HaiKIT splits tool results: 67-token digests for the model, full data stays in the browser
Head_Inspector2200 · reddit · 2026-10-30
A developer open-sourced HaiKIT (MIT, TypeScript), a pattern for keeping bulky tool results out of LLM context. Every tool result splits into a required model-facing digest (47 flight rows compressed to 67 tokens, with cheapest/fastest precomputed plus a handle) and a UI channel streaming full rows to the browser. The model can query the handle via queryui, but only through cap functions so filters can't refill context; an 'elicit' mechanism suspends turns until user clicks supply the result; the browser can only send {handle, action, value}, forming an injection boundary. Adapters for Claude and OpenAI Responses API, plus a Postgres store with fenced turn leases.
More from coding & agent
- NVIDIA shows developers building Omniverse simulations with frontier AI agents — RevLebaredian · 2026-10-30
- Dev builds MCP server to visualize and manage parallel coding agents in a web UI — Shreevenkr · 2026-10-30
- Claude Projects is Anthropic's best idea and worst implementation, argues power user — Many_Region8176 · 2026-10-30
- Asking Anthropic Staff to Explain Claude Connectors vs MCP Gets You "RL'd" — AaronBergman18 · 2026-10-30
- Hot Take: The Figma File Is Becoming the Spec Doc Nobody Reads — talkaboutdesign · 2026-10-30
- Growing collection of CSS view transitions page effects, with copy-paste code for agents — jh3yy · 2026-10-30