pi-session-mentions

简体中文

Reference local Pi, Claude Code, and Codex conversations in a Pi message.

Type @@ for recent sessions, then type a keyword to filter. Select a result with the arrow keys and Tab or Enter, then send your message as usual. For several keywords, use quotes: @@"build cache".

It adds model context only when you send a selected reference.

The ordinary @ file menu has one Sessions entry that inserts @@. Session results have their own menu showing the title, agent, project, and date. Use the arrow keys to scroll through all matches. File results keep their order.

Install

Requires Pi 1.2.0 or later and Node.js 22.19.0 or later. Install from npm:

pi install npm:pi-session-mentions

Run /reload in an existing Pi session. This is a personal installation, so the extension works across projects. It has no runtime dependencies outside Pi and Node.js, and does not require ah.

Search

Use a short prefix to choose a source:

Type a keyword after the prefix, for example @@cx:cache or @@cc:"build cache". Type @@: to choose All, CC, CX, or Pi from a small menu. Tab or Enter inserts the source prefix; then type a keyword or press Tab to browse its sessions. Typing @@c also suggests CC and CX.

Search matches titles, user messages, project paths, and session IDs. Matching ignores case and supports Chinese text. All query words must match. Title matches come first, followed by the current project and recent activity. The current session and subagent sessions are excluded.

Assistant messages, tool output, and thinking are not indexed. This keeps the index smaller and avoids matches on generated text.

The index loads in a background worker and caches unchanged files. During the first scan, the status line shows the number of indexed sessions. Results appear when that scan finishes, and the status clears. You can keep typing while it runs. A complete cache makes results available as soon as it loads. Later file refreshes run in the background, at most once every 15 seconds when you search. Typing searches the in-memory index. There are no model calls or embeddings.

Search time depends on the indexed text and the number of returned matches. The worker keeps user text in memory, so memory use grows with your history. An uncached first scan also reads the source files. Changing or closing the Pi session stops the worker and releases its index.

Pi's current session directory is also included when using a custom directory. There is no settings screen or extra command.

What gets sent

Selecting a session inserts a short reference containing its title. On submission, the extension replaces it with a numbered reference and adds its agent, project, original path, and a local transcript path. Multiple references to the same session share one attachment.

The transcript contains user: and assistant: text with source information. It omits tool calls, tool results, images, system instructions, and thinking. The original session stays unchanged. Each export is a snapshot, so a later edit to the original does not change a reference that was already sent.

Only the reference and paths enter the initial model context. The model can read the transcript as needed using its file tool. This reduces the initial token cost; reading the conversation still uses tokens. The model needs a local file reader to use the reference.

References inside fenced or inline code and references escaped with \ are left as text. An unavailable source is explicitly reported in the sent message. Existing image attachments are preserved.

Pi exports follow the latest recorded branch. Claude Code exports follow parent links, and Codex exports apply recorded rollbacks. If a branch ancestor is missing, the transcript explicitly marks the available history as incomplete. A file that changes during export must be selected again after its current write finishes.

Local data

The search cache lives in ~/.pi/agent/session-mentions/, or under PI_CODING_AGENT_DIR. Transcript snapshots are written to /tmp/pi-sessions-mentions/. Both contain private conversation text and use private directory and file permissions. No data is sent to a separate search service. Referenced text can reach the model provider when the model reads it; this is not secret redaction.

Each submitted session reference generates a fresh transcript. Multiple references to the same session in one message share one snapshot. Deleting the search cache causes it to be rebuilt. Deleting the temporary snapshots, including during system cleanup, makes their paths in older messages unreadable. Selecting and sending a session again creates a new snapshot. Original sessions are never modified.

Development

From the repository root:

npm install
npm run check --workspace pi-session-mentions

MIT; see LICENSE.