Skip to content

v0.50.0: Sliding-window analyzer, updated model defaults

Choose a tag to compare

View all tags
@hughpyle hughpyle released this 19 Feb 18:13
· 890 commits to main since this release
v0.50.0
645e497

What's new

  • Sliding-window analyzer is now the default for keep analyze. Token-budgeted sliding windows with XML target marking and hash-based dedup across windows — works well with both local and cloud models.
  • Model-aware context budgets — the analyzer automatically selects the right context budget based on your configured model (3k for llama3.2:3b, 12k for claude-haiku-4.5, etc.)
  • Updated default models — Anthropic→claude-haiku-4.5, OpenAI→gpt-4.1-mini (best quality/cost from eval)
  • GPT-5 model support — OpenAI provider handles max_completion_tokens and temperature constraints for GPT-5 family
  • Default prompt: temporal-v4 — improved synthesis quality with stricter no-copy rules
  • Multi-provider test harness for comparing analysis across Ollama, Anthropic, OpenAI, and Gemini

Model comparison results

Model $/MTok Parts Quality
gpt-5-nano $0.05 0 Failed
gpt-4o-mini $0.15 46 No synthesis
gpt-5-mini $0.25 18 Good
gpt-4.1-mini $0.40 13 Excellent
claude-haiku-4.5 $1.00 12 Excellent
gpt-4o $2.50 12 Excellent
llama3.2:3b free 3 Poor
Assets 2
Loading