[Iris] Add babysitting MCP server

.agents/skills/babysit-job/SKILL.md

@@ -70,17 +70,46 @@ If any required field is missing, ask for it before proceeding.
 - Sleep must be foreground (max ~10 min due to tool timeout).
 - Loop control is at agent level, not bash.
 
+## MCP-Assisted Monitoring
+
+When testing or using `marin-mcp-babysitter`, keep the MCP server resident and
+verify the job through MCP tools, not only through Iris CLI commands.
+
+- Keep the controller tunnel and MCP server in named, restartable sessions
+  (`screen`, `tmux`, or one long-running exec session). Record session names,
+  ports, and log paths in the state file.
+- Start MCP with a stable local controller URL and streamable HTTP transport:
+  `uv run --package marin marin-mcp-babysitter --controller-url <URL> --cluster <CLUSTER> --transport streamable-http --host 127.0.0.1 --port <PORT>`
+- Verify with `iris_job_summary` and `iris_tail_logs`. For heartbeat-style
+  monitoring, report: job state, latest progress/tick/log line, timestamp, and
+  error signal.
+- If the MCP server is reachable but tool calls fail with connection refused to
+  the controller URL, restart only the smoke-test tunnel/session. Do not restart
+  or mutate the Iris cluster.
+- If a sandbox blocks direct localhost TCP probes, run the probe inside an
+  existing long-lived session and write a small JSON result under `scratch/`.
+- For bounded smoke tests, create a thread heartbeat only after the job is
+  submitted, MCP is reachable, and at least one expected log/progress line has
+  appeared. Delete the heartbeat and stop the smoke-test sessions/listeners when
+  the job reaches the expected terminal state.
+
 ## State File
 
 Write to `scratch/<create_timestamp>_monitoring_state.json`, create the `scratch`
 directory if needed. `<create_timestamp>` should have format `YYYYMMDD-HHMM`.
-Track `restart_count` to detect flapping. State file allows resume after context reset.
+Track `restart_count` to detect flapping. Add MCP fields when a resident MCP
+server is part of the monitoring setup. State file allows resume after context reset.
 
 ```json
 {
   "ts": <timestamp_ms>,
   "job_id": "<JOB_ID>",
   "config": "<IRIS_CONFIG_PATH>",
+  "mcp_url": "http://127.0.0.1:<PORT>/mcp",
+  "tunnel_session": "<SESSION_NAME>",
+  "server_session": "<SESSION_NAME>",
+  "tunnel_log": "scratch/<TUNNEL_LOG>",
+  "server_log": "scratch/<SERVER_LOG>",
   "resubmit_command": "<IRIS_JOB_RUN_COMMAND_WITH_NO_WAIT>",
   "restart_count": 0
 }

lib/marin/pyproject.toml

@@ -17,6 +17,7 @@ dependencies = [
     "fasteners>=0.19",
     "flask>=3.1.3",
     "marin-fray",
+    "marin-iris",
     "marin-rigging",
     "fsspec>=2025.3.0",
     "gcsfs",
@@ -31,6 +32,7 @@ dependencies = [
     "lxml-html-clean>=0.4.4",
     "markdownify>=0.14.1",
     "multiprocess==0.70.16",
+    "mcp>=1.25.0",
     "numpy",
     "openai",
     "pandas>=2.0",
@@ -52,6 +54,9 @@ dependencies = [
 [project.license]
 file = "../../LICENSE"
 
+[project.scripts]
+marin-mcp-babysitter = "marin.mcp.babysitter:main"
+
 [dependency-groups]
 test = [
     "pytest>=8.3.2",

lib/marin/src/marin/mcp/__init__.py

@@ -0,0 +1,4 @@
+# Copyright The Marin Authors
+# SPDX-License-Identifier: Apache-2.0
+
+"""MCP servers for Marin operational workflows."""