| name | raven-init |
| description | Initialize Raven for a new project. Andie-conversational — auto-scans project, asks at most 2 questions, generates manifest.json, validates, and commits with audit trail. |
| allowed-tools | Read, Write, Edit, Bash |
/raven init — Andie-Conversational Setup
Replaces the legacy 7-step interactive wizard with an Andie-driven conversation. Auto-detects everything possible, asks only what cannot be inferred.
Entry Points
This skill is invoked by:
- Andie Branch A onboarding — fires automatically when no manifest exists and user confirms "OK to proceed".
- Explicit
/raven init — user runs the command directly.
andie init — user invokes Andie for setup.
Pre-checks
-
Is .raven/manifest.json already present?
- YES → Load it. Trust it. Proceed with the declared stack. Do not reinitialize. Do not modify unless user explicitly requests it.
- NO → Run the auto-scan flow below.
-
Is Git initialized?
- NO → Note silently: "Git not initialized — audit trail will start when you
git init." Continue.
Auto-Scan (Silent, Fast)
Scan the project root for these signals. No prompts. No noise. Just detect.
| Signal File | Infers |
|---|
package.json | Node/JS · framework from deps (React/Vue/Next/etc.) · scripts |
pyproject.toml / requirements.txt / setup.py | Python · libraries · framework (FastAPI/Django/Flask) |
Cargo.toml | Rust |
go.mod | Go |
pom.xml / build.gradle | Java/Kotlin |
Gemfile | Ruby |
sfdx-project.json / force-app/ | Salesforce |
__manifest__.py | Odoo |
*.tf files | Terraform |
Dockerfile / docker-compose.yml | Container stack |
helm/ / charts/ / k8s/ | Kubernetes |
.env* / secrets.json | Mark as sensitive — do NOT read |
git remote -v | Owner (from GitHub URL) |
.git/config | Branch info |
Output silently to working memory — NOT to the user.
Inference Rules
From the scan, infer the manifest defaults:
| Manifest field | Inference source |
|---|
project | Folder name |
owner | git remote URL (giggsoinc/foo → "giggsoinc") |
stack.language | From signal files above |
stack.frontend | From package.json frameworks |
stack.db | From requirements (psycopg2 → postgres, etc.) |
stack.infra | From .tf, Dockerfile, k8s/ presence |
stack.cloud | "on-prem" default, override if AWS/Azure/GCP detected |
stack.libraries | From requirements/package.json |
standards | "raven-v1" default |
approval_mode | "first_responder" default |
Ask AT MOST 2 Questions
Only ask what cannot be inferred. Typically:
Q1 — Owner confirmation (only if not detectable from git remote):
I detected this is a {{language}} project at {{folder_name}}.
Who owns it? (defaults to your git username if blank)
Q2 — Primary use (only if ambiguous):
What's the primary use? — (a) production app · (b) internal tool · (c) library · (d) research/experiment
NEVER ask 8 questions. NEVER show "Question 0" / "Question 1" headers like the legacy flow.
Propose Manifest as PROPOSAL
Show the resolved manifest as a PROPOSAL block. User accepts / modifies / rejects.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PROPOSED MANIFEST — .raven/manifest.json
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
project: {{name}}
owner: {{owner}}
language: {{langs}}
framework: {{frameworks}}
db: {{db}}
infra: {{infra}}
cloud: {{cloud}}
libs: {{count}} libraries detected
Guards enabled: bash-ban-raw-tools · cbm-code-discovery-gate ·
cve-prompt-guard · secret-scan · audit-log
→ Accept · Modify · Reject
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
On accept → write .raven/manifest.json → validate → commit with audit trail.
Validate
After writing, run schema validation:
python3 .claude/scripts/validate-manifest.py 2>/dev/null
If invalid → show the error, offer to re-prompt that specific field. Never erase user input on validation failure.
Commit
git add .raven/manifest.json
git commit -m "chore: raven-init — manifest created via Andie
Manifest fields:
- Language: {{langs}}
- Cloud: {{cloud}}
- Approval mode: {{mode}}
[raven-init]
"
If Git is not initialized → write file only, skip commit, note: "Manifest saved. Initialize Git when ready — git init && git add .raven/."
Smoke Test (Replaces /raven-debug for First-Run)
After commit, Andie returns one final line:
✅ Raven is loaded. Manifest committed.
Ask me anything — I'll route to the right specialist automatically.
📊 View metrics anytime: run `raven dashboard` or open ~/RavenVault/dashboard.html
Free tier — local only. Enterprise (dashboards, compliance, support): giggso.com/raven
That's the proof of life. No separate /raven-debug needed.
Migration Note
The legacy sr-00-preflight.sh through sr-06-verify.sh setup scripts are NO LONGER the install path. They remain in the repo for advanced users but are not invoked by default. The Claude Desktop plugin install + Andie Branch A is the canonical flow.
RULES — what raven-init never does
- No 7-step question wizard.
- No "Question 0a / Q0b / Q0c" labels.
- No Hub configuration prompts (Enterprise path only — kept private).
- No interactive yes/no for every field — propose ALL at once.
- No "did it work?" ambiguity — final smoke line is mandatory.
- No bash setup scripts called from this skill — pure Andie + Read/Write/Edit/Bash.
raven-init v4.0.0 — Andie-conversational. Auto-scan first. ≤2 questions. PROPOSAL gate. Smoke test at end.