`flue run` terminates long-running local agents after ~300s of no SSE events

[skylarbpayne]

flue run appears to terminate local agent runs after ~300 seconds if the agent is inside a long-running tool call that does not emit Flue-level SSE events before completing.

This shows up as:

[flue] Agent error: terminated

The tool’s own timeout can be much higher than 300s, but the run still terminates around the 300s mark.

Minimal repro

Create a Flue agent that runs a long shell command:

import type { FlueContext } from '@flue/sdk/client';
import { defineCommand } from '@flue/sdk/node';

export const triggers = {};

const bash = defineCommand('bash');

export default async function (ctx: FlueContext<unknown>) {
  const { init } = ctx;

  const agent = await init({
    sandbox: 'local',
    commands: [bash],
  });

  const session = await agent.session('long-bash-probe');

  return await session.shell(
    "bash -lc 'sleep 360; echo done'",
    {
      timeout: 480,
      commands: [bash],
    },
  );
}

Then run:

flue run long-bash-probe \
  --target node \
  --id long-bash-probe \
  --workspace .flue \
  --output /tmp/flue-long-bash-probe

Observed behavior

A 240s sleep succeeds:

session.shell("bash -lc 'sleep 240; echo done'", {
  timeout: 360,
  commands: [bash],
});

Result:

[flue] Done.

A 360s sleep fails at roughly 301 seconds:

session.shell("bash -lc 'sleep 360; echo done'", {
  timeout: 480,
  commands: [bash],
});

Result:

[flue] Agent error: terminated

My interpretation of what's happening here is that the SSE server used in flue has a default timeout of 300s which isn't directly overrideable, and there isn't any continuous keepalive.

P.S. Great work on flue! I love the abstraction here, and super excited to start driving more usage with it! And I'm happy to contribute a fix to this, but would love to discuss what authors/maintainers the preferred approach :)