Skip to content

Reasoner timeout defaults (2h) are shorter than a normal feature build (observed 7h27m) #176

Open

Description

@Jack11111eee

Summary

swe-planner.plan and swe-planner.execute each run an entire build phase inside one reasoner call. AgentField enforces a wall-clock watchdog per reasoner using AsyncConfig.default_execution_timeout, which defaults to 7200 s, and SWE-AF does not override it. A healthy feature build routinely needs longer than that, so the watchdog kills builds that are still making progress.

Separately, agent_timeout_seconds defaults to 2700 s (45 min), which is shorter than observed single-agent runtimes.

Steps to Reproduce

  1. Run swe-planner.build for a feature that decomposes into roughly 30–40 issues.
  2. Leave the node at default config (agentfield>=0.1.130, SWE-AF main).

Expected Behavior

A build that keeps producing commits and notes completes, or fails on a real error and can be resumed.

Actual Behavior

From the control plane's workflow_executions table for one 35-issue build:

Execution Result
execute failed: Reasoner 'execute' timed out after 7200.0s of active time (twice)
execute (after raising the budget via env) failed at 21600s once
execute (with a 12h budget) succeeded in 26845.7 s (7h27m)
run_qa succeeded, max observed 3087 s (51 min)
run_coder succeeded, max observed 31416 s; 4 executions hit execution timed out (no activity)

So the SDK's 2-hour default is below the normal duration of the pipeline it is wrapping, and the 45-minute per-agent cap cut healthy QA/coder calls mid-edit:

QA agent failed: gate-c: Agent call 'qa:gate-c:iter1' timed out after 2700s
Coder agent failed: gate-d iter 1: Agent call 'coder:gate-d:iter1' timed out after 2700s

Environment

  • OS: macOS (Apple Silicon), 8 GB RAM
  • Python: 3.13
  • AI provider: claude_code runtime
  • Deployment: local control plane + node
  • SWE-AF: main @ da78a77, agentfield 0.1.145

Logs or Trace

exec_..._4b1mtd05  execute  failed  7200.1s  Reasoner 'execute' timed out after 7200.0s of active time
exec_..._m1ed98cv  execute  failed  21602.6s Reasoner 'execute' timed out after 21600.0s of active time
exec_..._d6tfdqtg  execute  succeeded 26845.7s

A node-level AsyncConfig subclass with a 12h budget (overridable through the existing AGENTFIELD_ASYNC_MAX_EXECUTION_TIMEOUT / AGENTFIELD_ASYNC_DEFAULT_EXECUTION_TIMEOUT variables), plus a 60-minute agent_timeout_seconds, would fit the pipeline's real runtime while keeping both watchdogs as last-resort bounds.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions