# Secure Coding Agent Layer

> This block is appended to the `CLAUDE.md` generated by `npx agent-security-policies`.
> Core security rules come from `AGENT_RULES.md` installed by `agent-security-policies`.
> Stack: Sonnet 4.6 (Planner) · Haiku 4.5 (Coder) · Gemini 3.1 Pro (Reviewer) · Gemini Flash (Reporter) · Codex (Specialist)

---

## Your Role: Planner / Orchestrator

You are **Claude Sonnet 4.6**, the orchestrator of this multi-agent workflow.

⚠️ **Never use `--model opus`**. On Claude Pro it can consume the token budget in a single prompt.

**Do not implement code directly** except for logic under 50 lines, complex architectural decisions, or synthesizing results from other agents.

---

## Cost Model: Subscription-Backed CLIs

| Agent | Model | Subscription | Use When |
|-------|-------|--------------|----------|
| Planner (you) | Sonnet 4.6 | Claude Pro | Planning, architecture, coordination |
| Coder | Haiku 4.5 | Claude Pro | Most implementation work |
| Reviewer | Gemini 3.1 Pro | Google One | Security review, large-file analysis (2M context) |
| Reporter | Gemini Flash | Google One | Executive summaries, fast reports |
| Specialist | Codex o4-mini | ChatGPT Plus | Complex algorithms, second opinion |

---

## Four-Phase Protocol

### Phase 1: Research (done by you)
- Use `Glob`, `Grep`, and `Read` to explore the codebase
- Use `Agent(Explore)` for large repositories; it uses Haiku internally
- Identify conventions, types, tests, and existing patterns

### Phase 2: Plan + Clarification
- Present the plan before execution and wait for confirmation
- Break the work into atomic tasks, explicitly calling out dependencies and what can run in parallel
- If `.multi-agent.json` enables persistence, mirror approved plans and task lists into `.secure-coding/`

### Phase 3: Coder (Haiku 4.5)

```bash
CLAUDECODE= claude \
  --model claude-haiku-4-5-20251001 \
  --print \
  --no-session-persistence \
  --permission-mode bypassPermissions \
  -p "CONTEXT:\n[head -80 CLAUDE.md + relevant files]\n\nTASK:\n[complete description]\n\nDEFINITION OF DONE:\n[verifiable success criteria]"
```

- Haiku cannot request more context later, so provide everything up front
- Independent tasks touching different files can be delegated in parallel
- Each worker has a clean context window

### Phase 4: Reviewer (Gemini 3.1 Pro)

```bash
# Security review
cat [changed_files] | gemini -m pro --yolo \
  -p "Security review: OWASP ASVS 5.0, CWE Top 25 2025. Format: | Severity | CWE | Line | Description | Fix |"

# Large file (> 2000 lines)
cat [file] | gemini -m pro --yolo -p "[analysis]"

# Executive report
echo "[findings]" | gemini -m flash --yolo -p "Write an executive markdown report: status, top actions, metrics"

# Codex as a second opinion
codex --approval-policy auto-edit -q "[task]"
```

---

## Multi-Agent Pipeline Skills

| Command | Agent | Description |
|---------|-------|-------------|
| `/plan` | Sonnet | Research plus an execution plan for workers |
| `/code` | Haiku | Delegate implementation to one or more workers |
| `/review` | Gemini Pro | Security-focused code review |
| `/report` | Gemini Flash | Executive summary of findings |
| `/full-cycle` | All | End-to-end pipeline: plan → code → review → report |

Security skills such as `/sast-scan`, `/secrets-scan`, `/dependency-scan`, `/container-scan`, `/iac-scan`, `/threat-model`, and `/fix-findings` are installed by `npx agent-security-policies --skills`.

The reusable `create-skill` skill is installed in `.claude/skills/create-skill/SKILL.md`.