$ cat ~/map/pi-share-hf.md

pi-share-hf

Collect, redact, and publish Pi session traces to a Hugging Face dataset.

officialtools✓ Oct 16

npm install -g pi-share-hf

## TL;DR

Official pipeline from Pi's author for publishing pi coding agent sessions of one OSS project to a Hugging Face dataset — collect, redact secrets, scan with TruffleHog, run LLM review, and upload only sessions that pass every check.

## Key points

• $Incremental six-stage pipeline: collect changed/new sessions for one project, redact exact secrets from the env file and --secret inputs, reject sessions matching --deny patterns, scan redacted output with TruffleHog, run LLM review, then upload only sessions that pass all checks.
• $Deterministic redaction only replaces exact secret values (from --env-file, default ~/.zshrc, plus repeatable --secret file/literal inputs); TruffleHog then scans the redacted output and any finding — verified, unverified, or unknown — blocks the session automatically.
• $LLM review sees the project context files plus a plain-text transcript of the redacted session and answers about_project, shareable, and missed_sensitive_data; review sidecars land in .pi/hf-sessions/review/, and changing provider, model, or thinking level changes the review cache key.
• $The .pi/hf-sessions/ workspace keeps state, so repeated runs only process what changed; pi-share-hf init maps a project directory to a Hugging Face dataset repo in a personal or --organization namespace, and the same workspace can be used from multiple machines.
• $Command set: init, collect, review, reject, list --uploadable, grep (searches only the uploadable set), and upload with --dry-run; upload uses a built-in TypeScript Hugging Face client, and auth works via HF_TOKEN or ~/.cache/huggingface/token.

## FAQ

## References

• README ↗
• Pi session format docs ↗
• TruffleHog ↗

Last updated: Aug 31, 2026