provider: OpenCode Go Anthropic route masks upstream errors as "Provider request failed with HTTP 400" #51530

Activity

1. github-actions

   assigned

   fwang

   on Sep 26, 2026
2. github-actions commented on Sep 26, 2026

   

   github-actionsBot

   on Sep 26, 2026 – with GitHub Actions

   Contributor

   More actions

   ---

   This issue might be a duplicate of existing issues. Please check:

   • #50446: OpenCode Go Qwen3.8 Anthropic-route 400s where the provider error body is not surfaced.
   • #51348: OpenCode Go /v1/messages returns a 400 with an empty error payload; the report asks for a descriptive error.
3. a-d-gordienko commented on Sep 29, 2026

   

   a-d-gordienko

   on Sep 29, 2026

   More actions

   Same failure here with opencode-go/qwen3.8-flash — adding data points from a different platform, in case they help narrow it down:

   Environment

   • opencode 1.18.33 on Windows 11 (build 26200), PowerShell 7
   • Go subscription; model opencode-go/qwen3.8-flash
   • UI shows: Bad Request: event: error data: {}

   Deterministic — 5/5 attempts on 2026-09-29 (both plan and build agents):

   timestamp=2026-09-29T03:21:23.709Z level=ERROR message="stream error" providerID=opencode-go modelID=qwen3.8-flash session.id=ses_f14f124b0ffeKNd3m5ZK8V1TWp agent=plan error.error="AI_APICallError: Bad Request"
   timestamp=2026-09-29T03:34:00.450Z level=ERROR message="stream error" providerID=opencode-go modelID=qwen3.8-flash agent=plan error.error="AI_APICallError: Bad Request"
   timestamp=2026-09-29T04:27:57.036Z level=ERROR message="stream error" providerID=opencode-go modelID=qwen3.8-flash agent=plan error.error="AI_APICallError: Bad Request"
   timestamp=2026-09-29T06:42:40.440Z level=ERROR message="stream error" providerID=opencode-go modelID=qwen3.8-flash agent=build error.error="AI_APICallError: Bad Request"
   timestamp=2026-09-29T07:15:12.727Z level=ERROR message="stream error" providerID=opencode-go modelID=qwen3.8-flash agent=plan error.error="AI_APICallError: Bad Request"
   

   Control: in the same session/environment opencode-go/deepseek-v4.1-flash works fine, so the Go subscription/auth is not the issue — only qwen3.8-flash fails.

   This matches the empty event: error + data: {} payload described in the issue. Happy to run further diagnostics (fresh session, local recording endpoint for the request body) if that would help.
4. VictorHMH18713 commented on Oct 27, 2026

   

   VictorHMH18713

   on Oct 27, 2026

   More actions

   Related data point on the same route, different masking shape.

   • Client: Claude Code 2.1.293, ANTHROPIC_BASE_URL=https://opencode.ai/zen/go, model deepseek-v4.1-flash, stream: true.
   • Observed 2026-10-27 07:24–07:44Z and 07:53Z UTC.
   • Instead of an SSE event: error with {}, we get HTTP 400 with a 0-byte body (curl reports Content-Length absent, 0 bytes read), which Claude Code renders as API Error: 400 status code (no body).
   • The same request body sent with stream: false returns the real upstream reason:

   Upstream request failed: [invalid_request_error] The `content[].thinking` in the thinking mode must be passed back to the API.
   (request_id: 65871f21-0908-4a0f-bc02-89f7c5fb6d1f) (request id: 202610270826085305488478268d9d6RBCXETWj)
   

   • Sending the identical 67 KB body again minutes later returns 200, so the client has no way to distinguish "your request is invalid" from "this attempt hit a bad route". Full analysis in Go/Zen gateway: intermittent 400 "reasoning_content must be passed back" for clients that don't replay DeepSeek reasoning (store-miss on call_id) #58064.

   Since Claude Code does not retry 400 (408 / 409 / 429 / 5xx / x-should-retry: true only), the empty 400 is terminal there: it killed 8 of 25 workflow subagents in a 20-minute window. Any upstream error text (SSE event or plain JSON body) would at least let the client surface and report it.
5. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   Contributor

   More actions

   Thanks for the detailed report. I reproduced this with a local mock in place of the OpenCode Go Anthropic route. The mock returns what you and #57122 observed: HTTP 400 with event: error / data: {}. A provider using @opencode/ai/providers/anthropic pointed at it gives:

   • 2.0.18 and 2.2.0: Provider request failed with HTTP 400, with no further detail.
   • Latest dev (86ce19d): Provider request failed with HTTP 400: event: error data: {}. Since fix(ai): show provider error bodies when no message is recognized #51978 the raw response body is appended, which is your first suggested fix. Because the gateway sends an empty payload, the real upstream reason still never reaches you.

   The remaining problem looks like the Go gateway dropping the upstream error on the Anthropic route for qwen3.8 (your second suggestion). That gateway isn't in this repo, so I couldn't test it directly. Your workaround (overriding the model's package to @opencode/ai/providers/openai-compatible) is the way to see the real message for now.

   Related: #57122 (a deterministic curl reproduction of the same empty data: {} error) and #52791 (probably the same empty error frame showing up as an invalid stream event). The script below reproduces this against a mock server.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduces #51530 client-side: a gateway that rejects an Anthropic Messages request with
   # HTTP 400 and an empty SSE error payload (`event: error` / `data: {}`), as the OpenCode Go
   # route does for qwen3.8-flash. Shows what opencode reports to the user.
   # Usage: bash repro.sh            (uses `opencode` from PATH)
   #        OC_DEV=/path/to/repo bash repro.sh   (runs from a source checkout)
   set -u
   DIR=$(mktemp -d)
   PORT=${PORT:-48151}
   cat > "$DIR/mock.ts" <<EOF
   Bun.serve({
     port: $PORT,
     async fetch(req) {
       const url = new URL(req.url)
       const body = await req.text()
       let stream = true
       try { stream = JSON.parse(body).stream !== false } catch {}
       console.error("mock:", req.method, url.pathname, "stream=" + stream)
       if (stream) return new Response("event: error\ndata: {}\n\n", { status: 400, headers: { "content-type": "text/event-stream" } })
       return Response.json({ model: "qwen3.8-flash" }, { status: 400 })
     },
   })
   EOF
   bun "$DIR/mock.ts" & MOCK=$!
   trap 'kill $MOCK 2>/dev/null; rm -rf "$DIR"' EXIT
   sleep 1
   mkdir -p "$DIR/project" && cd "$DIR/project" && git init -q
   cat > opencode.jsonc <<EOF
   {
     "providers": {
       "mockgo": {
         "package": "@opencode/ai/providers/anthropic",
         "settings": { "baseURL": "http://127.0.0.1:$PORT/v1", "apiKey": "test" },
         "models": { "qwen3.8-flash": {} }
       }
     }
   }
   EOF
   if [ -n "${OC_DEV:-}" ]; then
     OC=(bun --conditions=browser --preload="$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js" "$OC_DEV/packages/cli/src/index.ts")
   else
     OC=(opencode)
   fi
   echo "opencode version: $("${OC[@]}" --version)"
   SERVER=()
   if [ -n "${OC_DEV:-}" ]; then
     # From source, start the server with the same preload (a spawned server would miss it).
     export OPENCODE_PASSWORD=repro-pass
     "${OC[@]}" serve --port 48152 >"$DIR/serve.log" 2>&1 & SRV=$!
     trap 'kill $MOCK $SRV 2>/dev/null; rm -rf "$DIR"' EXIT
     for i in $(seq 60); do curl -s -o /dev/null http://127.0.0.1:48152/ && break; sleep 1; done
     SERVER=(--server http://127.0.0.1:48152)
   fi
   timeout 120 "${OC[@]}" run "${SERVER[@]}" --model mockgo/qwen3.8-flash "hello" 2>&1 | tail -20

   Run it: bash repro.sh
6. opencodejr

   added

   reproduced

   on Oct 28, 2026