Reproduction: model alias mismatch with OpenCode Go (related to #526)

[sathwick-p]

Summary

This is a concrete reproduction of the issue described in #526 — model ID mismatch when using a custom OpenAI-compatible upstream (OpenCode Go) via OPENROUTER_BASE_URL.

The router sends slash-form model IDs (deepseek/deepseek-v4-flash) in the request body, but OpenCode Go expects bare names (deepseek-v4-flash). There is no env-var mechanism to remap model names at runtime.

Steps to reproduce

1. Start the router

docker compose up -d

2. Point OpenRouter at OpenCode Go

Add to .env (docker compose auto-reads this):

OPENROUTER_API_KEY=sk-your-opencode-go-key
OPENROUTER_BASE_URL=https://opencode.ai/zen/go/v1

Then recreate:

docker compose up -d --force-recreate router-server

3. Create a router API key

Login at http://localhost:8080/ui/login (password: admin), then create a key via Settings → API Keys.

4. Send a test request

curl -s -N "http://localhost:8080/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -H "x-api-key: rk_YOUR_KEY" \
  -d '{"model":"deepseek/deepseek-v4-flash","messages":[{"role":"user","content":"say hello in one word"}]}'

5. Actual result

{"type":"error","error":{"type":"ModelError","message":"Model deepseek/deepseek-v4-pro is not supported"}}

The error format ({"type":"error",...}) confirms the request reached OpenCode Go. It was rejected because the model name sent was the router's public slash-form name instead of OpenCode Go's bare name.

Root cause analysis

The OpenRouter client (openaicompat.Client) sends the body's "model" field verbatim — no modelIDMap rewriting happens for OpenRouter:

Provider	Has modelIDMap?	Constructor in main.go
OpenRouter	No (public slug = upstream ID)	NewClient(key, url)
DeepInfra	Yes	NewClientWithModelIDMap(key, url, upstreamIDsForProvider("deepinfra"))
Fireworks	Yes	NewClientWithModelIDMap(key, url, upstreamIDsForProvider("fireworks"))
Bedrock	Yes	NewClientWithModelIDMap(key, url, upstreamIDsForProvider("bedrock"))

The modelIDMap rewrites the body's "model" field at request time in rewriteModelField() (internal/providers/openaicompat/client.go:119). For OpenRouter this is a no-op because the map is nil.

OpenCode Go's /v1/models returns bare names (deepseek-v4-flash, deepseek-v4-pro, kimi-k2.6, etc.) but the router catalog uses slash-form IDs (deepseek/deepseek-v4-flash, moonshotai/kimi-k2.6, z-ai/glm-5.2, etc.).

The existing rewriteModelField infrastructure is correct — it just needs an env-var hook so self-hosters can supply runtime mappings.

Proposed fix direction

An env var (e.g. ROUTER_MODEL_ID_MAP) — comma-separated from=to pairs — parsed at boot and merged into the OpenRouter client's modelIDMap. Example:

ROUTER_MODEL_ID_MAP="deepseek/deepseek-v4-flash=deepseek-v4-flash,deepseek/deepseek-v4-pro=deepseek-v4-pro,moonshotai/kimi-k2.6=kimi-k2.6,xiaomi/mimo-v2.5-pro=mimo-v2.5-pro,z-ai/glm-5.2=glm-5.2"

Environment

• Router: main branch, docker compose stack
• Upstream: OpenCode Go (https://opencode.ai/zen/go/v1)

[vignesh-chaturvedi]

Opened #900, which closes this along with #526.

Your reproduction is what the regression test is built from — an httptest stub stands in for the gateway and the assertion is the model field it actually receives, so the three cases are: no alias configured gives deepseek/deepseek-v4-flash (the rejection you hit), the alias set gives deepseek-v4-flash, and an unrelated model goes through untouched.

Your root-cause table was right that the gap is OpenRouter having no modelIDMap. Worth adding that #892 has since landed per-key model_aliases, but it's BYOK-credential-only (internal/proxy/credentials.go:50), so the OPENROUTER_API_KEY + OPENROUTER_BASE_URL path you used still had nothing. The env var fills that layer, and a BYOK key's aliases still win over it.

If you still have the OpenCode Go setup handy, a confirmation against the real endpoint would be more convincing than my stub.