no-slop-harness 0.9.0

Deterministic, local-first LLM orchestration framework implementing the CIV pattern.

No-Slop Harness

Deterministic, local-first LLM orchestration framework implementing the CIV (Coordinator-Implementor-Verifier) pattern for zero-slop, high-fidelity software engineering.

Overview

No-Slop Harness is an agentic framework that enforces structured, verifiable LLM workflows. It rejects the "black-box agent" model in favor of a three-phase pipeline where each phase has explicit constraints, validated schemas, and deterministic handoffs.

The CIV Pattern

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│ Coordinator │ ──▶ │ Implementor │ ──▶ │  Verifier   │
│   (plan)    │     │  (execute)  │     │  (validate) │
└─────────────┘     └─────────────┘     └─────────────┘
       │                                       │
       └─────────── feedback loop ─────────────┘

• Coordinator: Decomposes user requests into a DAG of typed Task objects
• Implementor: Executes tasks using a constrained toolset (read_file, write_file, edit_file_ast, bash_execute)
• Verifier: Validates output via tests, linting, type checking — rejects slop before it merges

Key Features

• Deterministic scheduling — Kahn's algorithm with priority-aware topological sort
• Sandboxed execution — command allowlisting, blocklisting, timeout enforcement, output truncation
• Structured inter-agent protocol — CIVMessage schema enforces typed communication between phases
• AST-aware editing — tree-sitter powered code modifications with syntax validation fallback
• Pydantic schema enforcement — every tool call, task, and message is validated at the boundary
• Zero external dependencies for core — only pydantic, rich, click, tree-sitter

Quick Start

Installation

Recommended: pipx (isolated, works on all distros including Arch, Debian 12+, Ubuntu 24.04+, Fedora)

# Install pipx if you don't have it
sudo pacman -S pipx        # Arch
sudo apt install pipx       # Debian/Ubuntu
sudo dnf install pipx       # Fedora

# Install from GitHub
pipx install git+https://github.com/iknowkungfubar/no-slop-harness.git

# Or with inference support (adds httpx for LLM API calls)
pipx install git+https://github.com/iknowkungfubar/no-slop-harness.git[inference]

uv (fast, modern, venv-based)

uv tool install git+https://github.com/iknowkungfubar/no-slop-harness.git[inference]

venv (traditional, works everywhere)

python -m venv .venv
source .venv/bin/activate
pip install git+https://github.com/iknowkungfubar/no-slop-harness.git[inference]

pip --break-system-packages (quick, not recommended)

pip install --break-system-packages git+https://github.com/iknowkungfubar/no-slop-harness.git

GitHub install. Not yet on PyPI. pipx install no-slop-harness[inference] from GitHub works today.

Extras reference:

Extra	Adds	Size
(none)	Core: schemas, orchestration, CLI, verifier, plugin system	~5MB
[inference]	httpx — OpenAI-compatible LLM API client	+3MB
[dev]	pytest, ruff, mypy — development tools	+15MB

Basic Usage

# Initialize a pipeline session
no-slop init --sandbox-allowlist echo --sandbox-allowlist python

# Check pipeline status
no-slop status

Programmatic Usage

Synchronous (core only):

from no_slop_harness.orchestrator import PipelineOrchestrator
from no_slop_harness.schemas import Task, SandboxConfig

sandbox = SandboxConfig(
    allowed_commands=["echo", "python", "pytest"],
    timeout_seconds=120,
)

pipeline = PipelineOrchestrator(sandbox_config=sandbox)

tasks = [
    Task(task_id="add_model", description="Create User model", action="Add SQLAlchemy model"),
    Task(task_id="add_tests", description="Add unit tests", action="Write pytest suite", dependencies=["add_model"]),
]
msg = pipeline.ingest_tasks(tasks)

while task := pipeline.next_task():
    pipeline.report_result(task.task_id, "done", success=True)
    pipeline.verify_task(task.task_id)
    pipeline.verification_complete(task.task_id, passed=True)

print(pipeline.status())

Full pipeline with LLM (requires [inference]):

import asyncio
from no_slop_harness.runner import CIVPipeline

async def main():
    pipeline = CIVPipeline(
        base_url="http://localhost:1234/v1",
        model="qwen/qwen3.6-35b-a3b",
    )
    result = await pipeline.run("Add a hello() function to demo.py")
    print(result["success"], result["summary"])

asyncio.run(main())

Architecture

src/no_slop_harness/
├── __init__.py              # Package version
├── cli.py                   # Click-based CLI (init, status, list, verify, report)
├── runner.py                # End-to-end CIV pipeline runner
├── schemas.py               # Pydantic models (Task, CIVMessage, ToolCall, SandboxConfig, PipelineState)
├── orchestrator.py          # CIV PipelineOrchestrator lifecycle
├── async_orchestrator.py    # Async pipeline for parallel task execution
├── dag.py                   # Topological sort (Kahn's) + DAG validation
├── pipeline_scheduler.py    # TaskScheduler + ResultCollector
├── sandbox.py               # Sandboxed command execution (allowlist, blocklist, timeout)
├── ast_editor.py            # Tree-sitter AST editor with regex fallback
├── verifier.py              # Test/lint/typecheck runner
├── worktree.py              # Git worktree isolation per task
├── sdlc.py                  # .sdlc/ context injection (ADRs, standards, memory)
├── constrained.py           # llguidance grammar-enforced JSON output
├── rag.py                   # RAG + self-healing hallucination detection
├── advanced_metrics.py      # Token entropy, variance penalty, inter-step timing
├── tla_bridge.py            # TLA+ formal verification bridge (spec gen + TLC)
├── llm_client.py            # LLM provider abstraction with retry logic
├── logging_config.py        # Structured logging (JSON formatter, PipelineLogger)
├── metrics.py               # Observability (counters, timers, histograms)
├── plugin.py                # Plugin system (discovery, registration, lifecycle hooks)
├── errors.py                # Exception hierarchy
├── agents/                  # CIV agent implementations
│   ├── coordinator.py       # Task decomposition agent
│   ├── implementor.py       # Task execution agent with constrained toolset
│   └── verifier.py          # Automated verification agent
├── providers/               # LLM provider backends
│   └── openai_compatible.py # OpenAI-compatible API (LM Studio, OpenRouter, vLLM, Ollama)
└── prompts/                 # Agent system prompt templates
    ├── coordinator.txt
    ├── implementor.txt
    └── verifier.txt

Development

See CONTRIBUTING.md for development setup and guidelines.

Running Tests

pip install -e ".[dev]"
python -m pytest tests/ -v

Code Quality

python -m ruff check src/ tests/
python -m mypy src/ --ignore-missing-imports

Documentation

• ARCHITECTURE.md — Design decisions, data flow, and phase lifecycle
• AGENTS.md — AI agent operating rules and context conventions
• CONTRIBUTING.md — Development workflow and PR process
• CHANGELOG.md — Version history

Contributing

Contributions are welcome! Please read CONTRIBUTING.md for detailed guidelines on our development process, coding standards, PR workflow, and code of conduct.

License

Apache 2.0 — see LICENSE for details.