Stall/non-convergence on trivial repo_path smoke task exits -2 #12
Description
Summary
During ODIN-136 Phase 0 smoke testing, SWE-AF completed one trivial repo_path run, but a second bounded rerun stalled and never finalized.
The stalled run appears to get stuck in coder-side acceptance-criteria debugging loops and exits with -2 without producing a final successful completion for the issue.
Environment
- SWE-AF base:
f63b9e744ad922c0a914201020a0e9df9d4264fb - Stabilization branch tested:
fix/claude-sdk-rate-limit-event-compat(includes SDK protocol mismatch fail-fast) - AgentField:
v0.1.41 - Mode:
repo_path
Reproduction (high-level)
- Run trivial smoke workflow against a local repo requiring a small README edit and AC verification commands.
- Observe planner/coder execution begin normally.
- Coder enters acceptance-criteria checks and repeated debugging behavior.
- Run does not converge to completion; eventually aborts/fails.
Evidence snippets
From coder_append-smoke-check-2-section_iter_1.jsonl:
- Start prompt includes ACs and explicit focus on AC fix: isolate concurrent builds via build_id namespace on git resources #2 awk sensitivity.
- Assistant reports it is running AC checks.
- Terminal event:
event: endis_error: trueerror: Command failed with exit code -2 (exit code: -2)
From checkpoint.json at failure time:
completed_issues: []in_flight_issues: ["append-smoke-check-2-section"]- integration branch exists but issue not finalized.
Expected behavior
For trivial bounded smoke tasks, execution should either:
- converge and finalize reliably, or
- fail fast with actionable failure classification and cleanly terminate issue state.
Actual behavior
Execution remained in a non-convergent loop and exited with -2 while the issue remained in-flight.
Request
Can maintainers advise whether this is a known orchestration failure mode and recommend guardrails/fixes (e.g., bounded retry/iteration controls or better AC-debug loop exit conditions)?