Repository navigation
Data folder and settings
Everything Dive keeps lives in one folder: ~/dive, the dive folder in your home folder. This page lists what's inside, which settings belong to one mode and which are shared, and what Dive keeps elsewhere.
Set DIVE_DATA_DIR before Dive or its server starts:
DIVE_DATA_DIR=~/dive-test npm run dev:webDive reads the variable once, at start-up. The desktop app passes its environment on to the server it starts, so setting it once covers both. The tests use it too, so they never touch real data.
| Folder | What it holds |
|---|---|
conversations/ |
Conversation history, one file per mode: <mode>-conversations.json
|
attachments/ |
Uploaded images, named img_ plus a SHA-256 hash of the content |
notes/ |
One Markdown file per note, plus .active, which names the note open in the Notes panel |
lessons/ |
One file per mode: <mode>-lessons.md
|
system-prompts/ |
System prompts you've edited: <mode>-dboff.md for DATABASE off and <mode>-dbon.md for DATABASE on |
skills/ |
Your Agent Skills |
plugins/ |
External tools |
plugin-drafts/ |
Drafts written by propose_plugin. Nothing loads from here |
workspace/ |
The folder file_operations works in. fetch_paper saves papers to workspace/papers, and the browser saves screenshots to workspace/browser
|
| File | What it holds |
|---|---|
ui-settings.json |
Appearance for each mode (palette, font, font size), whether the thinking panel opens expanded, the enabled modes and the default mode |
skills_config-<mode>.json |
That mode's tool and skill switches and INPUT buttons |
custom_skills-<mode>.json |
That mode's custom tools |
cloud-settings.json |
Cloud provider, models, base URLs, API keys, maximum output tokens and Agent Mode |
local-model-settings.json |
llama.cpp and LM Studio: server address, model, sampling parameters, native tool calling and Agent Mode |
ollama-settings.json |
The Ollama server address |
llamacpp.json |
The llama.cpp manager: models folder, port, binary, extra arguments and each model's settings |
pi-settings.json |
Pi mode's settings, and Dive's own server port |
library-config.json |
Library sources, search settings and where the library database lives |
book-search.json |
Book search provider keys and the Calibre server |
skill-paths.json |
Extra folders to scan for skills |
allowed-dirs.json |
Extra folders for the code tools. It doesn't exist until you create it |
coding-settings.json |
Code tool settings, such as pythonVenv for run_python
|
prompts.json |
Saved prompts |
| File | What it holds |
|---|---|
library.sqlite |
The library index. Its location can be changed with databasePath in library-config.json
|
library-index-job.json |
Progress of an index run, so it can resume after a restart |
library-index-errors.jsonl |
Files and embeddings that failed during indexing |
security-events.jsonl |
Permission requests and what happened to them |
deleted-tombstones.json |
Recently deleted conversations, so a reply that is still running can't bring one back |
daemon.log, daemon.error.log
|
Output of the background server that install-launchagent.sh sets up for a copy run from source |
Dive reads these to move older installs to the current layout: conversations.json, skills_config.json, custom_skills.json, lessons.md, notes.json and the llamacpp-models/ folder, whose models now belong in ~/models.
Kept separately for each mode:
- conversations and the chosen model
- system prompts and lessons
- tool and skill switches, INPUT buttons and custom tools
- MCP servers
- Agent Mode, the tool budget and native tool calling
- colour palette, font and font size
Shared by every mode:
- Dive's server port
- Cloud API keys
- Pi settings
- library sources and the index
- allowed folders, book search providers and extra skill folders
- the external tools installed in
~/dive/plugins, each of which still has a switch per mode
| Where | What |
|---|---|
~/Library/Application Support/dive-desktop/runtime |
macOS: the copy of the server code the app runs, rebuilt on every launch |
~/Library/Application Support/dive-desktop/launchagent.out.log and launchagent.err.log
|
macOS: output of the background server |
~/Library/LaunchAgents/com.dive.server.plist |
macOS: the LaunchAgent that keeps the background server running |
~/models |
llama.cpp models, unless you change MODELS FOLDER |
~/.agents/skills |
Skills Dive shares with pi and other agents. Dive reads it but doesn't own it |
~/.pi/agent |
pi's own settings, sign-ins, sessions, skills and packages |
| The window's local storage | Each mode's MCP server list (mcpConfigByMode), light or dark theme (ollama-pi-chat-dark-mode), and a copy of the enabled modes |
Copying ~/dive copies everything Dive needs: conversations, settings, notes, skills, external tools and the library. Two things to know:
-
cloud-settings.jsoncontains your API keys in plain text, so treat the backup like a password file. - The library index can be large.
To restore, quit Dive first. On macOS the background server keeps running after the window closes, so stop it too:
launchctl bootout gui/$(id -u) ~/Library/LaunchAgents/com.dive.server.plistCopy the folder back, then open Dive, which starts the server again.
MCP server lists live in the window's local storage, not in ~/dive, so copy their JSON from Settings if you want to keep them.
-
data-dir.jsdefinesDATA_DIR,PLUGINS_DIRandWORKSPACE_DIR, andisOverriddentells you whetherDIVE_DATA_DIRis set. - Most file names are constants next to the code that uses them:
server.js(conversations, attachments, security events, local-model settings),mode-state.js(the per-mode files),server/cloud.js,routes/llamacpp.js,routes/library.js,library/store.js,agent-skills.jsandskills/sandbox.js. -
docs/data-directory.mdcovers the same ground for contributors.
These pages describe Dive 5.0.8. Source code: DuckTapeKiller/dive. Spotted a mistake? Open an issue.
Get started
Tools and skills
Build your own
Share and stay safe
How Dive works
Help