Bug: DeepSeek V4 Pro model unstable on Command Code

[doutorprestes]

DeepSeek V4 Pro model unstable on Command Code

Environment:

• Command Code v0.27.0
• Model: DeepSeek V4 Pro (-m deepseek-v4-pro)
• Platform: Linux (x64), Node v24.15.0

Steps to reproduce:

1. Run cmd -m deepseek-v4-pro or specify DeepSeek V4 Pro via config
2. Attempt to use any skill or submit a coding/design task
3. Model produces inconsistent results — sometimes works, often fails or hangs

Observed behavior:

• Responses are unpredictable: short prompts may work, longer prompts fail
• High variability between runs with identical inputs
• In some cases the model appears to connect but produces no output
• In other cases the session hangs similarly to the /design long-prompt issue
• Switching to the default model (taste-1) resolves the stability issues

Expected behavior:

• DeepSeek V4 Pro should be a stable, usable model option within Command Code
• Model selection should not introduce reliability regressions
• Failures should be reported with actionable error messages

Impact:
DeepSeek V4 Pro is effectively unusable in Command Code today. Skills that would benefit from this model (e.g., /design, complex refactoring) cannot rely on it.

Additional context:

• This may be related to the /design long-prompt hang (separate issue #)
• Only the default taste-1 model is currently stable
• Consider surfacing model instability warnings to users at connection time