Repository navigation
deepseek-v4-flash via Console Go (opencode.ai/zen/go) stops after text output: no tool_calls, finish_reason stop #40176
Description
Summary
deepseek-v4-flash served through Console Go (opencode.ai/zen/go/v1) ends the response with finish_reason: stop whenever the model first emits a text passage before tool calls. The tool calls never arrive. An agent loop exits mid-task right after the model says what it is about to do.
Environment
- Client: Grok CLI → cliproxy (Responses→Chat bridge) → Tailscale Aperture → provider
opencode-go(Console Go,opencode.ai/zen/go/v1) →deepseek-v4-flash - Observed with
reasoning_effort: xhighandmax; streaming and non-streaming.
Evidence
Batch analysis of 15 captured responses on the gateway (one session, same request shape):
| finish_reason | count | text before tool calls |
|---|---|---|
tool_calls |
6 | 0/6 |
stop, no tool_calls |
7 | 7/7 |
Every early stop has text first; every normal response skips text. The stopped texts end with a colon, for example "…check the tests directory and summarize:".
Control experiment
Same model, same prompt, same tools, same multi-turn history against the official DeepSeek API (api.deepseek.com/v1/chat/completions), with instructions forcing text-then-tool output: 4/4 returned finish_reason: tool_calls, text followed by complete tool calls. The model itself continues normally after text.
Reasoning_content contract is satisfied
Assistant tool-call messages carry reasoning_content (229/229 on a 465-message request; empty string when no reasoning was stored). The early stop persists. This points to Console Go terminating its generation at the text→tool_calls transition, independent of request shape.
Related
- deepseek-v4-flash-free stops the agent loop after nearly every tool call #39204: same symptom ("model says it is about to continue" then stops)
- DeepSeek silently stops executing (interleaved reasoning_content dropped in tool call messages) #35689: reasoning_content contract root cause. Our data shows the stop persists with the contract satisfied, so this looks like a separate defect in the zen/go serving path.
Ask
Please check Console Go streaming generation for mixed text + tool_calls output from DeepSeek V4 (text deltas followed by tool-call deltas). It appears to end the response at the text boundary instead of continuing into tool calls.
Activity
LZHcode1986 commentedon Aug 8, 2026
Same symptom here, and I can add strong evidence that this is a zen/go serving-path defect, not a client issue.
Timeline
Observed pattern (matches your analysis exactly)
What I ruled out
Please prioritize the streaming text→tool_calls transition in the zen/go serving path. A control run against api.deepseek.com with the same prompt/tools completes tool calls 4/4, so the model itself is fine.
RicSchonfelder commentedon Aug 19, 2026
+1 — I am experiencing the exact same issue with OpenCode Go / DeepSeek (deepseek-v4-flash). The agent is coding along and then suddenly stops mid-task with no error, no retry, nothing. I have to keep sending "continue" commands to get it going again, and it stops again shortly after. It happens repeatedly within the same session.
OpenCode version: 1.18.18, using the OpenCode Go gateway.
GreyElaina commentedon Sep 19, 2026
i do never use opencode / opencode go so close this issue for me