[xxwtiancai]

HUMAN:

Problem: Re-prompting after failed context condensation loops indefinitely until hitting MaxIterationsReached.

Fix: Added trailing-event detection to mark runs STUCK on repeated condensation failures while ignoring proactive compaction.

Verification: Verified via scripted repro: the run now safely halts at the 5th rejection instead of spinning for 15 turns.

---

AGENT:

Why

When every agent LLM call is rejected with a context-window error while the condenser keeps succeeding, the conversation loops forever: each rejected call emits a CondensationRequest, the condenser emits a Condensation, and the next call is rejected again. StuckDetector._is_stuck_context_window_error was an unconditional stub (return False), so the cycle only stopped at max_iteration_per_run (default 500), wasting one rejected agent call plus one summarizer call per iteration in exactly the long conversations that are most expensive to restart. Fixes #6104.

Summary

• Implemented scenario 5 in openhands/sdk/conversation/stuck_detector.py: the run is stuck when every event in the trailing window after the last user message belongs to rejected-call/condensation cycles (CondensationRequest/Condensation), with no action, observation, or agent message between them. At least one CondensationRequest must be present, so proactive condensation (token/event pressure, which emits only Condensation) never matches.
• Reused the existing scenario-5 hook, the existing len(events) >= 10 gate in is_stuck(), and the existing STUCK run-loop exit — no new plumbing, no new configuration field, no changes to scenarios 1–4 or the nudge.
• Removed the stale TODO: blocked by agent-sdk#282 stub.

Issue Number

Fixes #6104

How to Test

uv run --frozen pytest tests/cross/test_stuck_detector.py \
  tests/cross/test_stuck_detector_config.py \
  tests/sdk/conversation/local/test_stuck_detector_nudge.py \
  tests/sdk/conversation/local/test_stuck_detector_context_window.py -q

New coverage (all verified to fail without the product change and pass with it):

• tests/cross/test_stuck_detector.py: trailing run of 6 cycles → stuck; 4 cycles (8 events) below the 10-event gate → not stuck; 12 proactive Condensation events with no CondensationRequest → not stuck; cycles interrupted by an ActionEvent → not stuck; user message after the cycles resets the window → not stuck.
• tests/sdk/conversation/local/test_stuck_detector_context_window.py: the issue's scripted repro (agent LLM always raises LLMContextWindowExceedError via TestLLM, summarizer always succeeds) ends STUCK with no ConversationErrorEvent and agent calls strictly below max_iteration_per_run; with stuck_detection=False the same loop still ends ERROR/MaxIterationsReached.

The reproduction is the mocked scripted run the issue and its triage specify (no real LLM involved); the condenser, git-less run loop, and stuck plumbing are the real code paths.

Video/Screenshots

Reproduction notes instead of media (non-visual bug):

• Without the fix, the scripted repro ends error after 15 rejected agent calls + 15 summarizer calls with MaxIterationsReached (matches the issue's reported log).
• With the fix, the same scripted run ends stuck after 5 rejected calls (agent_llm.call_count == 5 < 30) with no ConversationErrorEvent.

Type

• Bug fix
• Feature
• Refactor
• Breaking change
• Docs / chore