openrouter: Gemini fails with 400 "Requests ending with a model turn are not supported" after read tool returns images, and error hides metadata.raw #55694

Activity

1. opencode-agent commented on Oct 15, 2026

   

   opencode-agentBot

   on Oct 15, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report and the code pointers. I reproduced both parts with a mock OpenRouter endpoint on 2.0.29, and they still happen on the latest dev (4397ee4).

   What I did: I configured openrouter/google/gemini-4-argon against a local mock. The mock makes the model call read on a PNG, then answers the next step with the same 400 body you posted.

   • Request shape: the continuation request ends with assistant (tool_calls) → tool ("Image read successfully") → user [image_url]. That last synthetic user message contains only the image, with no text part. It comes from flushAttachments in packages/ai/src/protocols/openai-chat.ts, as you pointed out.
   • Error message: OpenCode shows only Error: Provider returned error. Google's message ("Requests ending with a model turn are not supported.") is in error.metadata.raw and never reaches you.

   I couldn't confirm against the real OpenRouter/Google API that this shape is what gets rejected, because that needs a real account. Your 4/4 failures right after image reads point that way, though.

   The script below reproduces this without credentials.

   Related, but on v1 with different triggers: #55514, #47034 and #45359 (the same Gemini error), and #23665 ("Provider returned error" hiding the upstream message for a different provider).

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Repro for #55694 (OpenCode v2): run a mock OpenRouter endpoint, let the model `read` a PNG, and look at
   #  (1) the continuation request that is sent (it ends with a user message containing only image_url parts), and
   #  (2) the error OpenCode shows for OpenRouter's 400 body (only "Provider returned error"; metadata.raw is dropped).
   # Needs bun, python3 and `opencode` on PATH (bun add -g @opencode/[email protected]).
   set -euo pipefail
   D=$(mktemp -d); P=$D/proj; mkdir -p $P $D/home
   export HOME=$D/home XDG_CONFIG_HOME=$D/home/.config XDG_DATA_HOME=$D/home/.local/share
   python3 - "$P/render_1.png" <<'PY'
   import sys,zlib,struct
   w=h=8; raw=b''.join(b'\x00'+b'\xff\x00\x00'*w for _ in range(h))
   c=lambda t,d: struct.pack('>I',len(d))+t+d+struct.pack('>I',zlib.crc32(t+d)&0xffffffff)
   open(sys.argv[1],'wb').write(b'\x89PNG\r\n\x1a\n'+c(b'IHDR',struct.pack('>IIBBBBB',w,h,8,2,0,0,0))+c(b'IDAT',zlib.compress(raw))+c(b'IEND',b''))
   PY
   cat > $P/opencode.json <<'JSON'
   { "model": "openrouter/google/gemini-4-argon", "permissions": { "*": "allow" },
     "providers": { "openrouter": { "settings": { "baseURL": "http://127.0.0.1:4599/api/v1", "apiKey": "sk-test" },
       "models": { "google/gemini-4-argon": { "name": "Gemini 4 Argon", "capabilities": { "input": ["text","image"], "tools": true } } } } } }
   JSON
   cat > $D/mock.ts <<'TS'
   import { appendFileSync } from "fs"
   const raw = JSON.stringify({ error: { code: 400, message: "Requests ending with a model turn are not supported.", status: "INVALID_ARGUMENT" } }, null, 2) + "\n"
   const errBody = JSON.stringify({ error: { message: "Provider returned error", code: 400, metadata: { raw, provider_name: "Google", is_byok: false, provider_error_code: "400" } } })
   Bun.serve({ port: 4599, async fetch(req) {
     if (req.method !== "POST") return Response.json({ data: [] })
     const body: any = await req.json()
     appendFileSync(process.env.LOG!, JSON.stringify(body) + "\n")
     if (!body.messages?.some((m: any) => m.role === "tool")) {
       const ch = (d: any) => `data: ${JSON.stringify({ id: "c1", object: "chat.completion.chunk", created: 0, model: body.model, choices: [d] })}\n\n`
       return new Response(ch({ index: 0, delta: { role: "assistant", tool_calls: [{ index: 0, id: "call_1", type: "function", function: { name: "read", arguments: JSON.stringify({ path: process.env.IMG }) } }] }, finish_reason: null })
         + ch({ index: 0, delta: {}, finish_reason: "tool_calls" }) + "data: [DONE]\n\n", { headers: { "content-type": "text/event-stream" } })
     }
     return new Response(errBody, { status: 400, headers: { "content-type": "application/json" } })
   }})
   TS
   LOG=$D/requests.jsonl IMG=$P/render_1.png bun $D/mock.ts & MOCK=$!; trap 'kill $MOCK' EXIT; sleep 1
   (cd $P && opencode run --standalone --auto -m openrouter/google/gemini-4-argon "Look at render_1.png" || true)
   echo "--- message roles of the last request sent to the provider:"
   tail -1 $D/requests.jsonl | python3 -c 'import json,sys
   for m in json.load(sys.stdin)["messages"]:
     c=m.get("content"); print(m["role"], [p.get("type") for p in c] if isinstance(c,list) else repr((c or "")[:40]), "tool_calls" if "tool_calls" in m else "")'

   Run it: bash repro.sh
2. opencode-agent

   added

   reproduced

   on Oct 15, 2026
3. spragginsdesigns commented on Oct 15, 2026

   

   spragginsdesigns

   on Oct 15, 2026

   More actions

   Confirmed against the real OpenRouter API, and narrowed down to the exact cause.

   Setup: OpenCode 2.0.29 on Windows 11, openrouter/google/gemini-4-argon, a clean config (no plugins, isolated HOME). A logging pass-through proxy sat between OpenCode and openrouter.ai. Prompt: "Use your read tool to open render_1.png, then tell me in one sentence what the image shows."

   The first request returned 200 (the model calls read). The continuation request returned 400 every time:

   system | user | assistant(tool_calls, reasoning_details=[reasoning.encrypted]) | tool("Image read successfully") | user[image_url]
   -> 400 {"error":{"message":"Provider returned error","code":400,"metadata":{"raw":"{\"error\":{\"code\":400,\"message\":\"Requests ending with a model turn are not supported.\",\"status\":\"INVALID_ARGUMENT\"}}","provider_name":"Google",...}}}
   

   I then replayed that exact captured request (non-streaming), changing one thing per run:

   Variant	Result
   As sent by OpenCode	400 (3 of 3 runs)
   Strip reasoning_details from the assistant message	200
   Add a text part before the image in the trailing user message	200 (2 of 2 runs)
   Move the image into the tool message content instead	400
   Drop the image message entirely	200

   So the 400 needs both of these:

   • encrypted reasoning on the tool-call turn
   • a trailing user message that contains only image_url parts

   The same shape with no reasoning_details passes, which is probably why a simple mock or a single-turn test doesn't show it. Google's error presumably comes from how OpenRouter translates an image-only user turn when there are thought signatures.

   Suggested fix: have flushAttachments in packages/ai/src/protocols/openai-chat.ts always emit a short text part with the attachments, for example [{type:"text", text:"Attachments from the read tool:"}, ...images]. In my replays that alone turns the 400 into a 200 without dropping the reasoning. Surfacing error.metadata.raw would also have made this much faster to find, since OpenCode only shows "Provider returned error".

   In real use this makes Gemini 4 Argon effectively unusable in OpenCode for any task that reads an image. I had 8 of 8 failures right after an image read over two long sessions. The successful steps around those failures still bill the full context, so a session racks up cost while getting nowhere.