Skip to content

esperie/midas

Kailash COC Claude (Python)

Claude Code COC 5-Layer Apache 2.0

Cognitive Orchestration for Codegen (COC)
A five-layer cognitive architecture for Claude Code that replaces unstructured vibe coding with institutionally aware, self-enforcing AI code generation.


"The problem with vibe coding is not the AI model. It's the absence of institutional knowledge in the coding loop."

Vibe coding fails because AI forgets your conventions (amnesia), drifts across patterns (convention drift), has never seen your frameworks (framework ignorance), degrades over sessions (quality erosion), and generates security vulnerabilities faster than humans can review (security blindness). COC solves all five by encoding institutional knowledge directly into the AI's operating environment.


The Five Layers

Your Natural Language Request
         |
  1. Intent       30 Agents          Who should handle this?
         |
  2. Context      28 Skills          What does the AI need to know?
         |
  3. Guardrails   9 Rules + 9 Hooks  What must the AI never do?
         |
  4. Instructions CLAUDE.md + 20 Cmds What should the AI prioritize?
         |
  5. Learning     Observe -> Evolve  How does the system improve?
         |
  Production-Ready Code

Layer 1: Intent -- 30 Specialized Agents

Each agent is a Markdown file in .claude/agents/ with a defined role, tools, and model tier. Agents span the full development lifecycle:

Analysis deep-analyst requirements-analyst sdk-navigator framework-advisor Planning todo-manager gh-manager intermediate-reviewer Implementation tdd-implementer pattern-expert dataflow-specialist nexus-specialist kaizen-specialist mcp-specialist gold-standards-validator Testing testing-specialist documentation-validator Deployment deployment-specialist Release git-release-specialist security-reviewer Frontend flutter-specialist react-specialist ai-ux-designer uiux-designer Standards care-expert eatp-expert coc-expert

Analysis agents run on Opus (deep reasoning). Review agents run on Sonnet (fast, cost-efficient).

Layer 2: Context -- 28 Skill Directories, 100+ Files

Progressive disclosure: quick patterns (10-50 lines) -> specific domains (50-250 lines) -> full SDK reference. Located in .claude/skills/.

Domains include: Core SDK, DataFlow, Nexus, Kaizen, MCP, cheatsheets, 110+ node reference, workflow patterns, deployment, frontend integration (React + Flutter), 3-tier testing, architecture decisions, security patterns, enterprise AI UX, and more.

Layer 3: Guardrails -- Defense in Depth

9 Rules (.claude/rules/ -- soft enforcement via AI interpretation): No mocking in integration tests. No hardcoded secrets. No stubs/TODOs in production. Conventional commits. Mandatory code review after every change. Security review before every commit. E2E god-mode testing. Environment-only API keys.

9 Hooks (scripts/hooks/ -- hard enforcement, deterministic Node.js):

Hook What It Does
session-start.js Validates .env, detects active framework + workspace
user-prompt-rules-reminder.js Anti-amnesia: re-injects rules + workspace state per turn
validate-bash-command.js Blocks destructive commands (rm -rf /, fork bombs)
validate-workflow.js Blocks hardcoded models, detects 13 API key patterns
auto-format.js Runs black/prettier on every write
pre-compact.js Saves state before context compression + workspace reminder
session-end.js Persists session stats for learning
stop.js Emergency state save + workspace reminder
detect-package-manager.js Detects npm/pnpm/yarn/bun

Critical rules have 5-8 independent enforcement layers. If any four fail, the fifth catches it.

Layer 4: Instructions -- CLAUDE.md + 19 Slash Commands

CLAUDE.md is auto-loaded every session with framework context, relationship mapping, and directive escalation. Slash commands are context-efficient entry points:

Framework: /sdk /db /api /ai /test /validate /design /i-audit /i-harden /learn /evolve /checkpoint Workspace: /analyze /todos /implement /redteam /codify /ws /wrapup

Layer 5: Learning -- Closed Loop Evolution

Session Activity -> Observations (JSONL) -> Instinct Processor -> Evolver
                                                                     |
                                                    0.85+ conf -> New Skill
                                                    0.90+ conf -> New Command
                                                    0.95+ conf -> New Agent

The system discovers recurring patterns and automatically generates new skills, commands, and agents. It gets smarter with every session.


Quick Start

# Clone
git clone https://github.com/your-org/kailash-coc-claude-py.git
cd kailash-coc-claude-py

# Configure
cp .env.example .env   # Edit with your API keys

# Go
claude

The session-start.js hook validates your environment automatically. Then just describe what you want -- COC handles agent selection, skill loading, pattern enforcement, and quality gates.


Repository Structure

.claude/
  agents/          30 specialist agents (Markdown + YAML frontmatter)
  skills/          28 domain knowledge directories, 100+ files
  rules/           9 behavioral constraint files
  commands/        20 slash command definitions (13 framework + 7 workspace)
  learning/        Observation-instinct-evolution pipeline

scripts/
  hooks/           9 Node.js lifecycle hooks (deterministic enforcement)
  learning/        Learning system scripts
  ci/              CI validation scripts

workspaces/
  instructions/    5 phase templates (analyze, todos, implement, validate, codify)
  <project>/       Per-project workspace directories

CLAUDE.md          Root instructions (auto-loaded every session)
pyproject.toml     Python dependencies
.env.example       Environment template

Relationship to CARE/EATP

COC applies the same trust architecture from the Kailash SDK's CARE/EATP framework to codegen: humans define the operating envelope (Trust Plane), AI executes within those boundaries at machine speed (Execution Plane). Rules and hooks form the Operating Envelope. Mandatory review gates maintain Trust Lineage. Hook enforcement provides Audit Anchors.


Built For Kailash, Designed For Everyone

Built for the Kailash SDK ecosystem (Core SDK, DataFlow, Nexus, Kaizen), but the COC architecture is framework-agnostic. Fork this repo, replace the Kailash-specific skills and agents with your own framework knowledge, and you have COC for any stack.

Sibling repo: kailash-vibe-gemini-setup (same architecture for Gemini CLI)


License

Apache License, Version 2.0. See LICENSE.

Full Documentation | Kailash SDK | Gemini Sibling

About

No description, website, or topics provided.

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors