pi-event-search

Event-oriented cross-session history retrieval for the Pi coding agent.

pi-event-search treats Pi's persisted session entries as durable evidence. It builds a disposable local index of searchable event fragments, returns hits anchored to exact session entries, and lets callers inspect the surrounding branch and relationships when a snippet is not enough.

The MVP public tools are event_search, event_read, and event_trace. The core search/read/trace engine is implemented and tested; the extension entrypoint in extension.ts registers those tools with the Pi coding agent.

Install

pi-event-search requires Pi running on Node.js 24 or newer. Install the pinned package release globally for Pi:

pi install npm:pi-event-search@0.0.1

Start Pi normally in a trusted project. The package registers event_search, event_read, and event_trace; no -e flag is needed after installation.

For development from a checkout:

npm install
pi -e ./extension.ts

Why this exists

Most session-search tools answer “which old session mentions this?” They commonly flatten a complete session into one search document or expose event-shaped data internally while returning only a session summary.

This project instead starts from the event:

Pi session JSONL
  -> persisted session entry
  -> one or more semantic fragments
  -> ranked event hit
  -> exact entry read and relationship trace

A search result should retain what produced it: the session, entry, fragment kind, time, branch position, and source relationship. Search snippets are discovery aids, never substitutes for the underlying entry.

Principles

• Session logs are canonical. The index is derived and can be deleted and rebuilt.
• Every hit has provenance. Search results lead back to a stable (sessionId, entryId) source.
• Events remain typed. User text, assistant text, tool calls, tool results, compactions, and branch summaries are not flattened into an indistinguishable transcript.
• Pi's tree matters. Parentage, alternate branches, compaction, and the selected materialized branch remain queryable facts.
• Session forks remain connected. Pi --fork copies are linked back to their authorized parent session with full event identities.
• Retrieval is bounded and explicit. Search returns small snippets; exact reads report any presentation truncation.
• Model access is scoped. A Pi tool must not silently expose sessions from unrelated workspaces.
• Private reasoning is excluded from the MVP. Raw sessions remain untouched, but thinking content is neither indexed nor returned.
• Memory is a separate concern. This project retrieves historical evidence; it does not extract durable facts, rewrite history, or decide what a user should remember.

Design documents

• Core concepts
• Tool design
• Persistent index lifecycle

Implementation status

• src/parser.ts — JSONL parser for fixture and Pi persisted session formats; header-only reader for scoped discovery.
• src/projector.ts — typed semantic projection; private thinking excluded.
• src/tree.ts — append/branch order, materialized leaf, fork rules.
• src/relationships.ts — recorded/inferred relationship extraction.
• src/index/provider.ts — in-memory SQLite FTS5 search/read/trace provider.
• src/index/maintainer.ts — incremental append/rebuild/removal lifecycle; scoped root refresh plus hot-path stat fast path.
• src/auth/* — session discovery and workspace-root authorization.
• src/api/service.ts, src/tools.ts — bounded public tool layer.
• src/pi-adapter.ts — indexes the current session from Pi's SessionManager.
• extension.ts — Pi package entrypoint. Full scoped discovery runs only at startup/workspace change; turns and tools sync only the current session file.

Run the test suite and typecheck:

npm install
npm run check

npm run check runs tsc --noEmit plus the full node --test suite.

Startup scope and large histories

The extension resolves the current Git worktree as its default workspace scope. Outside Git it falls back to the current working directory. Historical startup indexing is capped at 32 MiB and 100 recent session files so a broad directory cannot freeze Pi while an in-memory index is built. The tool reports partial coverage whenever that cap is active; the current session is synchronized separately.

The startup policy can be configured with environment variables:

• PI_EVENT_SEARCH_WORKSPACE_ROOT — explicit authorization and discovery root.
• PI_EVENT_SEARCH_STARTUP_BUDGET_MB — source-byte budget; use unlimited to disable.
• PI_EVENT_SEARCH_STARTUP_FILE_LIMIT — session-count budget; use unlimited to disable.

Pi's own PI_CODING_AGENT_SESSION_DIR setting is honored for session discovery.

Initial boundary

The first useful version should provide lexical, event-level search and bounded source reads over local Pi JSONL sessions. It uses an automatically resolved Git worktree root (falling back to the current working directory) as its default authorization scope, with explicit root overrides for unusual layouts.

Embeddings, automatic recall injection, memory consolidation, session grouping, reasoning retrieval, and custom-event extractor registries can be considered after the event model and authorization boundary are proven.

License

MIT