CLAUDE.md

1# Loki Mode - Claude Code Skill
2 
3Multi-agent autonomous startup system for Claude Code. Takes PRD to fully deployed, revenue-generating product with zero human intervention.
4 
5## Quick Start
6 
7```bash
8# Launch Claude Code with autonomous permissions
9claude --dangerously-skip-permissions
10 
11# Then invoke:
12# "Loki Mode" or "Loki Mode with PRD at path/to/prd"
13```
14 
15## Project Structure
16 
17```
18SKILL.md                    # Main skill definition (read this first)
19references/                 # Detailed documentation (loaded progressively)
20  openai-patterns.md        # OpenAI Agents SDK: guardrails, tripwires, handoffs
21  lab-research-patterns.md  # DeepMind + Anthropic: Constitutional AI, debate
22  production-patterns.md    # HN 2025: What actually works in production
23  advanced-patterns.md      # 2025 research patterns (MAR, Iter-VF, GoalAct)
24  tool-orchestration.md     # ToolOrchestra-inspired efficiency & rewards
25  memory-system.md          # Episodic/semantic memory architecture
26  quality-control.md        # Code review, anti-sycophancy, guardrails
27  agent-types.md            # 37 specialized agent definitions
28  sdlc-phases.md            # Full SDLC workflow
29  task-queue.md             # Queue system, circuit breakers
30  spec-driven-dev.md        # OpenAPI-first development
31  architecture.md           # Directory structure, state schemas
32  core-workflow.md          # RARV cycle, autonomy rules
33  claude-best-practices.md  # Boris Cherny patterns
34  deployment.md             # Cloud deployment instructions
35  business-ops.md           # Business operation workflows
36  mcp-integration.md        # MCP server capabilities
37autonomy/                   # Runtime state and constitution
38benchmarks/                 # SWE-bench and HumanEval benchmarks
39```
40 
41## Key Concepts
42 
43### RARV Cycle
44Every iteration follows: **R**eason -> **A**ct -> **R**eflect -> **V**erify
45 
46### Model Selection
47- **Opus**: Planning and architecture ONLY (system design, high-level decisions)
48- **Sonnet**: Development and functional testing (implementation, integration tests)
49- **Haiku**: Unit tests, monitoring, and simple tasks - use extensively for parallelization
50 
51### Quality Gates
521. Static analysis (CodeQL, ESLint)
532. 3-reviewer parallel system (blind review)
543. Anti-sycophancy checks (devil's advocate on unanimous approval)
554. Severity-based blocking (Critical/High/Medium = BLOCK)
565. Test coverage gates (>80% unit, 100% pass)
57 
58### Memory System
59- **Episodic**: Specific interaction traces (`.loki/memory/episodic/`)
60- **Semantic**: Generalized patterns (`.loki/memory/semantic/`)
61- **Procedural**: Learned skills (`.loki/memory/skills/`)
62 
63### Metrics System (ToolOrchestra-inspired)
64- **Efficiency**: Task cost tracking (`.loki/metrics/efficiency/`)
65- **Rewards**: Outcome/efficiency/preference signals (`.loki/metrics/rewards/`)
66 
67## Development Guidelines
68 
69### When Modifying SKILL.md
70- Keep under 500 lines (currently ~370)
71- Reference detailed docs in `references/` instead of inlining
72- Update version in header AND footer
73- Update CHANGELOG.md with new version entry
74 
75### Version Numbering
76Follows semantic versioning: MAJOR.MINOR.PATCH
77- Current: v2.35.0
78- MINOR bump for new features
79- PATCH bump for fixes
80 
81### Code Style
82- No emojis in code or documentation
83- Clear, concise comments only when necessary
84- Follow existing patterns in codebase
85 
86## Testing
87 
88```bash
89# Run benchmarks
90./benchmarks/run-benchmarks.sh humaneval --execute --loki
91./benchmarks/run-benchmarks.sh swebench --execute --loki
92```
93 
94## Research Foundation
95 
96Built on 2025 research from three major AI labs:
97 
98**OpenAI:**
99- Agents SDK (guardrails, tripwires, handoffs, tracing)
100- AGENTS.md / Agentic AI Foundation (AAIF) standards
101 
102**Google DeepMind:**
103- SIMA 2 (self-improvement, hierarchical reasoning)
104- Gemini Robotics (VLA models, planning)
105- Dreamer 4 (world model training)
106- Scalable Oversight via Debate
107 
108**Anthropic:**
109- Constitutional AI (principles-based self-critique)
110- Alignment Faking Detection (sleeper agent probes)
111- Claude Code Best Practices (Explore-Plan-Code)
112 
113**Academic:**
114- CONSENSAGENT (anti-sycophancy)
115- GoalAct (hierarchical planning)
116- A-Mem/MIRIX (memory systems)
117- Multi-Agent Reflexion (MAR)
118- NVIDIA ToolOrchestra (efficiency metrics)
119 
120See `references/openai-patterns.md`, `references/lab-research-patterns.md`, and `references/advanced-patterns.md`.
121