feat(website_browser): record the browser with computer-recording/v1 (FD-4497)

.github/scripts/infra_build.py

@@ -50,8 +50,8 @@
 # Docker Hub send.
 _RATE_LIMITED = re.compile(r"429 Too Many Requests|toomanyrequests", re.IGNORECASE)
 # What pip in an agent image's build answers for the agentenv-framework-protocol its release published moments earlier,
-# until PyPI's index, which a CDN caches, lists it.
-_NOT_YET_SERVED = re.compile(r"No matching distribution found for agentenv-framework-protocol(\[[^\]]*\])?==")
+# until PyPI's index, which a CDN caches, lists it: pinned (an agent's), or a floor it names (the browser's recorder).
+_NOT_YET_SERVED = re.compile(r"No matching distribution found for agentenv-framework-protocol(\[[^\]]*\])?(==|>=)")
 # How long to wait before each try again, so a burst of builds starting together spreads out.
 _RETRY_WAITS_SECONDS = (30, 60, 120)
 

.github/scripts/infra_images.py

@@ -5,8 +5,8 @@
 ``plan`` prints each image's name, repository, tag, Dockerfile and context as JSON, ``--missing`` only those the
 registry hasn't published: it has no image under the tag, or refuses to show one, as GHCR does for a package that
 doesn't exist yet, so the first publish can start. ``--only infra`` lists only the infra images, and ``--only agents``
-only the agents' images, which a release builds after it publishes the protocol, as an agent's protocol floor may name
-it. ``write PATH`` records the digest the registry serves under every
+only the images that install the protocol, the agents' and the website browser's, which a release builds after it
+publishes the protocol, as their protocol floor may name it. ``write PATH`` records the digest the registry serves under every
 image's tag as a release wheel carries them, and fails if any is missing. ``public`` warns about each image a user, who
 pulls anonymously, can't. The registry is read with ``GHCR_TOKEN`` (as ``GITHUB_ACTOR``) when it's set, anonymously
 otherwise, and always anonymously by ``public``.
@@ -26,9 +26,10 @@
 from agent_env.store.image_store.registry_api import NoSuchImageError, RegistryRefusedError, pin_digest
 
 REPO = Path(__file__).resolve().parents[2]
-# "agents" is the shipped bundles' images, the agents' and an example env's, which install the protocol a release
-# publishes.
-_GROUPS = {"infra": tuple(image for image in PUBLISHED_IMAGES if image not in BUNDLE_IMAGES), "agents": BUNDLE_IMAGES}
+# "agents" is the images that install the protocol a release publishes: the shipped bundles', the agents' and an
+# example env's, and the website browser's, whose recorder installs it.
+_AFTER_PROTOCOL = BUNDLE_IMAGES + tuple(image for image in PUBLISHED_IMAGES if image.name == "agentenv-website-browser")
+_GROUPS = {"infra": tuple(image for image in PUBLISHED_IMAGES if image not in _AFTER_PROTOCOL), "agents": _AFTER_PROTOCOL}
 
 
 def main() -> int:

src/agent_env/env/envs/multi_env.py

@@ -39,6 +39,7 @@ def _snapshot_after_load_default() -> bool:
 )
 from agent_env.env.envs.mcp_server import MCPServerEnv
 from agent_env.env.envs.website import WebsiteEnv
+from agent_env.env.envs.website_browser import WEBSITE_BROWSER_RECORDER_PORT
 from agent_env.store.ids import derive_id
 from agent_env.providers.sandbox_providers.sandbox_provider import SANDBOX_MODE_CONTAINER
 from agent_env.attribution import Attribution
@@ -615,7 +616,8 @@ async def _load_from_snapshot(self, snapshot) -> None:
         ]
         if website_configs:
             website_browser_env = Env.get(config.default_website_browser_env_id)
-            mcp_servers = list(mcp_servers) + [MCPServerConfig(image=website_browser_env.docker_image_artifact.image_name, environment_name=website_browser_env.environment_name)]
+            mcp_servers = list(mcp_servers) + [MCPServerConfig(image=website_browser_env.docker_image_artifact.image_name, environment_name=website_browser_env.environment_name,
+                                                               card_port=WEBSITE_BROWSER_RECORDER_PORT)]
 
         compose_content = gw.create_docker_compose(
             mcp_servers=mcp_servers,

src/agent_env/env/envs/website_browser/Dockerfile

@@ -1,9 +1,10 @@
 FROM public.ecr.aws/docker/library/node:22-slim@sha256:813a7480f28fdadac1f7f5c824bcdad435b5bc1322a5968bbbdef8d058f9dff4
 
 # Chromium runtime dependencies for headless mode, and python3 for the socket probe the gateway's compose
-# health-checks every MCP server with.
+# health-checks every MCP server with and for the recorder.
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    python3-minimal \
+    python3 \
+    python3-venv \
     libnss3 \
     libnspr4 \
     libatk1.0-0 \
@@ -38,7 +39,12 @@ RUN npm init -y && \
     npm install @playwright/mcp@${PLAYWRIGHT_MCP_VERSION} && \
     npx playwright install chromium
 
-COPY entrypoint.sh /app/entrypoint.sh
+# The recorder (recorder.py): computer-recording/v1 of the browser's page, idle until a recording starts.
+COPY recorder-requirements.txt /app/
+RUN python3 -m venv /opt/recorder && \
+    /opt/recorder/bin/pip install --no-cache-dir -r /app/recorder-requirements.txt
+
+COPY recorder.py entrypoint.sh /app/
 RUN chmod +x /app/entrypoint.sh
 
 CMD ["/app/entrypoint.sh"]

src/agent_env/env/envs/website_browser/__init__.py

@@ -6,7 +6,11 @@
 # The environment_name used when creating MCPServerEnv.
 WEBSITE_BROWSER_ENVIRONMENT_NAME = "website-browser"
 
+# The port the image's recorder serves the browser's env card on (recorder.py, entrypoint.sh's RECORDER_PORT).
+WEBSITE_BROWSER_RECORDER_PORT = 18766
+
 __all__ = [
     "WEBSITE_BROWSER_IMAGE_TAG",
     "WEBSITE_BROWSER_ENVIRONMENT_NAME",
+    "WEBSITE_BROWSER_RECORDER_PORT",
 ]

src/agent_env/env/envs/website_browser/entrypoint.sh

@@ -3,6 +3,16 @@ set -euo pipefail
 
 # MCP_PORT and MCP_HOST are injected by docker-compose (gateway convention).
 # PLAYWRIGHT_MCP_VERSION is baked into the image at build time.
+
+# The recorder serves the browser's card on RECORDER_PORT, restarted whenever it exits; the MCP server starts once
+# it listens, so the card is there by the time the server is healthy, or after RECORDER_WAIT_S without it.
+export RECORDER_PORT="${RECORDER_PORT:-18766}"
+(while true; do /opt/recorder/bin/python /app/recorder.py || true; sleep 1; done) &
+for _ in $(seq 1 "$(( ${RECORDER_WAIT_S:-20} * 10 ))"); do
+    (exec 3<>"/dev/tcp/127.0.0.1/${RECORDER_PORT}") 2>/dev/null && break
+    sleep 0.1
+done
+
 exec npx @playwright/mcp@${PLAYWRIGHT_MCP_VERSION} \
     --browser chromium \
     --headless \

src/agent_env/env/envs/website_browser/recorder-requirements.txt

@@ -0,0 +1,4 @@
+# The recorder installs the newest agentenv-framework-protocol at or above this floor, the first with computer-recording/v1;
+# raise it when the recorder needs a newer protocol.
+agentenv-framework-protocol[computer]>=0.1.382
+websockets==15.0.1