hackernewsagent-harness
Keep a coding-agent conversation append-only to preserve prompt caching
At a glance
Thorsten builds Caudra after finding that plugins could not add the control points he needed. This workflow is listed under agent-harness, prompt-caching, developer-tools, context-management.
- Agents / tools
- Caudra, Maki
Building a tool for these workflows? See weekly rates.
What they did
He extended a Rust coding harness with queued and mid-turn steering, cached side questions and nudges for smaller models that stop early. The conversation stays append-only to preserve the provider's prompt cache.
How it went
From his own use, the author reports about 97% cache hits with Anthropic models and 94% with OpenAI models.
Worth knowing
The project is still in preview, and the cache figures are the author's observations rather than a controlled comparison.
Save ideas in this browser. Upvotes count across the board.