pip install fitz-graveyard

fitz-graveyard plan "Add OAuth2 authentication with Google and GitHub providers"
fitz-graveyard run   # let it cook overnight
fitz-graveyard get 1 # full architectural plan in the morning

About 🧑‍🌾

Solo project by Yan Fitzner (LinkedIn, GitHub).

• ~7k lines of Python
• 400+ tests (402)
• Zero LangChain/LlamaIndex dependencies — built from scratch

The Problem

Claude Code costs $100/month to run semi-productively — and that's heavily subsidized. When subsidies shrink, prices go up. The single most expensive operation in agentic LLM coding is the planning phase: understanding a codebase, reasoning about architecture, producing a structured plan. Every token of that burns through your API budget.

What if the planning phase could run on local hardware instead? What if you could do it with a machine you already own?

The Insight 💡

Running LLMs locally means balancing three things: tokens per second, quantization quality, and model intelligence. A 70B model at high quant gives you excellent reasoning but crawls at 2-5 tok/s on consumer hardware. That feels unusable — until you realize planning doesn't need to be interactive.

Queue a job. Go to sleep. Let it run overnight.

Suddenly tok/s doesn't matter. You can run a large, intelligent model purely in RAM at 10 tok/s and that's fine.

10 tok/s × 60s × 60min × 8 hours = 288,000 tokens

That's enough for a full architectural plan — reasoning, self-critique, structured extraction — from a model running on hardware you already own. No API costs. No data leaving your network.

And the best part: as local models improve, your plans improve for free.

Why fitz-graveyard?

Runs on modest hardware 🖥️

A 35B model at Q6 on a single GPU produces plans in ~15 minutes. A 70B model in RAM takes a few hours. You don't need a datacenter — you need patience and a machine that can stay on overnight.

Reads your codebase first 🔍

An agent builds a structural index of your codebase (classes, functions, imports), navigates it using keyword extraction to pick task-relevant files, summarizes them, and synthesizes a context document. An implementation check then verifies whether the task is already built before planning begins. Every planning stage sees your actual code, not a hallucinated version of it.

Per-field extraction that small models can handle 🧩

Each stage does 1 reasoning pass + 1 self-critique + N tiny JSON extractions (<2000 chars each). Even a 3B model can reliably produce structured output at this scale. Failed extractions get Pydantic defaults instead of crashing the stage — partial plan > no plan.

Crash recovery built in 🔄

Jobs checkpoint to SQLite. Machine crashes mid-plan? retry picks up from the last checkpoint. Power goes out overnight? Resume in the morning.

Claude where it counts, local everywhere else 🎯

The local model does the heavy lifting — 95% of the tokens. But the pipeline knows what it's uncertain about. Per-section confidence scoring flags weak spots, and those sections can pause for an Anthropic API review pass before the plan finalizes. You get Claude-grade quality on the parts that matter, at a fraction of the token cost. Fully optional — off by default, zero API calls unless you opt in.

Two interfaces, same engine 🔌

CLI for background job queues, MCP server for Claude Code / Claude Desktop integration. Both wrap the same tools/ service layer and SQLite job store.

Other features at a glance 🃏

1. Two LLM providers. Ollama (with OOM fallback to smaller model) or LM Studio (OpenAI-compatible API).
2. Cross-stage coherence check. Post-pipeline pass verifies context → architecture → roadmap consistency.
3. Section-specific confidence scoring. Each section type (context, architecture, design, roadmap, risk) scored against its own criteria with 1-10 granularity.
4. Implementation detection. Surgical check prevents planning to build what already exists.

How It Works

An agent pre-stage followed by 3 merged planning stages. Each stage uses per-field extraction: one reasoning prompt produces analysis, a self-critique pass catches scope inflation and hallucinated files, then small JSON extractions pull structured data from the reasoning.

  [Agent]    map file tree → build structural index → navigate by keywords → summarize → synthesize
                 |
                 v
  [Check]    implementation check — is this task already built?
                 |
                 v
  [Stage 1]  Context — requirements, constraints, assumptions (4 field groups)
  [Stage 2]  Architecture + Design — merged stage (6 field groups)
  [Stage 3]  Roadmap + Risk — merged stage (3 field groups)
                 |
                 v
  [Post]     coherence check → confidence scoring (section-specific criteria) → optional API review → render markdown

# Install
pip install fitz-graveyard

# Queue a job
fitz-graveyard plan "Build a plugin system for data transformations"

# Start the background worker
fitz-graveyard run

# Check on it
fitz-graveyard status 1

# Read the plan
fitz-graveyard get 1

pip install "fitz-graveyard[api-review]"    # Anthropic API review pass
pip install "fitz-graveyard[lm-studio]"    # LM Studio provider (openai SDK)
pip install "fitz-graveyard[dev]"          # pytest, build tools

fitz-graveyard plan "description"   # Queue a planning job
fitz-graveyard run                  # Start background worker (Ctrl+C to stop)
fitz-graveyard list                 # Show all jobs
fitz-graveyard status <id>          # Check progress
fitz-graveyard get <id>             # Print completed plan as markdown
fitz-graveyard retry <id>           # Re-queue failed/interrupted job
fitz-graveyard confirm <id>         # Approve optional API review
fitz-graveyard cancel <id>          # Skip API review, finalize plan
fitz-graveyard serve                # Start MCP server

QUEUED → RUNNING → COMPLETE
                 → AWAITING_REVIEW → QUEUED (confirm) / COMPLETE (cancel)
                 → FAILED / INTERRUPTED (both retryable)

{
  "mcpServers": {
    "fitz-graveyard": {
      "command": "fitz-graveyard",
      "args": ["serve"]
    }
  }
}

# LLM provider: "ollama" or "lm_studio"
provider: ollama

ollama:
  base_url: http://localhost:11434
  model: qwen2.5-coder-next:80b-instruct
  fallback_model: qwen2.5-coder-next:32b-instruct  # OOM fallback (null to disable)
  timeout: 300
  memory_threshold: 80.0  # RAM % threshold to abort

lm_studio:
  base_url: http://localhost:1234/v1
  model: local-model
  timeout: 300

agent:
  enabled: true
  max_summary_files: 15
  source_dir: null  # null = cwd at runtime

confidence:
  default_threshold: 0.7
  security_threshold: 0.9

anthropic:
  api_key: null  # null = API review disabled
  model: claude-sonnet-4-5-20250929

output:
  plans_dir: .fitz-graveyard/plans
  verbosity: normal

CLI (typer)   ──→ tools/ ──→ SQLiteJobStore ←── BackgroundWorker ──→ PlanningPipeline
MCP (fastmcp) ──→ tools/ ──→ SQLiteJobStore

fitz_graveyard/
├── cli.py                     # Typer CLI (9 commands)
├── server.py                  # FastMCP server + lifecycle
├── __main__.py                # python -m fitz_graveyard (MCP stdio)
├── tools/                     # Service layer
├── models/                    # JobStore ABC, SQLiteJobStore, JobRecord
├── background/                # BackgroundWorker, signal handling
├── llm/                       # LLM clients (Ollama, LM Studio), retry, memory monitor
├── planning/
│   ├── pipeline/stages/       # 3 merged stages + orchestrator + checkpoints
│   ├── agent/                 # Multi-pass codebase context gatherer
│   ├── prompts/               # Externalized .txt prompt templates
│   └── confidence/            # Per-section confidence scoring
├── api_review/                # Anthropic review client + cost calculator
├── config/                    # Pydantic schema + YAML loader
└── validation/                # Input sanitization

git clone https://github.com/yafitzdev/fitz-graveyard.git
cd fitz-graveyard
pip install -e ".[dev]"  # editable install for development
pytest  # 400 tests

License

MIT

Links

• GitHub
• PyPI