DeepPi

Reasonix-style DeepSeek price/performance for the Pi coding agent.

DeepPi targets the direct DeepSeek API only. It stabilizes cacheable request prefixes, reports measured cache economics from Pi's real usage records, and reduces paid retries with loop guards and hash-verified edits.

Supported models

• deepseek-v4-flash
• deepseek-v4-pro

DeepPi is dormant for every other provider and model.

Install

pi install npm:@arter/deep-pi@1.0.0

To try it for a single run without installing:

pi -e npm:@arter/deep-pi

Reload Pi, select a supported direct DeepSeek model, and run /deeppi.

What /deeppi measures

• cache-read and uncached input tokens;
• cache-hit rate;
• actual input cost from Pi;
• estimated savings against fully uncached input;
• detected local prefix churn;
• guarded retry loops and hashline edit outcomes.

DeepPi does not promise a fixed hit rate. Provider cache expiry and backend state can produce misses even when the local request prefix is stable.

Development

npm install --ignore-scripts
npm run verify

Default verification never calls an external API. The paid smoke benchmark is explicitly opt-in:

DEEPPI_LIVE=1 npm run benchmark:live

Attribution

DeepPi is derived from jrimmer/pi-deepseek-optimized and is licensed under the Apache License 2.0. The original project implements techniques described by Howard Chen and Can Akay; their credits are retained.

DeepPi's additions are its exact direct-V4 boundary, measured Pi cache/cost telemetry, prefix-churn diagnostics, batch-aware retry guards, atomic hashline edits, removal of destructive rewind behavior, and tests against current Pi event shapes.