> ## Documentation Index
> Fetch the complete documentation index at: https://docs.commanderplugin.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Intelligence Layer: How CC Commander Thinks Before It Acts
> How CC Commander scores task complexity, detects your stack, recommends skills, and compounds session knowledge automatically — zero configuration needed.
Before dispatching a single token to Claude Code, CC Commander runs four analysis modules that together decide the right model, budget, turn limit, and skills for your task. You never touch a config flag — the Intelligence Layer handles every dispatch automatically based on what your task actually needs.
The Intelligence Layer runs automatically — no configuration required.
## The four modules
### 1. Weighted complexity scoring
Every task you describe is scored from 0 to 100 using 47 keyword signals, word count, and fuzzy regex matching. The score maps directly to a model, turn limit, and dollar budget.
```
"fix typo" → score 0 → 10 turns, $1 budget, Haiku
"add dark mode" → score 25 → 20 turns, $3 budget, Sonnet
"refactor auth module" → score 60 → 35 turns, $6 budget, Sonnet
"build SaaS platform" → score 100 → 50 turns, $10 budget, Opus
```
The scorer starts at a baseline of 30 and adjusts up or down based on what it finds in your task description:
| Signal type | Example | Effect |
| ------------------------------ | -------------------------------------------- | -------------- |
| Trivial reducers | `fix typo`, `rename`, `one line` | −15 to −30 pts |
| Moderate adders | `add feature`, `refactor`, `integrate` | +10 to +20 pts |
| High-complexity adders | `build saas`, `from scratch`, `multi-tenant` | +25 to +40 pts |
| Long description (25–50 words) | Detailed task prompt | +10 pts |
| Long description (50+ words) | Very detailed prompt | +20 pts |
| File scope bonus | Matches many project files | 0 to +20 pts |
The four score ranges and what they produce:
| Score range | Label | Turns | Budget | Effort |
| ----------- | -------- | ----- | ------ | ------ |
| 0–25 | Trivial | 10 | \$1 | low |
| 26–50 | Simple | 20 | \$3 | low |
| 51–75 | Moderate | 35 | \$6 | medium |
| 76–100 | Complex | 50 | \$10 | high |
File scope estimation adds a 0–20 point bonus by scanning how many project files the task is likely to touch. You never need to set `--model`, `--budget`, or `--max-turns` manually.
### 2. Stack detection
Before every dispatch, CCC reads your project directory and detects your technology stack automatically. It parses `package.json`, `Dockerfile`, `go.mod`, and `requirements.txt`. It also reads your current git branch and the themes of your last five commits, and detects monorepo setups (workspaces, lerna, turbo, nx).
From those files it recognizes: nextjs, react, vue, docker, python, rust, go, github-actions, orm, billing, and testing frameworks. No configuration needed — CCC reads what is already in your project.
### 3. Skill recommendations
Once the stack and task are known, CCC's skill recommender combines three signals to surface the most relevant skills:
| Signal | Weight | What it does |
| ------------- | ---------------- | ---------------------------------------------------- |
| Stack match | 10 pts | Next.js project → `nextjs-app-router` ranks first |
| Keyword match | 2 pts per hit | "auth" task → `auth`, `jwt`, `better-auth` bubble up |
| Usage history | Boost multiplier | Skills that succeeded for you rank higher over time |
Trending skills (7-day window) surface automatically. Skills that led to successful sessions compound their ranking advantage the more you use them.
### 4. Knowledge compounding
Every completed session extracts a lesson — keywords, category, tech stack, error patterns, success patterns — and stores it in `~/.claude/commander/knowledge/`. The next time you dispatch a similar task, those lessons are injected into the session prompt automatically.
Relevance is weighted by time:
| Age | Multiplier |
| ------------------ | ---------- |
| Less than 7 days | 2× |
| 7–30 days | 1.5× |
| Older than 30 days | 1× |
Fuzzy keyword matching and cross-domain boosts (web↔react, api↔backend, testing↔bugfix) catch related concepts across categories. Outcome weighting gives successful sessions a 1.5× boost and error sessions a 1.2× boost, so the knowledge base gets more useful the more you build.
## Smart dispatch in action
Here is what CCC's analysis output looks like before a session starts:
```
Task: "Add authentication with JWT and OAuth"
Intelligence Analysis:
├─ Complexity Score: 72/100 (complex)
├─ Keyword signals: +15 (auth) +15 (implement) +20 (production)
├─ Word count: +10 (detailed description)
├─ Stack detected: nextjs, react, tailwind
└─ Related lessons: 2 found (JWT auth succeeded, OAuth2 pattern)
Auto-configured dispatch:
Model: opus | Turns: 35 | Budget: $6
Skills: nextjs-app-router, ccc-saas, auth-patterns
Knowledge: 2 past lessons injected
Dispatching... ████████████████████░░░░ 78%
```
## How it compounds over sessions
| Session | What changed |
| ------- | ------------------------------------------------------------------- |
| 1 | Dispatches based on complexity score alone |
| 5 | Knows your stack, recommends proven skills |
| 20 | Has learned your patterns — feels like a PM who knows your codebase |
## Smart retry
The Intelligence Layer also handles failures automatically so you do not need to babysit long sessions:
* **Rate limit hit** — waits 60 seconds, then retries automatically
* **Context overflow** — reduces turns to 60% of original, then retries
* **Budget exceeded** — surfaces a clear error with next steps
You can view your accumulated knowledge base stats at any time by running `ccc --stats`.