Sift – save AI tokens in Codex/Claude by summarizing command output

Name: Sift – save AI tokens in Codex/Claude by summarizing command output
Availability: InStock
Author: piotrwilczek

by piotrwilczek·Apr 22, 2026·2 points·3 comments

Visit Project View on HN

AI Analysis

●●SolidSolve My ProblemNiche Gem

Nested agent summarization cuts token costs ~45% for command-heavy workflows.

Strengths

•Targets specific pain point: verbose test/build output bloating agent context
•Honest about limitations—doesn't guarantee correctness, workflow-dependent savings
•Simple implementation with clear CLI options for Codex and Claude

Weaknesses

•Only helps when commands produce large noisy logs, minimal benefit for short outputs
•Summary may lose details that matter for debugging

Post Description

I made a small skill/script for agentic coding workflows:

https://github.com/panpeter/sift-skill

The idea is simple:

when a command like cargo test, pytest, npm test, or ./gradlew test prints a lot of output, that raw log often gets pulled into the context even though only a small part is actually useful.

Sift runs the command, captures the full output, and asks a nested cheaper agent call to return only a compact summary. The goal is to keep the main thread smaller and easier to work with.

In one command-heavy Codex workflow I used as a baseline, this looked like roughly ~45% lower total token cost. That number is an estimate, not a benchmark claim, and it will depend a lot on the workflow. It should help most when commands produce large noisy logs. It probably does very little for short commands.

Would be interested in how others handle this problem in Codex / Claude / other coding-agent setups.