Phase 2 of #452. Phase 1 (merged, #466) gives the agent a browser on the host via chrome-devtools-mcp behind --browser. That covers dev-machine runs.
Phase 2 is the sandboxed version: bake Chromium + chrome-devtools-mcp into the Docker runner image so the browser works when the run is isolated in a container.
Blocked on the agent-in-container move (#109). Today the agent runs on the host and only the app is served inside the runner, so an in-image browser has nothing pointing at it. This only pays off once the agent itself runs inside the sandbox.
Scope when unblocked:
Also worth deciding first, from Phase 1 real-world use: is browser access valuable enough to carry into the sandbox story at all?
Phase 2 of #452. Phase 1 (merged, #466) gives the agent a browser on the host via chrome-devtools-mcp behind
--browser. That covers dev-machine runs.Phase 2 is the sandboxed version: bake Chromium + chrome-devtools-mcp into the Docker runner image so the browser works when the run is isolated in a container.
Blocked on the agent-in-container move (#109). Today the agent runs on the host and only the app is served inside the runner, so an in-image browser has nothing pointing at it. This only pays off once the agent itself runs inside the sandbox.
Scope when unblocked:
chromium+ fonts +--no-sandbox.Also worth deciding first, from Phase 1 real-world use: is browser access valuable enough to carry into the sandbox story at all?