Sift Skill

Sift is a small skill for summarizing noisy shell command output before it gets pulled into the main agent thread.

It is useful for commands like:

• cargo test
• cargo check
• npm test
• pytest
• ./gradlew test
• long-running scripts that print a lot of logs

By default it uses Codex for the nested summary step. It can also use Claude.

Why

Large command output is often low value:

• it increases token usage
• it makes later turns heavier
• it makes the main thread harder to read

Sift runs the command, captures the full output, and returns only a compact summary.

Estimate

Conservative estimate for command-heavy multi-turn sessions: about 45% lower total token cost.

This number comes from a real Codex session where:

• command output was 65.2% of known tool-output tokens
• summarizing that slice was estimated to be about 70% cheaper

This estimate is mainly for noisy build/test/debug workflows. If your commands are already short, the benefit will be smaller.

Disclaimer

Sift is a simple helper around nested agent calls.

• it may reduce token usage, but the exact savings depend on your workflow, model, and provider
• it does not guarantee correctness of the summary
• it should not replace reading the raw output when details matter
• the Claude path is supported in the script, but local CLI behavior can vary by Claude Code version and environment

Main options

• --engine codex|claude
• --instructions
• --cwd
• --model
• --shell-command
• --keep-temp
• --summary-file

Run ./sift/scripts/sift.sh --help for the full list.

License

MIT. See LICENSE.