/screenpipe-chats

name: screenpipe-chats
description: Search what was said in existing screenpipe, Codex, Claude, Cursor, and Gemini CLI chats, then continue or steer one exact chat when the user explicitly asks. Use for requests about finding another agent conversation, recalling what an agent said, or sending work to one.

existing agent chats

screenpipe exposes `search_chats` and `send_to_chat` as agent tools. They use a private core capability passed into the current chat process; do not scrape transcript folders, guess runtime commands, or call a localhost app route.

search

1. Call `search_chats` with a short title, message, workspace, or exact id. The query also matches the transcript, so a phrase said mid-conversation finds the chat even when the title never mentions it. Omit the query to list recent chats. Filter `sources` only when the user named a runtime. 2. Show enough title, source, preview, and workspace information to distinguish ambiguous results. 3. Treat the returned `source` and `id` as the address. Never derive an id from a title or reuse a stale result after the target disappears. 4. Read the `warnings` array. Transcript reading is bounded to the most recent chats per runtime, so report that older chats were not searched instead of claiming a phrase was never said.

Searching is read-only and does not require confirmation.

send

1. Require an exact result from `search_chats` in the current turn. 2. Require explicit user authorization for that target and message. A draft, suggestion, or request to inspect chats is not permission to send. 3. Call `send_to_chat` with the exact `source`, `id`, message, and `confirmed: true`. 4. Use `mode: queue` by default. Use `mode: steer` only when the user asked to redirect a currently running screenpipe chat. 5. Report the returned status and exact target. Do not retry an uncertain or failed send until the result is reconciled.

Gemini CLI chats are search-only: Gemini has no non-interactive resume, so `send_to_chat` rejects them instead of opening an interactive session. Say so and offer to summarize what the chat already contains.

The tool refuses self-sends, guessed ids, dormant-chat steering, oversized messages, unsendable sources, and sends without confirmation. Cross-chat sending is disabled in unattended scheduled runs.