Skip to content

fix(passthrough): proxy provider-native error bodies, clean decode errors, cap stop sequences #966

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
SantiagoDePolonia wants to merge 3 commits into
base: main
Choose a base branch
from
Open

fix(passthrough): proxy provider-native error bodies, clean decode errors, cap stop sequences #966

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1075ae2
fix(passthrough): proxy provider-native error bodies, clean decode er…
SantiagoDePolonia Sep 11, 2026
a3d4dcb
fix(responses): attribute nested input decode failures to the input m…
SantiagoDePolonia Sep 11, 2026
36dc6d3
Merge origin/main into fix/passthrough-native-errors
SantiagoDePolonia Sep 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion docs/advanced/anthropic-messages-api.mdx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -236,7 +236,9 @@ features that have no canonical equivalent are not preserved end to end:
`stop_reason: "stop_sequence"` plus the `stop_sequence` value. OpenAI-family
providers conflate stop-parameter hits with natural stops in `finish_reason`,
so completions there report `stop_reason: "end_turn"` (output is still truncated
correctly).
correctly). OpenAI, Azure OpenAI, and Groq accept at most four stop sequences;
a longer list is truncated to the first four (with a warning in the gateway
log) instead of failing the request. Anthropic accepts the full list.
- **`thinking`** on a non-Anthropic model becomes a reasoning effort (budget
under 10000 tokens: `low`, under 20000: `medium`, otherwise `high`; adaptive:
`medium`), sent as each provider's own field: `reasoning_effort` on OpenAI,
Expand Down
5 changes: 4 additions & 1 deletion docs/features/passthrough-api.mdx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,7 +144,10 @@ Passthrough is intentionally narrow while the API is in beta.
that surface.
- GoModel does not translate passthrough request bodies or response bodies.
- Provider-native error bodies and status codes are proxied instead of converted
into OpenAI-compatible responses.
into OpenAI-compatible responses, so provider SDKs raise their own typed
errors. Errors raised by GoModel itself (authentication, rate limits, budgets,
a provider that is not enabled, an unreachable upstream) use the gateway's
error shape, as does an upstream error body that is not JSON.
- Features that depend on OpenAI-compatible request or response shapes may not
apply to passthrough traffic in the same way as `/v1` traffic.

Expand Down
2 changes: 1 addition & 1 deletion internal/core/chat_json.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ func (r *ChatRequest) UnmarshalJSON(data []byte) error {
type alias ChatRequest
var raw alias
if err := json.Unmarshal(data, &raw); err != nil {
return err
return wrapJSONDecodeError(err)
}

extraFields, err := extractUnknownJSONFieldsSet(data, chatRequestFields)
Expand Down
Loading