ctx v0.22.0

v0.22.0 — Dual-Model Dream Architecture

Dream Mode can now use a separate, larger model for background
cross-referencing while synthesis stays on the fast model. Evaluated
10 Gemma models — qwen3.5 family remains unbeaten. qwen3.5:27b is the
only model to pass all 4 dream tests including causal relationship
detection.

Per-function LLM configuration:

• OLLAMA_DREAM_MODEL (separate model for dream evaluation)
• OLLAMA_DREAM_NUM_CTX / OLLAMA_CHAT_NUM_CTX (per-function context)
• OLLAMA_DREAM_THINK / OLLAMA_THINK (per-function think mode)
• Removed global ThinkMode — think *bool threaded through full chain

Continuous dream loop:

• Replaced 10s ticker with back-to-back processing
• 120s idle wait when no blocks available
• 2s yield to active queries (demand interruption)
• 10s pause on error to prevent tight loops

Gemma evaluation results (Session 21):

• 10 models tested (gemma3:4b/12b/27b, gemma3n:e2b/e4b,
  gemma4:e2b/e4b/26b/31b, embeddinggemma)
• Best Gemma for dream: gemma4:e2b (3/4), but only 80% synthesis KW
• Larger models consistently worse (gemma3:12b 0/4, gemma4:31b 0/4)
• gemma4:26b initially scored 0/4 due to markdown fence wrapping
  (format:"json" not respected); actual quality 2/4 after strip

VRAM budget (24 GB Quadro RTX 6000):

• Query path: qwen3-embedding:8b (5.3 GB) + qwen3.5:9b (9.3 GB) = 14.6 GB
• Dream path: qwen3.5:27b (22.2 GB at num_ctx=16384) — Ollama swaps

Installation

With Go:

go install github.com/GottZ/ctx/cmd/ctx@v0.22.0

Binary download:
Download the binary for your platform, make it executable, move to PATH:

chmod +x ctx-*
sudo mv ctx-* /usr/local/bin/ctx