DeepSeek V4 Flash Free: output truncated mid-sentence without warning

[realsharpswing]

Issue

DeepSeek V4 Flash Free model frequently stops output mid-sentence without any error or warning message. This makes normal conversation nearly impossible.

Symptoms

• Output cuts off after 1-2 lines
• Stops mid-sentence
• No error code or warning returned
• Requires repeated retries to complete a response
• Happens consistently across different prompts

Steps to Reproduce

1. Use DeepSeek V4 Flash Free model
2. Enter any prompt
3. Output stops abruptly after ~1-2 lines without completion signal

Environment

• OS: Windows 11
• Terminal: PowerShell 7
• OpenCode: latest version

Suggestion

Consider adding client-side auto-retry or a friendly warning when output is truncated, so users know the response was incomplete.

[github-actions]

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

• BUG REPORT: Output Truncation During Long Responses #39393: Same symptom — output silently truncated mid-response specifically with the DeepSeek V4 Flash Free model on Windows PowerShell
• [FEATURE]: Auto-continue when model hits output token limit (finish_reason: "length") #17471: Feature request for auto-continue when model hits output token limit (finish_reason: length), which describes the same underlying cause
• Local vLLM + Mistral Medium 3.5 can exit the loop on truncated assistant text with finish=stop #36657: Local vLLM model exits loop on truncated assistant text with finish=stop

If your issue is the same as one of the above, please add a comment there instead so discussion stays focused.

[kai-kleinbard-tyto]

Closing — no template used, insufficient information to investigate. Please reopen with the Bug Report or Feature Request template filled out (version, OS, repro steps, logs).

[northdeamon]

I can reproduce this issue reliably and have additional protocol-level evidence that may help isolate the cause.

I originally observed the same symptom in an agent client: assistant text was frequently cut off mid-word immediately before a tool call.

To rule out the client, UI, session storage, and downstream parsing, I reproduced the behavior by bypassing them completely and sending OpenAI-compatible HTTP/SSE requests directly to:

POST https://opencode.ai/zen/v1/chat/completions

Model:
deepseek-v4-flash-free

Request configuration:

• stream: true
• max_tokens: 512
• one simple function/tool
• tool_choice: auto

The prompt explicitly instructed the model to:

1. Write one normal complete sentence in Russian describing what it is going to check.
2. MUST finish that sentence with a period.
3. Only after the period call the provided function.

I repeated the test 6 times.

Result: the truncation reproduced in 5/6 requests.

Examples:

Trial 1:
"Я собираюсь проверить корректность вызова диагностической функции check_value с те"
→ structured tool call immediately follows

Trial 2:
"Сейчас я собираюсь проверить работу инструмента check_value с тестовым значением trial"
→ structured tool call immediately follows

Trial 3:
"Я собираюсь проверить корректность вызова диагностической функции с зада"
→ structured tool call immediately follows

Trial 4:
"Проверяю корректность передачи строкового значения параметра в диагности"
→ structured tool call immediately follows

Trial 5 completed correctly:
"Я собираюсь проверить корректность вызова диагностической функции с указанным значением."
→ structured tool call follows normally

Trial 6:
"Я собираюсь проверить корректность передачи значения при вызове диагности"
→ structured tool call immediately follows

Important details:

• HTTP status: 200 in all 6 trials
• Tool call emitted: 6/6
• Failing responses still ended normally with:
  finish_reason = "tool_calls"
• There was no:
  finish_reason = "length"
• No HTTP/provider error occurred.

Most importantly, the incomplete text is already present in the raw SSE response from the OpenCode Zen endpoint.

Therefore this does not appear to be a frontend rendering issue or downstream client/parser truncation.

The protocol transition in failing cases is effectively:

incomplete delta.content
→ structured delta.tool_calls
→ finish_reason = "tool_calls"

From a downstream client's perspective the response looks completely valid, so the client cannot reliably detect that the natural-language portion was accidentally truncated.

This seems to narrow the issue to the DeepSeek V4 Flash Free <-> OpenCode Zen tool-call path. Two possibilities I can see are:

1. DeepSeek V4 Flash emits the tool-call transition before completing its natural-language text, or
2. the OpenCode Zen tool-call adapter/parser detects the tool-call boundary too early and drops the remainder of the text.

I am attaching a sanitized standalone reproducer and the observed results.

Related issue #40176 may also be relevant because it involves DeepSeek V4 Flash behavior around the text → tool_calls transition, although its exact symptom is different.

Reproduction environment:

• Windows 11
• PowerShell
• Direct HTTP/SSE request
• No local OpenCode CLI involved
• No downstream agent framework involved
• Endpoint: https://opencode.ai/zen/v1/chat/completions
• Model: deepseek-v4-flash-free
• Reproduction rate in this run: 5/6

I can also provide the raw SSE captures from all six trials if that would help.

repro_direct_sse.zip
OBSERVED_RESULTS.md