bug: Safeoutputs MCP sessions expire during long-running agent tasks (~30 min idle timeout)

[lpcox]

Summary

The MCP gateway drops safeoutputs sessions after approximately 30 minutes of inactivity. When an agent runs a long CPU-bound task (e.g., ML training, large builds) inside a shell tool call, no MCP requests are made during that period. The session expires server-side and becomes unrecoverable — all subsequent safeoutputs calls fail with session not found, and the agent has no way to deliver its results.

This is a correctness issue: safeoutputs is the only channel for the agent to report results, and it silently expires during the work the agent was asked to do.

Upstream issue

github/gh-aw#23153 — two independent reports:

• dsyme/fv-squad — 45-minute job, session expired
• githubnext/autoresearch_local — 30-minute training run, session expired

Reproduction timeline (autoresearch_local)

Time (UTC)	Event
22:29:07	safeoutputs MCP server started (v0.2.9), session established
22:29:32	First initialize + tools/list calls succeed
22:50–23:00	Agent runs ML training (~30 min, no MCP calls)
23:00:34	Training completes successfully
23:02:21	First safeoutputs failure: session not found
23:02–23:13	All subsequent calls fail — noop, create_pull_request, push_repo_memory, add_comment, missing_tool
23:13:48	Agent gives up: "safeoutputs MCP session is permanently expired"

Error

All calls return the same error:

✗ noop (MCP: safeoutputs)
  └ MCP server 'safeoutputs': Error: Streamable HTTP error: Error POSTing to endpoint: session not found

✗ create_pull_request (MCP: safeoutputs)
  └ MCP server 'safeoutputs': Error: Streamable HTTP error: Error POSTing to endpoint: session not found

Impact

• Zero safe outputs completed — no PR, no comments, no repo-memory
• Training succeeded (val_bpb 2.236 → 2.107) but results were lost
• Agent wasted ~11 minutes retrying with sleep waits before giving up
• Client-side workarounds (keepalive prompts) don't help because the agent can't send MCP calls while blocked on a long shell execution

Proposed fixes

Any of these would resolve the issue:

1. Remove or significantly extend session timeout for safeoutputs — these sessions should live for the duration of the workflow (up to 6 hours for autoloop). A 30-minute idle timeout is incompatible with long-running tasks.

2. Automatic keepalive from the gateway side — the gateway could ping/refresh sessions internally rather than relying on client activity.

3. Transparent session reconnect — allow the client to re-establish a session when it receives session not found, without requiring manual intervention from the agent.

References

• Upstream issue: MCP gateway drops out in long running jobs: Streamable HTTP error: Error POSTing to endpoint: session not found gh-aw#23153
• Comment with detailed analysis by @insop
• Client-side keepalive attempt (ineffective): https://github.com/githubnext/autoresearch_local/commit/b856479