Retrieval (recall)

hotl’s default retrieval is agentic: the model greps and reads the working tree, which is always current and never leaves your machine. The recall tool is for corpora that outgrow that — a large notes directory, team docs, anything you can’t grep because you don’t know the keywords.

Nothing is on by default. Configure a backend and the model gains one recall tool; configure none and the tool doesn’t exist.

Configuring a backend

P1 supports one backend kind: an MCP server that exposes a search tool.

~/.config/hotl/config.toml

[[retrieval]]
name = "notes"
kind = "mcp"
command = "/usr/local/bin/notes-rag"
args = ["--stdio"]
tool = "search"          # the MCP tool recall calls (default: "search")
description = "personal notes search"

The server’s tool is called with {"query": "...", "purpose": "...", "k": 8} and its text reply is returned to the model as the search result.

Trust and safety

• The first use of a backend raises the protected ask with the server binary’s SHA-256 — the same screen, and the same trust.toml, as the mcp tool. After that, each search is a plain y/n ask.
• Everything a backend returns is wrapped in the untrusted-content envelope with recall:<backend> provenance: retrieved text can inform the work but cannot authorize tool use or override your instructions.
• Results are capped at 50 KB; oversized results are spilled to a blob with a preview and a read-back pointer.
• hotl ships no cloud backend. A backend only reaches the network if the program you configured does — choose local ones.

Several backends

Add more [[retrieval]] sections; the model then picks with the backend argument. Keep descriptions specific (“personal notes”, “platform docs”) — the model routes on them.