fix: Streaming stalls mid-generation with Ollama models, requires user input to resume

[jstrockin]

Bug Description

When using Ollama models, streaming output stalls mid-generation.
The agent loop pauses and waits for user input before continuing.
Typing anything (Enter or a short message) causes generation to resume.
Expected behaviour: streaming continues uninterrupted without user intervention.

Steps to Reproduce

1. Run kit with any Ollama model (tested: ollama/gemma4:latest, ollama/qwen3:14b)
2. Ask a question that produces a long or multi-step response
3. Observe output stops mid-generation
4. Type anything and press Enter — generation resumes

Relevant Code / Configuration

temperature: 0.6
top-k: 20
top-p: 0.95
max-tokens: 8192
thinking-level: "off"
max-steps: 50

Affected Component

Streaming / Ollama provider

Kit Version

dev (go install github.com/mark3labs/kit@latest, built 2026-05-25)

Additional Context

Ollama version: 0.30.6
Hardware: Apple M4 Mac mini
No keepalive or poll-interval flags available in kit --help.
Issue occurs consistently across multiple sessions and models.

Checklist

• I've searched existing issues and this hasn't been reported yet
• I've tested with the latest version of Kit

[mark-iii-labs-huly]

Connected to Huly®: KIT-51

[ezynda3]

Are you sure this is not a max tokens issue? Try bumping max tokens to something like 16k. If the response is longer than 16 any model will generally just stop generating and finish their turn.

[jstrockin]

I have selected the context length in Ollama as 128k

[ezynda3]

You need to change you kit config specifically max-tokens

temperature: 0.6
top-k: 20
top-p: 0.95
max-tokens: 8192
thinking-level: "off"
max-steps: 50

[jstrockin]

I changed it to below and still feels model forgets it was working. In another instance it said it changed the file but in reality no changes were made and tool calls entirely failed.
temperature: 0.6
top-k: 20
top-p: 0.95
max-tokens: 32768
thinking-level: "off"
max-steps: 50