MCP gateway drops out in long running jobs: Streamable HTTP error: Error POSTing to endpoint: session not found

[dsyme]

@lpcox @Mossaka Do you know if there's some kind of 30-45min timeout for the MCP gateway or firewall or container network connections?

This was from a long running 45min job:

https://github.com/dsyme/fv-squad/actions/runs/23607702532/job/68754992942

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

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

[lpcox]

Filed an issue in gh-aw-mcpg github/gh-aw-mcpg#2596

[lpcox]

github/gh-aw-mcpg#2597

[github-actions]

🏷️ Issue Triaged

Hi @dsyme! I've categorized this issue as bug based on the following analysis:

Reasoning: This describes a runtime failure in existing functionality — the MCP gateway dropping with session not found errors during long-running (45+ min) jobs. This is a reproducible error with a clear stack trace, characteristic of a bug (connection/session timeout).

View Triage Details

Analysis

• Keywords detected: Error POSTing to endpoint, session not found, Streamable HTTP error, timeout, long-running job
• Issue type indicators: Concrete error messages with a reproducible scenario; failure of existing functionality; no new feature is being requested
• Confidence: High

Recommended Next Steps

• Investigate session lifetime configuration in the MCP gateway / safeoutputs server
• Check if there's a configurable keep-alive or session renewal mechanism for long-running jobs
• Consider adding retry logic for session not found errors if sessions can expire mid-job

References: Triage run §23651018257

Note

🔒 Integrity filter blocked 17 items

The following items were blocked because they don't meet the GitHub integrity level.

• gh aw add rewrites {{#import shared/X.md}} with incorrect cross-repo path (resolves from repo root instead of .github/workflows/) #23212 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Feature: allow restricting Copilot CLI built-in tools via frontmatter #23198 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Security: --env-all exposes secrets to prompt-injected agent in pull_request_target workflows #23197 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Update apm-action to v1.4.1 to fix per-dependency authentication #23178 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Harden APM auth: also set GITHUB_APM_PAT when github-app is configured #23148 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• BUG: 'GH_HOST' set with GITHUB_OUTPUT (not env or env-threaded), breaking Process Safe Outputs on GHES #23093 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• BUG: Downstream 'GH_AW_SAFE_OUTPUTS_CONFIG_PATH' and 'GH_AW_SAFE_OUTPUTS_TOOLS_PATH' variables not available as env vars due to GITHUB_OUTPUT-only write #23092 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• AWF allowlist not enforced for plain-HTTP connections to numeric IPs; web-fetch may bypass proxy enforcement #23079 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• BUG: AWF agent cannot reach Postgres started via services: #22939 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• bug: agentic workflows fail with token limit overflow when repository has comprehensive custom instructions #22929 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• gh-aw compiler must reject env.* expressions in markdown per documented safety policy #22914 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Failed to register MCP tools | HTTP 400: Bad Request #22913 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• MCP gateway should enforce tool allowlist at the gateway layer, not only at the Claude client layer #22908 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• cache-memory restore must clear or validate prior-run artifacts before agent launch #22902 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Assign to agent seems to be limited for 1 assignment per issue, I would like to extend it to multiple assignments so agent can work on multiple repos. #22897 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Feature request: force-rerun semantic for workflow_dispatch against the same bound issue #22585 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• ... and 1 more item

To allow these resources, lower min-integrity in your GitHub frontmatter:

tools:
  github:
    min-integrity: approved  # merged | approved | unapproved | none

Generated by Issue Triage Agent · ◷

[insop]

Another occurrence: safeoutputs MCP session expired during autoloop run (2026-04-02)

Failing run: https://github.com/githubnext/autoresearch_local/actions/runs/23924939318/job/69779658230
Repo: githubnext/autoresearch_local

Timeline

Time (UTC)	Event
22:29:07	safeoutputs MCP server started (v0.2.9), session established
22:29:32	MCP gateway routed first initialize + tools/list calls successfully
22:50–23:00	Agent ran training experiments (train.py) — no MCP calls during this ~30 min window
23:00:34	Training completed (105 steps × 4096 tokens = 430K tokens in 451s)
23:02:21	First safeoutputs failure: session not found
23:02:21–23:13:32	All subsequent safeoutputs calls fail — noop, create_pull_request, push_repo_memory, add_comment, missing_tool — every tool type returns session not found
23:13:48	Agent declares: "The safeoutputs MCP session is permanently expired"

Errors (all identical)

✗ noop (MCP: safeoutputs) · message: "Iteration 21 completed: val_bpb=2.107228 …"
  └ MCP server 'safeoutputs': Error: Streamable HTTP error: Error POSTing to endpoint: session not found

✗ missing_tool (MCP: safeoutputs) · reason: "safeoutputs MCP session expired during the 451s training run…"
  └ MCP server 'safeoutputs': Error: Streamable HTTP error: Error POSTing to endpoint: session not found

Impact

• Zero safe outputs completed — no PR, no issue comments, no repo-memory push
• The training itself succeeded (val_bpb improved from 2.236 → 2.107), but all results were lost because the session expired before the agent could report them
• The agent spent ~11 minutes (23:02–23:13) retrying with sleep 60 waits and different tool types, all failing identically

Root cause analysis

The MCP session was idle for roughly 30 minutes while the agent was running CPU-bound training inside the container. No MCP keepalive / heartbeat mechanism exists to maintain the session during long shell operations. Once the session expires on the gateway side, it is unrecoverable — there is no reconnect or session refresh capability.

Position: safeoutputs should never timeout

For autoloop workflows that run long training jobs (minutes to hours), the MCP session must not expire. This is a correctness issue, not just a convenience issue — the entire point of safeoutputs is to be the channel for the agent to report results, and if it silently expires during the work, the agent has no way to deliver its output.

Suggested fixes (any of):

1. No session timeout for safeoutputs sessions (or a very long one, e.g., 4+ hours)
2. Automatic session renewal on the client/gateway side via periodic keepalive pings
3. Session reconnect — allow the agent to re-establish a session transparently when it gets session not found

We already attempted a client-side workaround by adding MCP keepalive guidelines to our agent prompts, but the agent does not issue keepalive calls while blocked on a long shell tool execution, so this doesn't help for the main failure mode.

cc @lpcox @Mossaka

[insop]

@pelikhan

[lpcox]

@insop as noted we extended the timeout period and added keepalives last week. I'll take a closer look.

[pelikhan]

Can we disable timeouts?

[lpcox]

Possibly. There are default values but we can probably just disable them.