MCP tool calls hard-cap at 60s; warehouse queries abort

[regutierrez]

Summary

Executor MCP tool invocations abort at 60 seconds. There is no env, compose, or toolkit knob to raise that ceiling. Typical warehouse/db queries (Snowflake run_snowflake_query, Postgres via dbhub) often need longer and fail as transportFailure.

Observed error from the MCP SDK path (v1.6.8):

Error: MCP tool failed: The operation was aborted due to timeout

Live example: three parallel snowflake.org.blushift.run_snowflake_query calls through /mcp/toolkits/akkio died at ~60.3s and ~61.9s.

Where the 60s lives

v1.6.8 (ghcr.io/usefulsoftwareco/executor-selfhost:1.6.8, latest published tag 2026-09-05):

• MCP plugin does connection.client.callTool({ name, arguments }) with no timeout
• @modelcontextprotocol/sdk then applies DEFAULT_REQUEST_TIMEOUT_MSEC = 60000

main after UsefulSoftwareCo#1956 (unreleased):

• SDK timer is set to MCP_SDK_TIMEOUT_BACKSTOP_MS = 2_147_483_647
• real ceiling is still hardcoded MCP_ACTIVE_WORK_TIMEOUT_MS = 60_000 in packages/plugins/mcp/src/sdk/invoke.ts
• that timer pauses only for elicitation, not for a long SQL statement

Abort message on main would be: MCP tool invocation exceeded its active-work deadline. Same 60s wall.

What does not control this

Knob	What it actually bounds
EXECUTOR_SANDBOX_TIMEOUT_MS	QuickJS sandbox (default 5 min). Inner MCP still dies at 60s.
GraphQL plugin timeout	110s in the same binary; MCP is separate
dbhub query_timeout	tighter Postgres ceiling (e.g. 30s), not the MCP client
Pi / Flue / Slack client timeouts	outer HTTP; Executor already aborted the inner tool

What we need

A caller- or host-configurable MCP invoke timeout, e.g.:

• InvokeOptions.timeoutMs / maxTotalTimeoutMs / resetTimeoutOnProgress (mcp: expose request options through InvokeOptions UsefulSoftwareCo/executor#1988 — open, unreviewed, merge-dirty against current main)
• and/or a self-host env such as EXECUTOR_MCP_TOOL_TIMEOUT_MS so Code Mode / execute does not have to pass options on every warehouse call

Default 60s can stay. Warehouse toolkits need a way to wait minutes, not one minute.

Workaround until then

Small bounded reads (LIMIT, short window, one query at a time). Do not raise EXECUTOR_SANDBOX_TIMEOUT_MS expecting it to help.

Related

• mcp: expose request options through InvokeOptions UsefulSoftwareCo/executor#1988 — expose MCP request options through InvokeOptions
• fix(mcp): pause active timeout during elicitation UsefulSoftwareCo/executor#1956 — pause active-work deadline during elicitation (does not raise the 60s cap)