You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Error: 403 Authentication failed for Muse Spark 1.3 Contributor #945
Controlled reproduction on 2026-09-29 (Provider API, meta/muse-spark-1.3-contributor):
The same unmodified Command Code API key returns 200 for gpt-5.6-sol from a Windows laptop, and 200 for Muse from an Azure host.
From the laptop, the documented /provider/v1/chat/completions endpoint intermittently returns 403 permission_error / Authentication failed. Please check your credentials. for Muse. One captured failure: x-trace-id: bcf923313633b31501ecc91270739d17, cf-ray: a42a5dce5f7e6131-HKG.
Cloudflare's client-side trace reports loc=KR for the laptop. Calls routed through NRT/KIX succeeded, while HKG calls failed; a DNS/IP pin did not reliably select a good edge.
The official Command Code CLI 1.66.0, using that same key on the same laptop with --no-session --model meta/muse-spark-1.3-contributor --print, returned OK immediately after the captured Provider API 403.
The API failures can persist across multiple bounded retries, so a one-off retry is insufficient. The response contains no actionable error code distinguishing an edge/routing failure from a region or account restriction.
Could you inspect this trace and confirm the actual rejection reason, then fix the HKG Provider API route or its error mapping? No key or prompt data is included here.
Additional reproduction from Korea on 2026-10-16 (Provider API, GOAT plan). The edge correlation lee5312 described also hits Claude Haiku 5.5 and Claude Sonnet 5.5, not only Muse Spark.
Setup: same unmodified API key, POST /provider/v1/messages (Anthropic Messages), max_tokens: 16, client loc=KR. Both Claude models are listed as included on the GOAT plan.
Result (01:54-01:57 UTC, roughly one request every 2-4s):
Model
Edge (cf-ray suffix)
Result
claude-haiku-5-5
HKG
403 x16
claude-sonnet-5-5
HKG
403 x16
claude-haiku-5-5
NRT
200 x6
claude-sonnet-5-5
NRT
200 x5
Every HKG response was a failure and every NRT response was a success, with nothing else changed. The failure body is the credential-shaped error:
403 {"type":"error","error":{"type":"permission_error","message":"Authentication failed. Please check your credentials."}}
Controls in the same run:claude-opus-5-5 returns the correct MODEL_NOT_IN_PLAN (trace de0050fbfd358fbd2e7af0415ed410a5), and gpt-6-luna and deepseek/deepseek-v4.1-flash return 200. So the key and plan are valid.
Impact: third-party agents that use the Provider API retry the 403 and then stop with a generic provider error. From Korea, these models fail whenever the request lands on HKG.
Requests:
Fix the HKG Provider API route for Anthropic models (and Muse).
If this is an intended restriction, return a distinct error code instead of "Authentication failed".
Activity
iAziz786 commentedon Sep 28, 2026
Another user reported this from Discord:
lee5312 commentedon Sep 29, 2026
Controlled reproduction on 2026-09-29 (Provider API,
meta/muse-spark-1.3-contributor):gpt-5.6-solfrom a Windows laptop, and 200 for Muse from an Azure host./provider/v1/chat/completionsendpoint intermittently returns 403permission_error/Authentication failed. Please check your credentials.for Muse. One captured failure:x-trace-id: bcf923313633b31501ecc91270739d17,cf-ray: a42a5dce5f7e6131-HKG.loc=KRfor the laptop. Calls routed through NRT/KIX succeeded, while HKG calls failed; a DNS/IP pin did not reliably select a good edge.--no-session --model meta/muse-spark-1.3-contributor --print, returnedOKimmediately after the captured Provider API 403.Could you inspect this trace and confirm the actual rejection reason, then fix the HKG Provider API route or its error mapping? No key or prompt data is included here.
nerin81-netizen commentedon Oct 16, 2026
Additional reproduction from Korea on 2026-10-16 (Provider API, GOAT plan). The edge correlation lee5312 described also hits Claude Haiku 5.5 and Claude Sonnet 5.5, not only Muse Spark.
Setup: same unmodified API key,
POST /provider/v1/messages(Anthropic Messages),max_tokens: 16, clientloc=KR. Both Claude models are listed as included on the GOAT plan.Result (01:54-01:57 UTC, roughly one request every 2-4s):
Every HKG response was a failure and every NRT response was a success, with nothing else changed. The failure body is the credential-shaped error:
Captured failures (HKG):
x-trace-id: 19a0059f0f5c8dfcc955b40afff00f69,cf-ray: a4b371219fed5dec-HKG(01:55:33 UTC)x-trace-id: 56bed892ff189ffbebfde12d3257c840,cf-ray: a4b371241a135dec-HKG(01:55:33 UTC)x-trace-id: c2e1ea9527f679a8a11dbc82fcaf5fa9,cf-ray: a4b3713eaad309f4-HKG(01:55:37 UTC)Successes (NRT, same key, about 40s earlier):
x-trace-id: 23cbb0b00203b85d4ef5b58fa4c2d70e,cf-ray: a4b3702a4f053018-NRT(01:54:54 UTC)x-trace-id: 14e0d4306e0a9e72fd63c7721c10df59,cf-ray: a4b3703c7c813018-NRT(01:54:57 UTC)x-trace-id: 6a89f2f4524239212e0f1f9ebc66643b,cf-ray: a4b3704b5f263018-NRT(01:54:59 UTC)Controls in the same run:
claude-opus-5-5returns the correctMODEL_NOT_IN_PLAN(tracede0050fbfd358fbd2e7af0415ed410a5), andgpt-6-lunaanddeepseek/deepseek-v4.1-flashreturn 200. So the key and plan are valid.Impact: third-party agents that use the Provider API retry the 403 and then stop with a generic provider error. From Korea, these models fail whenever the request lands on HKG.
Requests:
No key or prompt data is included.