr; traces/screenshots/video are off. Failure artifacts remain private and must not be published. Sanitized evidence records booleans/statuses only, no hostnames/keys/passwords. Bounded retries are not automatic: a failing run fails; a separate rerun must reauthenticate and creates a new explicit pending plan. `RMM_E2E_MODE=execute` is an additional separate opt-in for the prepared positive approval/result test, to use only after a real authorized agent path is ready. This pending-only run is neither native VM evidence nor positive approval proof.