Setup Module Specification¶

Status: Draft
Fixes: Plan review issue #15 (reversibility) + issue #13 (install delivery)
Related: PRD · ADR-0006

Install delivery¶

Phase	Method	Notes
Phase A	`./scripts/install.sh` in repo	Checks Python 3.12, venv, Ollama, pulls default model
Phase A.1	`daari install` CLI command	Same logic as shell script in a clone; falls back to `daari onboard` after pip/brew
Phase A.1	`daari onboard --yes`	Pure-Python first-run: probe Ollama, pull L3 + embed, run doctor
Demo	`./scripts/demo.sh`	One-command smoke: serve + curl + stats + setup dry-run
Future	`brew install daari` / curl pipe	Requires domain + release artifacts — not MVP

Removed from MVP claims: curl https://daari.dev/install — domain does not exist yet.

Setup commands¶

Command	Phase	Behavior
`daari setup`	A.1	Interactive wizard — pick clients, models, run doctor (see below)
`daari setup cursor`	A.1	Patch Cursor custom model settings
`daari setup models`	A.1	Pick / map Ollama models to L3 tier (interactive or flags)
`daari setup claude-code`	B	Requires Anthropic-compat gateway
`daari setup openai-compat`	B	Print env vars for generic SDK
`daari setup intellij`	B.1	Register IntelliJ path for Lt tier (not AI client)
`daari setup --all`	B	Detect + run applicable recipes (non-interactive)
`daari setup --undo <tool>`	A.1	Restore latest backup for tool
`daari doctor`	A.1	Health check
`daari onboard`	A.1	pip/brew first-run (no git clone)
`daari service`	A.1	user systemd / launchd stay-up (no root)

Interactive wizard (`daari setup`)¶

Yes — users need a guided path. Not everyone knows what to configure or that IntelliJ is an Lt backend, not an AI client.

Two entry modes (same engine):

Mode	When
`daari setup`	First install, exploratory — prompts for choices
`daari setup <tool>`	Scriptable, docs, CI — skip prompts

Wizard flow (Phase A.1)¶

$ daari setup

  daari setup — configure your local stack

  Detected on this machine:
    ✓ Ollama (3 models: llama3.2:3b, llama3.1:8b, …)
    ✓ Cursor
    ✗ IntelliJ IDEA (not found)
    ✗ Claude Code (not supported in this version)

  What do you want to set up?  [Space to select, Enter to confirm]

  ❯ Cursor          — point AI chat at daari (OpenAI-compat)
    Local model     — choose default model for L3 tier
    Run health check — daari doctor after setup
    Skip for now

  → User picks Cursor

  Configure Cursor to use http://127.0.0.1:11435/v1 ?  [Y/n]
  → Backup + patch settings
  → Done. Run: daari serve

  Optional: configure local model now?  [y/N]
  → Lists `ollama list`, user picks default for L3
  → Writes ~/.daari/config.yaml

Phase B+ wizard additions¶

Step	Command / screen	Notes
IntelliJ (Lt)	`daari setup intellij`	Explain: "Registers IDE for refactor/lint — not your AI chat client"
OpenAI SDK	`daari setup openai-compat`	Print `export OPENAI_BASE_URL=…`
Claude Code	`daari setup claude-code`	Only when Anthropic gateway ships (C2)
All detected	`daari setup --all`	Non-interactive batch for power users

Model selection (`daari setup models`)¶

daari setup models              # interactive: ollama list → pick L3 default
daari setup models --tier l3 --model llama3.2:3b   # non-interactive
daari setup models --list       # show current tier → model map

Pulls from ollama list (live) — not hard-coded
Phase A.1: L3 only; Phase B adds L4/L5 mapping
If Ollama missing: wizard offers ollama pull llama3.2:3b hint

Help & discoverability¶

daari setup --help              # list subcommands
daari setup cursor --dry-run    # show diff, no writes
daari doctor                    # post-setup: daemon, ollama, cursor endpoint, sample route

Phase A (tracer bullet): Manual only — cursor.md. No wizard yet.

Phase A.1: Ship daari setup wizard + daari setup cursor + daari setup models + daari doctor.

Reversibility (required)¶

Every daari setup <tool> must:

Detect if tool is installed; exit with message if not
Backup affected config files to ~/.daari/backups/<tool>/<ISO8601>/
Apply changes (patch or write)
Print summary: files changed, backup path, undo command
Support dry-run: daari setup cursor --dry-run shows diff without writing

Undo¶

daari setup --undo cursor
# Restores most recent backup for cursor; restarts nothing

Keep last 5 backups per tool; prune older
Backup manifest: manifest.json listing original paths

Cursor recipe (Phase A.1)¶

Code: daari/clients/cursor/recipe.py
Manual doc: cursor.md

Changes (automated by daari setup cursor): - Backs up settings.json and state.vscdb under ~/.daari/backups/cursor/<timestamp>/ - Patches settings.json: openai.baseUrl, openai.apiKey, and daari.setup.cursor marker - Patches Cursor state.vscdb application storage: openAIBaseUrl, useOpenAIKey, availableAPIKeyModels - Base URL: http://127.0.0.1:11435/v1 (from user config) - API key: daari-local (or match DAARI_API_KEY in future) - Model name: daari

Known limitation: Cursor may store the OpenAI API key in OS secure storage on some versions. If chat still fails after setup, enter daari-local once in Cursor Settings → Models → OpenAI API Key. Restart Cursor after apply.

Manual fallback (Phase A): Document in docs/setup/cursor.md.

Idempotency¶

Re-running daari setup cursor when already configured: - Detects existing daari endpoint - No-op with message, or --force to re-apply

Claude Code (Phase B — honest limitation)¶

Claude Code uses Anthropic API shape. Setup deferred until: - Anthropic-compat gateway exists (Phase C1), OR - Claude Code adds OpenAI-compat mode

Document in setup output: "Claude Code not supported in this version."