pi-history-ingest mines the JSONL sessions that the Pi coding agent stores under ~/.pi/agent/sessions. Each session is a tree, so the skill finds the current leaf, walks back to the root and reads that active path, giving extra weight to compaction and branch summaries that are already distilled.
Every durable claim is tied to a session entry through an evidence ledger and a source comment, and the draft is checked against that ledger before a page is written. Facts that do not appear in the sessions are left out.
What You Get
- Topic pages in your vault whose claims each carry a source comment pointing to a session file and entry
- An evidence ledger step that traces extracted facts back to the session text before anything is saved
- Provenance markers for inferred and ambiguous claims, plus summary, confidence and lifecycle frontmatter
- A manifest entry per session with project, source type and the pages it created or updated
- Append and full ingest modes
Common Use Cases
- Capturing the decisions and workflows from months of Pi sessions in one project
- Turning recurring bash commands into documented build, test and deploy playbooks
- Keeping a wiki current by ingesting only the Pi sessions added since the last run
What Makes It Unique
It treats the session files as the only source of truth: background knowledge from training, docs or the current chat is never added as extracted fact. A source verification gate checks that every claim, proper noun, command and filename appears in the cited entry, and unverifiable content is deleted or marked as inferred or ambiguous. It also understands Pi’s tree layout, so abandoned branches contribute through their branch summaries.
Under The Hood
The skill is instructions for the agent. It lists session files with the find command, reads and parses JSONL from the Pi session folder, and searches them with grep when a claim needs checking. It writes Markdown pages and the manifest into the vault and runs the obsidian-wiki command line tool to update the index, log and hot cache. If a QMD collection is set it runs the qmd command to refresh the search index. It does not fetch web pages or call remote APIs.