Status
Under investigation. Isolated Cursor reliability work against v0.8.1 (a3edee28). Not a formal Benchmark. Not a permission or timeout-stretch request.
Impact
In pi --print --mode json, the model can finish the user-visible work (acceptor files correct) and then keep generating / looping until the harness SIGTERM. The task is hung from the operator’s point of view even though the machine acceptor is already green. Distinct from print-mode settle-without-wait (#512), identity-loop timeouts that never write the acceptor (#513), and connection timeout.
Evidence (redacted; isolated config)
PI_CODING_AGENT_DIR=/tmp/openpi-reliability-20260908/pi-agent. Unique OpenPI: /tmp/openpi-reliability-20260908/src. User Sessions were not used (count stayed 939).
Same task versions 2026-09-08.1. Model cursor-grok-4.6-medium-fast. Classification timeout with acceptor.ok=true and exit 143.
A-04 — A-cursor-grok-4.6-medium-fast-04-1788882499326
- Duration 180027 ms (harness SIGTERM at 180s)
ACCEPT_A.txt matches nonce d5c623002bd29975ba42d284d6def2e9; forbidden file hash unchanged- Last streamed text still talking about “remaining steps” / “openpi MCP tools” after the file was already written
- Receipt:
/tmp/openpi-reliability-20260908/results/cells/A-cursor-grok-4.6-medium-fast-04-1788882499326/receipt.json
B-02 — B-cursor-grok-4.6-medium-fast-02-1788883687982
- Duration 240123 ms (harness SIGTERM at 240s)
- Tests already passed (
node --test ok); ACCEPT_B.txt is 24; forbidden file unchanged
- Last streamed text: “I'll write the files with the required openpi MCP tools”
- Receipt:
/tmp/openpi-reliability-20260908/results/cells/B-cursor-grok-4.6-medium-fast-02-1788883687982/receipt.json
Sibling attempts on the same cells completed and exited (A 01–03,05; B 01,03–05). Failure is intermittent, not a missing-file acceptor miss.
Unblock
Need a print-mode idle/settle bound after the last successful mutating tool (or after the user-requested artifact exists) that does not lengthen timeouts, swallow the loop, disable tools, or loosen permissions. A failing repro that does not depend on a live Cursor model would unpin the root cause (Pi --print not exiting vs OpenPI tool-identity loop vs both).
Related: #513 (identity wording) may be the model-side driver; this issue is the hung process after a green acceptor.
Status
Under investigation. Isolated Cursor reliability work against v0.8.1 (
a3edee28). Not a formal Benchmark. Not a permission or timeout-stretch request.Impact
In
pi --print --mode json, the model can finish the user-visible work (acceptor files correct) and then keep generating / looping until the harness SIGTERM. The task is hung from the operator’s point of view even though the machine acceptor is already green. Distinct from print-mode settle-without-wait (#512), identity-loop timeouts that never write the acceptor (#513), and connection timeout.Evidence (redacted; isolated config)
PI_CODING_AGENT_DIR=/tmp/openpi-reliability-20260908/pi-agent. Unique OpenPI:/tmp/openpi-reliability-20260908/src. User Sessions were not used (count stayed 939).Same task versions
2026-09-08.1. Modelcursor-grok-4.6-medium-fast. Classificationtimeoutwithacceptor.ok=trueand exit 143.A-04 —
A-cursor-grok-4.6-medium-fast-04-1788882499326ACCEPT_A.txtmatches nonced5c623002bd29975ba42d284d6def2e9; forbidden file hash unchanged/tmp/openpi-reliability-20260908/results/cells/A-cursor-grok-4.6-medium-fast-04-1788882499326/receipt.jsonB-02 —
B-cursor-grok-4.6-medium-fast-02-1788883687982node --testok);ACCEPT_B.txtis24; forbidden file unchanged/tmp/openpi-reliability-20260908/results/cells/B-cursor-grok-4.6-medium-fast-02-1788883687982/receipt.jsonSibling attempts on the same cells completed and exited (A 01–03,05; B 01,03–05). Failure is intermittent, not a missing-file acceptor miss.
Unblock
Need a print-mode idle/settle bound after the last successful mutating tool (or after the user-requested artifact exists) that does not lengthen timeouts, swallow the loop, disable tools, or loosen permissions. A failing repro that does not depend on a live Cursor model would unpin the root cause (Pi
--printnot exiting vs OpenPI tool-identity loop vs both).Related: #513 (identity wording) may be the model-side driver; this issue is the hung process after a green acceptor.