YuanyuanMa03/dsh-funnel

README 摘要

dsh-funnel English 中文 Curate tool output before it enters the model's context — keep the lines that matter, spill the rest to disk with a pointer. Faster turns, smaller context, sharper attention, for every tool. The problem Every tool result is appended to the conversation and re-sent on every subsequent model request . A npm test run that prints 500 lines rides along for the rest of the session: you pay for it on each turn, it fills the context window, and it buries the one AssertionError the model actually needs. The existing layers don't cover this: Layer What exists What it doesn't do Capture (executor) 64 MB spill protects the process from OOM nothing about what the model sees Execution (bash only) maxOutputBytes byte-cap, head truncation shell tools only; blind cut, no error lines kept Compaction time official tool-result-pruner rewrites over-budget results on the surface fires only when compacting; blind head/tail; model can't read the original back Ingestion — every result, immediately — ← dsh-funnel Compaction-time pruning cleans up after the budget is already blown; dsh-funnel keeps it from being blown. They compose. What you get 1. Install and forget. One command, zero …

分类