Muse Spark on /chat/completions returns misleading 500/403-country instead of pointing to /responses endpoint

Summary

When calling muse-spark-1.2/1.3-contributor on POST /zen/go/v1/chat/completions, the gateway returns a bare {"type":"error","error":{"type":"error","message":"Internal server error"}} (HTTP 500) with no hint that these models are served only via /zen/go/v1/responses. The same models work fine on /responses with identical auth/headers. This cost multiple hours of debugging across several users (#47949, #43692, #39872, #41518) — a WrongEndpointError (or an automatic redirect) would eliminate the whole class of reports.

Related: when the same /chat/completions call is made without x-opencode-session, the gateway sometimes returns [403]: This model is not available in your country. instead of the (correct) 400 MissingSessionID. The 403 RegionError is also produced for clients that hit /chat/completions from otherwise-allowed regions — i.e. the region check appears to be evaluated before endpoint/session routing, and its error message is misleading when the actual problem is the endpoint.

Environment

• Endpoint: https://opencode.ai/zen/go/v1 (OpenCode Go subscription, active)
• Client: OmniRoute proxy + direct curl reproduction (no client-specific behavior involved)
• Date: 2026-09-08, multiple hours of testing
• Auth: sk-... Bearer key

Reproduction (single API key, same egress IP, same minute)

1. /chat/completions — fails 500 regardless of headers:

curl -s -X POST "https://opencode.ai/zen/go/v1/chat/completions" \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -H "User-Agent: opencode" \
  -H "x-opencode-client: desktop" -H "x-opencode-project: global" \
  -H "x-opencode-session: $UUID" -H "x-opencode-request: $UUID" \
  -d '{"model":"muse-spark-1.3-contributor","messages":[{"role":"user","content":"say ok"}],"max_tokens":512}'
# → {"type":"error","error":{"type":"error","message":"Internal server error"}}  (HTTP 500)

Same result with muse-spark-1.2-contributor, with/without session headers, stream/non-stream.

2. /responses — works with the same key/IP/headers:

curl -s -X POST "https://opencode.ai/zen/go/v1/responses" \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -H "User-Agent: opencode" \
  -H "x-opencode-client: desktop" -H "x-opencode-project: global" \
  -H "x-opencode-session: $UUID" -H "x-opencode-request: $UUID" \
  -d '{"model":"muse-spark-1.3-contributor","input":[{"role":"user","content":"say ok"}],"max_output_tokens":512,"stream":false}'
# → HTTP 200, normal response object

3. /responses without session header — correct error, good message:

{"type":"error","error":{"type":"MissingSessionID","message":"Request is missing x-opencode-session..."}}

This proves the gateway can produce precise, actionable errors — it just doesn't for the wrong-endpoint case.

4. Misleading 403 region message:

/chat/completions calls for muse models from a supported-region egress sometimes return:

{"error":{"message":"[403]: This model is not available in your country."}}

while the identical request via /responses succeeds. When the actual cause is the endpoint (and/or missing session header), surfacing a country error sends users chasing proxies/DNS for hours — see #47949 where 12 exits across 7 countries were tested, and #43692.

Expected behavior

Any of the following would fix the whole class of confusion:

1. Best: route muse-family models to /responses automatically when they are requested on /chat/completions (the model registry already knows the required endpoint), or
2. Return a distinct, actionable error, e.g. {"type":"WrongEndpointError","message":"muse-spark-* is only served via /zen/go/v1/responses"} (HTTP 400/404), or
3. At minimum, document in docs/go (#Endpoints) that muse models return 500/403 country when called on /chat/completions, so the error text can be trusted.

Also worth noting: RegionError firing for /chat/completions muse calls from allowed regions (while /responses works) suggests the region gate runs before endpoint routing — moving it after endpoint resolution (or tagging its error with the resolved endpoint) would make the 403 text trustworthy.

Observed error surface summary

Call	Result
muse on /chat/completions, full CLI headers + session	500 Internal server error
muse on /chat/completions, no session	403 not available in your country (misleading) or 500
muse on /responses, no session	400 MissingSessionID (good error)
muse on /responses, with session	200 (or honest rate_limit_exceeded when quota hit)
non-muse (glm-5, hy3, kimi-k3...) on /chat/completions	200 — endpoint is correct for them

All checks performed on 2026-09-08 with a paid Go subscription; happy to provide request IDs if needed.