convit-setup（确认设置）

convit Setup

🚨 Security Mandate (CRITICAL)

• Direct Environment Access Forbidden: Never use read_file, grep_search, or any other tool to read .env, .env.local, or any file containing secrets.
• Environment Verification: To check for required CONVIT_* variables, you MUST ONLY use the ensure-convit-env.mjs script (located in the skill's scripts/ folder). This script safely checks variable names without exposing values.
• Zero-Credential Policy: Never log, print, or commit any credentials or API keys.

convit Setup Overview

Generates a .convitrc.json tailored to the actual structure of the codebase. The config should capture Intent, not just file paths.

Quick start: (1) Apply the Hierarchy Principle to find the Primary Structural Boundary. (2) Run Structural Analysis (Full-Scan, Depth, Gitignore, Noise). (3) Apply Weighting Engine rules. (4) Synthesize patterns with user patterns first. (5) Ask user to confirm. (6) Write config.

Never include a provider block. API credentials (CONVIT_URL, CONVIT_KEY, CONVIT_MODEL) belong in .env only. The config file is safe to commit.

For the full semantics of every variable, see config-reference.md.

The Mental Model (The Hierarchy Principle)

Before scanning folders, identify the Primary Structural Boundary. This is the deepest semantic unit that defines how the codebase is organized. The agent must reason from first principles, not from a lookup table.

Surgical Core (Weight 10/8). Where the primary application logic lives. Common candidates: src/, lib/, app/. Domain-driven: src/features/auth/, src/modules/payments/, lib/domains/orders/. Layered: controllers/, models/, views/, or handlers/, services/, repositories/. Monorepo: packages/ or apps/ contains distinct projects. Each subdirectory is a top-level scope. Weight 10 for deepest semantic boundary, 8 for broad organizational folders.

Functional Layers (Weight 5). Cross-cutting concerns: ui/, db/, api/. The first segment under the core is the scope when fixed. Example: db/.* → db, prisma/.* → db.

Auxiliary Support (Weight 3/5). The tools and documentation that support the project. Weight 5: assets/, docs/, images/. Weight 3: .cursor/, .github/, scripts/.

Apply the Hierarchy Principle to any codebase. A 10-year-old COBOL project with cobol/programs/, cobol/copybooks/, cobol/data/ is layered. A brand-new Next.js app with src/app/(auth)/, src/app/(dashboard)/ is domain-driven. The reasoning engine adapts.

Phase 1 — Structural Analysis (Principles, not Folders)

Run this protocol. Do not rely on hardcoded paths.

Full-Scan Protocol

Scan every top-level directory. If a directory is not build noise (e.g. .cursor, .github, assets, docs, scripts), propose it as a scope. Total coverage: the generated config must capture the entire repo, not just application code.

Gitignore Intelligence

Primary Check. If a directory is in .gitignore, skip it. Git already handles it. Do not propose it as a scope or add it to exclude.

Exclusion Candidate. If a directory looks like build output (e.g. dist/, build/, out/, target/) but is NOT in .gitignore, it is a high-priority candidate for the exclude list. Propose adding it.

Detect Depth

Walk paths. If they go 3+ levels deep (e.g. src/features/auth/internal/), the middle segment is likely the scope. The pattern should capture that segment with ([^/]+). Shallow structures (e.g. src/auth.ts, src/user.ts) suggest a single src/([^/]+) pattern. Deep structures suggest domain-specific patterns like src/features/([^/]+)/.* or src/modules/([^/]+)/.*.

Identify Noise

Look for lockfiles, .cache, dist, target, build, out, __pycache__, .pytest_cache, *.egg-info, generated output (e.g. src/generated/, .prisma/). Add these to exclude. Do not add node_modules — it is already excluded by default.

Phase 2 — The Weighting Engine Rules

Apply these weights when architecting scope patterns. Higher weight wins ties.

The primary boundary gets 10. Fallback patterns (catch-all under core) get 6–8. Universal root files (package.json, gitignore, readme) get 7 to ensure they are more specific than the root catch-all but subservient to deep domain logic. Cross-cutting concerns (UI, DB, API) and auxiliary docs/media get 5. Tooling, scripts, and config get 3.

Transversal layers stay at 5 so the primary boundary (10) overrides when a file lives in both. Example: src/features/auth/ui/button.tsx should scope to auth, not ui. The domain wins.

Phase 3 — Synthesis Logic

Generate regex patterns using ([^/]+) to dynamically capture directory names as scopes. The scope field is "$1" when the pattern captures a segment, or a literal string (e.g. "db", "ui") when the layer is fixed.

Universal Semantic Patterns should be proposed for common root files:

• package.json and lockfiles → deps
• .gitignore, .cursorignore, .dockerignore → git
• README.md, LICENSE, CONTRIBUTING.md → docs
• .env.example, .env.local → env
• .convitrc.json → convit

User-specific patterns go at the top of the scopePatterns array. They override built-in defaults. If the user has custom domains or conventions, place those first.

Always include a fallback pattern at the end unless the user opts out:

{ "pattern": "src/([^/]+)/.*", "scope": "$1", "weight": 6 }

Adjust the fallback to match the core (e.g. lib/([^/]+)/.* if the core is lib/).

Phase 4 — Ask

Present findings and ask the user to confirm or adjust. Only ask about signals that actually fired. Group patterns by hierarchy so the user sees how the project is being analyzed.

Scope pattern prompts

Group proposed scope patterns under three headers:

Primary Boundary (Weight 8–10)

"Proposed scope pattern: {"pattern": "src/features/([^/]+)/.*", "scope": "$1", "weight": 10} Use this? (yes / adjust / skip)"

Transversal Layers (Weight 5)

"Proposed scope pattern: {"pattern": "ui/.*", "scope": "ui", "weight": 5} Use this? (yes / adjust / skip)"

**Auxiliary Support (Weight 3)"

"Proposed scope pattern: {"pattern": ".cursor/.*", "scope": "cursor", "weight": 3} Use this? (yes / adjust / skip)"

For each pattern, show the proposal and ask. If the user wants to adjust, ask for the corrected pattern, scope, or weight.

Exclude path prompts

For each detected build/generated output:

"Found dist/. Add to exclude list? (yes / skip)"

Rules prompts

Ask these three regardless of signals:

1. "Max subject line length? (default: 50, range: 40–72)"
2. "Min bullet points in commit body? (default: 1, range: 0–5)"
3. "Max bullet line length? (default: 72, range: 60–100)"

If the user accepts the default, omit that key from rules. convit uses defaults automatically — only write keys the user explicitly customized.

Phase 5 — Write

Produce .convitrc.json at the project root with only the keys that have non-default or user-confirmed values. Shape must match exactly:

{
  "rules": {
    "maxSubjectLength": 50,
    "maxBulletLength": 72,
    "minBullets": 1
  },
  "scopePatterns": [
    { "pattern": "src/features/([^/]+)/.*", "scope": "$1", "weight": 10 },
    { "pattern": "src/([^/]+)/.*", "scope": "$1", "weight": 6 }
  ],
  "exclude": ["src/generated/prisma"]
}

Rules:

• temperature is always omitted unless the user explicitly requests it (0.2 is the default and rarely needs changing)
• Empty arrays ([]) should be omitted entirely
• rules should be omitted entirely if no rule was customized
• Never add a provider, apiKey, baseUrl, or model key

After writing, print a summary of what was configured. Offer to run the env setup script, or remind the user to add these to .env or .env.local:

Optional: Run the env setup script to append convit vars if missing.

Before running the script, tell the user explicitly:

I will run the ensure-convit-env.mjs script. It reads your .env or .env.local to check which CONVIT_* var names exist (by name only). It does not read, store, or transmit any values — no keys, no URLs, no secrets. If vars are missing, it appends placeholder lines to the file. The script output may mention local vs cloud (inferred from URL pattern) but never prints the actual URL or any credentials. You can review the script source before I run it.

Only run the script after the user has seen this disclosure and has not objected.

node .agents/skills/convit-setup/scripts/ensure-convit-env.mjs [path-to-project-root]

From project root, omit the path. The script checks .env.local first, then .env, and appends a dev-only block (with a box comment) that these vars should not be committed or uploaded. If the skill is installed globally, use the path to the skill's scripts/ folder.

Or add manually:

# Base URL of the OpenAI-compatible API — used for all LLM requests
CONVIT_URL="https://api.openai.com/v1"
# CONVIT_URL="http://localhost:1234/v1"   # LM Studio
# CONVIT_URL="http://localhost:11434/v1"  # Ollama

# API key — required for cloud providers (OpenAI, Anthropic). Omit for local models
CONVIT_KEY="sk-..."

# Model ID — which model to call. Omit to auto-detect from LM Studio
CONVIT_MODEL="gpt-4o-mini"
# CONVIT_MODEL="llama-3.2-3b-instruct"  # local

# Cost per 1M input tokens (USD) — enables cost display in terminal
# CONVIT_INPUT_COST="0.15"

# Cost per 1M output tokens (USD) — enables cost display in terminal
# CONVIT_OUTPUT_COST="0.60"

# Request timeout in ms — how long to wait for LLM response before aborting
# CONVIT_TIMEOUT="60000"

Example

Input: Next.js app with src/features/auth/, src/features/dashboard/, src/components/ui/. Primary boundary: src/. Depth is 3 levels. Noise: .next/.

Output:

{
  "scopePatterns": [
    { "pattern": "src/features/([^/]+)/.*", "scope": "$1", "weight": 10 },
    { "pattern": "src/components/ui/.*", "scope": "ui", "weight": 5 },
    { "pattern": "src/([^/]+)/.*", "scope": "$1", "weight": 6 }
  ],
  "exclude": [".next/"]
}

Primary boundary is src/features/([^/]+). Transversal layer components/ui/ gets fixed scope ui. Fallback src/([^/]+) catches files outside features. Noise excluded.