Season 3 stage: "gate unreachable, failing closed: timed out" on every run

#3
by azyg0s - opened

Hi, on the Season 3 stage (inbox/new_mail.md, fixed request) every run ends with the agent's own requested read_file being denied:

kernel DENY the requested read β€” reason: gate unreachable, failing closed: timed out

It happened on three runs today, 2026-11-18, around 11:44–12:05 UTC, and the planted content was the guide's example mail. The Agent Guard app on the desktop also shows "Backend offline: recorded". Is the kernel gate down for the stage, or is it something on my side? Thanks!

Same here. After a hard reload and a fresh session (syfeg5whi), a benign read_file(inbox/new_mail.md) still failed with gate unreachable, failing closed: timed out (no audit). The desktop also switched to Backend offline: recorded.

FINAL_Bench org

Thanks for the detailed report, and for the exact times and the audit reference. That was a real issue on our side, not something on yours.

What happened: the policy gate on the backend stopped responding for a period, so the kernel could not reach it within the time limit. The kernel is built to fail closed, which means that when it cannot confirm an action is safe in time, it denies the action rather than allowing it. That is why even the read_file your own run requested was denied with "gate unreachable, failing closed: timed out", and why the desktop app showed "Backend offline: recorded". The deny-by-default behaviour is intended, but a legitimate read being blocked is not, and the cause was the gate being unreachable.

Status now: we restarted the gate backend and it is responding normally again (verified just now, well under one second). Please start a fresh Season 3 session and run the planted-mail example again. It should produce verdicts as usual.

If it happens again, reply here with the audit number and the UTC time and we will look right away.


μ‹ κ³  κ°μ‚¬ν•©λ‹ˆλ‹€. μ •ν™•ν•œ μ‹œκ°κ³Ό 감사 λ²ˆν˜Έκ°€ 큰 도움이 λμŠ΅λ‹ˆλ‹€. 이 건은 μ°Έκ°€μžλΆ„ ν™˜κ²½μ΄ μ•„λ‹ˆλΌ 저희 μͺ½ λ¬Έμ œμ˜€μŠ΅λ‹ˆλ‹€.

원인: λ°±μ—”λ“œμ˜ μ •μ±… κ²Œμ΄νŠΈκ°€ ν•œλ™μ•ˆ 응닡을 멈좰, 컀널이 μ œν•œ μ‹œκ°„ μ•ˆμ— κ²Œμ΄νŠΈμ— λ‹Ώμ§€ λͺ»ν–ˆμŠ΅λ‹ˆλ‹€. 컀널은 μ•ˆμ „μ„ μ œλ•Œ ν™•μΈν•˜μ§€ λͺ»ν•˜λ©΄ ν—ˆμš©μ΄ μ•„λ‹ˆλΌ κ±°λΆ€λ‘œ 떨어지도둝(fail closed) 섀계돼 μžˆμ–΄, μ°Έκ°€μžλΆ„μ΄ μš”μ²­ν•œ read_file κΉŒμ§€ κ±°λΆ€λ˜κ³  λ°μŠ€ν¬ν†±μ— "Backend offline" 이 λ–΄μŠ΅λ‹ˆλ‹€. κ±°λΆ€ κΈ°λ³Έκ°’ μžμ²΄λŠ” μ˜λ„λœ λ™μž‘μ΄μ§€λ§Œ, 정상 읽기가 λ§‰νžŒ 것은 κ²Œμ΄νŠΈκ°€ λ‹Ώμ§€ μ•Šμ•˜κΈ° λ•Œλ¬Έμž…λ‹ˆλ‹€.

ν˜„μž¬: 게이트 λ°±μ—”λ“œλ₯Ό μž¬μ‹œμž‘ν–ˆκ³  μ§€κΈˆμ€ 정상(1초 이내 응닡) ν™•μΈν–ˆμŠ΅λ‹ˆλ‹€. μ‹œμ¦Œ 3 μ„Έμ…˜μ„ μƒˆλ‘œ μ‹œμž‘ν•΄ λ‹€μ‹œ μ‹œλ„ν•΄ μ£Όμ„Έμš”. μž¬λ°œν•˜λ©΄ 감사 λ²ˆν˜Έμ™€ UTC μ‹œκ°μ„ 적어 μ£Όμ‹œλ©΄ λ°”λ‘œ ν™•μΈν•˜κ² μŠ΅λ‹ˆλ‹€.

μ•ˆλ…•ν•˜μ„Έμš”. 이전 μ‹œμ¦Œ 3 μ •μ±… 게이트 νƒ€μž„μ•„μ›ƒμ€ λ³΅κ΅¬λμ§€λ§Œ, μ§€κΈˆμ€ λͺ¨λΈ λ‹¨κ³„μ—μ„œ λ‹€λ₯Έ μž₯μ• κ°€ λ³΄μž…λ‹ˆλ‹€.

  • 2026-11-19 15:21:07 UTC, task=mail, μ„Έμ…˜ s3-singleton-path-0020
  • 2026-11-19 15:22:26 UTC, task=mail, μ„Έμ…˜ s3-longaux-0025

두 /api/ag/s3/run μš”μ²­ λͺ¨λ‘ λ°”κΉ₯ HTTP 200 μ•ˆμ˜ μ΄λ²€νŠΈκ°€ μ‚¬μš©μž μš”μ²­ 직후 model error: HTTP Error 503: Service Unavailable둜 λλ‚¬μŠ΅λ‹ˆλ‹€. 첫 read_file μ œμ•ˆλ„ μ—†μ–΄ 감사 λ²ˆν˜ΈλŠ” μ—†μŠ΅λ‹ˆλ‹€. 15:39 UTC의 /api/ag/healthλŠ” ok:true, backend:openmodel, live:falseμ˜€μŠ΅λ‹ˆλ‹€. λͺ¨λΈ λ°±μ—”λ“œλ₯Ό ν™•μΈν•˜κ³  볡ꡬ해 μ£Όμ‹€ 수 μžˆμ„κΉŒμš”? κ°μ‚¬ν•©λ‹ˆλ‹€.

FINAL_Bench org

@taeng0204 μ‹ κ³  κ°μ‚¬ν•©λ‹ˆλ‹€. 15:21 UTC 와 15:22 UTC 의 model error: HTTP Error 503 은 κ·Έ μ‹œκ° λͺ¨λΈ μ„œλ²„κ°€ 멈좰 μžˆμ—ˆκΈ° λ•Œλ¬Έμž…λ‹ˆλ‹€. λͺ¨λΈ μ„œλ²„λŠ” λ‹€μ‹œ κ°€λ™ν–ˆκ³ , 같은 경둜(/api/ag/chat)둜 μš”μ²­μ„ 보내 λͺ¨λΈμ΄ 도ꡬ ν˜ΈμΆœμ„ μ •μƒμœΌλ‘œ μ œμ•ˆν•˜λŠ” κ²ƒκΉŒμ§€ ν™•μΈν–ˆμŠ΅λ‹ˆλ‹€. 같은 μš”μ²­μ„ λ‹€μ‹œ 보내 μ£Όμ‹œλ©΄ λ©λ‹ˆλ‹€.

@taeng0204 thank you for the report. The model error: HTTP Error 503 at 15:21 and 15:22 UTC happened because the model server was stopped at that time. It is running again, and we confirmed through the same path (/api/ag/chat) that the model proposes tool calls normally. Please send the same requests again.

μ•ˆλ…•ν•˜μ„Έμš”. λͺ¨λΈ λ‹¨κ³„μ˜ 503이 였늘 λ‹€μ‹œ λ°œμƒν•©λ‹ˆλ‹€. 2026-11-20 07:02:46–07:12:55 UTC에 /api/ag/chat의 μ„œλ‘œ λ‹€λ₯Έ SID 8건이 μ™ΈλΆ€ HTTP 200 응닡 μ•ˆμ—μ„œ model error: HTTP Error 503: Service Unavailable둜 첫 도ꡬ μ œμ•ˆ 전에 μ’…λ£ŒλμŠ΅λ‹ˆλ‹€. 예: s2lexn4bp2b2r0(07:02:46 UTC), s2lexbefore7f220261120a(07:12:55 UTC). 이 νšŒμ°¨λ“€μ€ 감사 λ²ˆν˜Έκ°€ μ—†μŠ΅λ‹ˆλ‹€. λͺ¨λΈ μ„œλ²„ μƒνƒœ 확인 λΆ€νƒλ“œλ¦½λ‹ˆλ‹€.

@taeng0204 μ‹ κ³  κ°μ‚¬ν•©λ‹ˆλ‹€. 2026-11-20 07:02~07:13 UTC 의 model error: HTTP Error 503 은 κ·Έ μ‹œκ° λͺ¨λΈ μ„œλ²„κ°€ 멈좰 μžˆμ—ˆκΈ° λ•Œλ¬Έμž…λ‹ˆλ‹€.

07:21 UTC 에 λͺ¨λΈ μ„œλ²„λ₯Ό λ‹€μ‹œ κ°€λ™ν–ˆμŠ΅λ‹ˆλ‹€. 같은 경둜(/api/ag/chat)와 μ‹œμ¦Œ 3 λ¬΄λŒ€(/api/ag/s3/run, 과제 mail)둜 μš”μ²­μ„ 보내 λͺ¨λΈμ΄ 도ꡬ ν˜ΈμΆœμ„ μ œμ•ˆν•˜κ³  μ‹€ν–‰κΉŒμ§€ λλ‚˜λŠ” 것을 ν™•μΈν–ˆμŠ΅λ‹ˆλ‹€. 07:30 UTC 에 μ›λž˜ μ‚¬μ–‘μœΌλ‘œ 되돌렸고, μ§€κΈˆμ€ ν•œ 번 싀행에 20초쯀 κ±Έλ¦½λ‹ˆλ‹€. 멈좰 있던 λ™μ•ˆμ˜ νšŒμ°¨λŠ” 감사 λ²ˆν˜Έκ°€ μ—†μœΌλ―€λ‘œ 같은 μš”μ²­μ„ λ‹€μ‹œ 보내 μ£Όμ‹œλ©΄ λ©λ‹ˆλ‹€.

@taeng0204 thank you for the report. The model error: HTTP Error 503 between 07:02 and 07:13 UTC on 2026-11-20 happened because the model server was stopped at that time.

The model server was started again at 07:21 UTC. We sent requests through the same path (/api/ag/chat) and through the Season 3 stage (/api/ag/s3/run, task mail) and confirmed that the model proposes a tool call and the run completes. At 07:30 UTC it was put back on its original hardware, and a run now takes about 20 seconds. The runs made while it was stopped have no audit numbers, so please send the same requests again.

Sign up or log in to comment