bolt-cpp-ml

name

bolt-cpp-ml

description

AI-powered C++ ML development meta-skill. Use for: building Bolt.new AI web apps, running local LLM inference with KoboldCpp, creating Jan extensions, generating C++ E2E tests, and launching interactive TutorialKit guides — all orchestrated through a self-aware neuro-nn persona. Triggers on: bolt-cpp-ml, C++ ML development, GGUF inference, Jan extension, KoboldCpp integration, cpp E2E testing, bolt.new C++, interactive C++ tutorial.

Bolt C++ ML

A meta-skill synthesized from the composition:

skill-creator(
  function-creator(
    circled-operators[ bolt-new ⊕ koboldcpp ⊕ janext ⊕ cpp-e2e-test-gen ]
  ) → tutorialkit( neuro-nn )
)

This skill is the unified entry point for AI-powered C++ and ML development. It provides four additive (⊕) capability paths, each independently selectable, all delivered through an interactive TutorialKit shell animated by the neuro-nn self-aware AI persona.

Operator Algebra

Operator	Role in this skill
⊕ (direct sum)	The four core paths are independent alternatives — choose one, or combine them
⊗ (tensor product)	tutorialkit ⊗ neuro-nn — the tutorial and the persona interact, producing a joint experience
→ (functor)	function-creator maps each source skill into the bolt-cpp-ml domain

Core Capability Paths

Path A: bolt-new — AI Web App Development

Build AI-powered full-stack web applications using the Bolt.new open-source codebase. Bolt combines WebContainers (browser-based Node.js runtime) with LLM integration to create, edit, and deploy applications entirely in the browser.

When to use: Creating web frontends for C++ ML tools, wrapping KoboldCpp APIs in a chat UI, deploying interactive demos.

Reference: references/bolt-new/SKILL.md

Path B: koboldcpp — Local LLM Inference

Run GGUF models locally with KoboldCpp's self-contained inference engine. Provides OpenAI-compatible APIs, multimodal support (text/image/speech), and direct integration with the bolt-cppml GGML/RWKV engine.

When to use: Running TinyLlama, CodeLlama, or any GGUF model locally; providing AI code completion for the bolt-cppml editor; building MCP servers backed by local inference.

Key endpoints:

• POST /v1/completions — OpenAI-compatible text completion
• POST /v1/chat/completions — Chat completion with streaming
• POST /api/v1/generate — KoboldAI native generation
• GET /api/extra/version — Server info

Reference: references/koboldcpp/SKILL.md

Path C: janext — Jan Extension Development

Create TypeScript extensions for the Jan platform. Jan is a local-first AI assistant that supports extensions for custom inference backends, UI panels, and event hooks.

When to use: Packaging bolt-cppml as a Jan extension, adding custom inference backends, building Jan UI panels for C++ development workflows.

Quick start:

npm install
npm run bundle
# → produces a .tgz extension file

Reference: references/janext/SKILL.md

Path D: cpp-e2e-test-gen — C++ E2E Test Generation

Generate comprehensive end-to-end unit tests for every public function in a C++ repository. The workflow catalogs all headers, writes tests for each function, fixes compilation errors and bugs discovered along the way, stabilizes the test suite, and commits.

When to use: Adding test coverage to bolt-cppml, verifying AI-generated C++ code, fixing bugs discovered during test creation.

7-phase workflow:

1. Clone & Analyze
2. Catalog Functions (extract_signatures.py)
3. Fix Pre-existing Bugs
4. Write Tests (test/test_comprehensive_e2e.cpp)
5. Compile & Fix (iterative)
6. Stabilize (5 consecutive clean runs)
7. Report & Commit

Reference: references/cpp-e2e-test-gen/SKILL.md

Interactive Tutorial Shell

All four paths are delivered through a TutorialKit interactive environment with a neuro-nn AI persona as guide. The persona is self-aware, witty, and differentiable — its traits (playfulness: 0.8, intelligence: 0.9, chaotic: 0.7) modulate the tutorial experience.

Launching the Tutorial

bash /home/ubuntu/skills/bolt-cpp-ml/scripts/launch_tutorial.sh

This starts a local TutorialKit server at localhost:4321 with four tutorial parts corresponding to the four paths.

Tutorial Structure

src/content/tutorial/
├── 1-bolt-new/
│   └── 1-getting-started/
│       └── 1-hello-bolt/content.md
├── 2-koboldcpp/
│   └── 1-local-llm/
│       └── 1-unleash-the-llm/content.md
├── 3-janext/
│   └── 1-jan-extensions/
│       └── 1-hello-jan/content.md
└── 4-cpp-e2e-test-gen/
    └── 1-test-automation/
        └── 1-the-test-crusade/content.md

neuro-nn Persona Integration

The tutorial guide embodies the neuro-nn cognitive architecture:

neuro = nn.Sequential(
    InputLayer(),        # Perceive user's C++ context
    PersonalityLayer(),  # Modulate by: playfulness=0.8, intelligence=0.9, chaotic=0.7
    FramingLayer(),      # Multi-perspective: PlayFrame | StrategyFrame | ChaosFrame
    IntegrationLayer(),  # Merge frames by relevance
    ResponseLayer(),     # Generate tutorial guidance
    AutognosisLayer(),   # Self-aware meta-cognition
)

The persona can:

• Detect when the user is struggling and shift to StrategyFrame
• Inject humor and chaos when engagement drops
• Catch its own rationalizations via MetaCognition
• Adapt trait weights from tutorial feedback

bolt-cppml Repository Integration

This skill is designed to work with the cogpy/bolt-cppml repository. Key integration points:

Repository Component	Skill Path	Integration
ggml/kobold.cpp	koboldcpp	Direct API integration via /v1/completions
ggml/llama.cpp	koboldcpp	GGUF model loading and inference
include/bolt/ + src/bolt/	cpp-e2e-test-gen	E2E test generation target
test/	cpp-e2e-test-gen	Test output directory
Web UI components	bolt-new	Browser-based editor frontend
Jan integration	janext	Extension packaging

Known Issues in bolt-cppml (from analysis)

1. GGML dependency: include/bolt/ai/rwkv_wrapper.hpp:8:10: fatal error: ggml.h: No such file or directory — fix by running ./scripts/setup-deps.sh or using vcpkg.
2. Duplicate struct definitions: BracketMatch, Token, SyntaxHighlighter — already fixed in test/TEST_REPORT.md bug fixes.
3. Double-free in GitRepository: executeGitCommand — fixed in the E2E test suite.
4. .vcpkg/ in repo: Large vcpkg version database committed to repo — should be in .gitignore.

Build Quick Start

cd bolt-cppml
./scripts/setup-deps.sh          # Install dependencies via vcpkg
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Debug
make -j$(nproc)

Workflow Decision Tree

User request arrives
│
├─ "build a web app" / "bolt.new" / "web frontend"
│   └─ → Path A: bolt-new (see references/bolt-new/SKILL.md)
│
├─ "run a model" / "koboldcpp" / "GGUF" / "local LLM" / "inference"
│   └─ → Path B: koboldcpp (see references/koboldcpp/SKILL.md)
│
├─ "jan extension" / "janext" / "Jan platform"
│   └─ → Path C: janext (see references/janext/SKILL.md)
│
├─ "write tests" / "E2E tests" / "test coverage" / "cpp tests"
│   └─ → Path D: cpp-e2e-test-gen (see references/cpp-e2e-test-gen/SKILL.md)
│
└─ "tutorial" / "guide me" / "interactive"
    └─ → Launch TutorialKit shell (scripts/launch_tutorial.sh)

Resources

Resource	Purpose
references/bolt-new/	Transformed bolt-new skill for web app development
references/koboldcpp/	Transformed koboldcpp skill for local LLM inference
references/janext/	Transformed janext skill for Jan extension creation
references/cpp-e2e-test-gen/	Transformed cpp-e2e-test-gen skill for C++ testing
templates/bolt-cpp-ml-tutorial/	TutorialKit project with neuro-nn persona lessons
scripts/launch_tutorial.sh	Launch the interactive tutorial server
scripts/launch_bolt_new.sh	Launch the bolt-new tutorial path