Description

Requests to OpenCode Go models qwen3.8-flash / qwen3.8-max on https://opencode.ai/inference/go/anthropic/v1/messages that are rejected by the upstream backend (e.g. thinking.budget_tokens >= max_tokens) return HTTP 400 with the SSE body event: error\ndata: {} — the real upstream error message is dropped. opencode then shows the useless "Provider request failed with HTTP 400: event: error data: {}" after its retries.

The same class of validation error is NOT masked on other Go paths: claude-haiku-5-5 returns a readable invalid_request_error message, and a request without x-opencode-session returns a readable MissingSessionID JSON. So the masking appears specific to the Qwen 3.8 backend route.

Control requests on the exact same key/plan (thinking with max_tokens > budget_tokens, no-thinking, tool-heavy payloads: $defs/$ref, enum/pattern, 40 tools, long system, cache_control) succeed 100% — so this is not a limits/auth/client issue and is reproducible with plain curl.

Expected: the Go gateway propagates the upstream provider's error payload for all backends, so clients surface the actual reason instead of data: {}.

User-visible symptom in opencode: frequent opaque failures on opencode-go/qwen3.8-flash with a thinking variant (variants none/low/medium/xhigh) — most turns succeed, some die in ~2s; with the payload masked, users cannot tell validation errors from outages. Related opaque-400 classes: #37771, #37937.

Plugins

One local plugin (command-log.js) — irrelevant, bug reproduces with no plugins and no opencode at all (pure curl).

OpenCode version

2.0.30 (also reproducible via plain curl/python, client-independent)

Steps to reproduce

A. Deterministic masked 400 (boundary budget == max):

curl -sN https://opencode.ai/inference/go/anthropic/v1/messages
-H "Authorization: Bearer $OPENCODE_GO_KEY"
-H "anthropic-version: 2023-06-01"
-H "x-opencode-session: ses_00000000000000000000000000"
-H "Content-Type: application/json"
-d '{"model":"qwen3.8-flash","max_tokens":1024,"stream":true,"thinking":{"type":"enabled","budget_tokens":1024},"messages":[{"role":"user","content":"hi"}]}'

=> HTTP 400, body exactly: "event: error\ndata: {}"

(same with budget > max, and same for model qwen3.8-max)

B. Control — valid combination works (thinking blocks stream fine):

same as A but "max_tokens":4096,"budget_tokens":1024 => HTTP 200 + message_stop

C. Same validation-error class on another backend is passed through readably:

claude-haiku-5-5 with max_tokens 256 + budget 1024 => HTTP 400

{"type":"error","error":{"type":"invalid_request_error","message":"Upstream request failed: [invalid_request_error] "thinking..." }}

Screenshot

No response

Operating System

Ubuntu 24.04.5 LTS (Linux 6.8.0-147-generic x64)

Terminal

TERM=xterm-256color (KDE Plasma 5.27 workstation, not a server)