specsoloist 0.7.0

Spec-as-Source Framework: AI-driven development where specifications are the source of truth.

SpecSoloist

SpecSoloist is a "Spec-as-Source" AI coding framework. It treats specifications as the source of truth and uses AI agents to compile them into executable code. Compose systems from natural language, conduct parallel builds, and detect spec-vs-code drift.

Why SpecSoloist?

Code is often messy, poorly documented, and prone to drift from original requirements. SpecSoloist flips the script:

1. Write Specs: You write requirements-oriented specifications (Markdown).
2. Compile to Code: AI agents read your specs and write implementations directly.
3. Self-Healing: If tests fail, agents analyze the failure and patch the code.
4. Orchestrate: Define complex workflows where agents collaborate, share state, and pause for human input.

Code is a build artifact. Specs are the source of truth.

Installation

pip install specsoloist

Quick Start

1. Set your API Key (Gemini or Anthropic):

   export GEMINI_API_KEY="your_key_here"
   # or
   export ANTHROPIC_API_KEY="your_key_here"
   

2. Scaffold a new project:

   sp init my-app                                  # blank Python arrangement
   sp init my-app --template python-fasthtml       # FastHTML + pytest
   sp init my-app --template nextjs-vitest         # Next.js + vitest
   sp init my-app --template nextjs-playwright     # Next.js + Playwright E2E
   sp init --list-templates                        # see all options
   

3. Write a brief and vibe:

   cd my-app
   sp vibe "A todo app with auth"          # compose → conduct in one command
   sp vibe brief.md                        # read brief from a Markdown file
   sp vibe brief.md --pause-for-review     # pause to edit specs before building
   sp vibe brief.md --resume               # treat brief as addendum (skip compiled specs)
   

   Or run the steps separately:

   sp compose "A todo app with auth"       # draft specs
   sp conduct specs/                       # build
   

Manual workflow (single spec)

sp create calculator "A simple calculator with add and multiply"
sp compile calculator
sp test calculator
sp fix calculator  # if tests fail

Orchestration

SpecSoloist allows you to chain multiple specs into a workflow.

1. Draft Architecture: Use sp compose to vibe-code your system.

   sp compose "A data pipeline that fetches stocks and calculates SMA"
   

   This generates a component architecture and draft specs.

2. Conduct Build: Compile all components via agent orchestration.

   sp conduct
   

   The conductor agent resolves dependency order and spawns soloist agents to compile each spec in parallel.

Adding SpecSoloist to an Existing Project

Already have a FastHTML or Next.js app? You don't have to start over. sp respec extracts specs from your existing code, and the spec layer can coexist with hand-written code indefinitely.

See the Incremental Adoption Guide for a step-by-step walkthrough, and examples/fasthtml_incremental/ for a concrete before/after example.

sp respec src/mymodule.py     # extract a spec from existing code
sp validate specs/mymodule.spec.md
sp conduct specs/mymodule.spec.md --arrangement arrangement.yaml

CLI Reference

Command	Description
sp init [name]	Scaffold a new project (--template, --list-templates)
sp vibe [brief]	Single-command pipeline: compose → conduct (--pause-for-review, --resume)
sp compose	Draft architecture & specs from natural language
sp conduct [dir]	Build project via conductor/soloist agents
sp respec	Reverse engineer code to spec
sp fix	Auto-fix failing tests (Agent-first)
sp validate	Check spec structure
sp status	Show compilation state of each spec
sp schema [topic]	Show annotated schema for arrangement.yaml (--json for JSON Schema)
sp help [topic]	Show bundled reference guides (arrangement, spec-format, conduct, overrides, specs-path)
sp doctor	Check environment health (API keys, CLIs, tools, env vars)
sp --version	Print version and exit
sp compile	Compile single spec to code + tests
sp test [--all]	Run tests for a spec (or all compiled specs)
sp diff	Compare spec against compiled implementation (drift detection)
sp build	Compile all specs (direct LLM, no agents)
sp list	List all specs
sp create	Create a new spec manually
sp graph	Export dependency graph (Mermaid.js)
sp verify	Verify schemas and interface compatibility

Commands that use agents (vibe, compose, conduct, respec, fix) default to detecting an available agent CLI (Claude Code or Gemini CLI). Use --no-agent to fall back to direct LLM API calls.

Running inside Claude Code? sp conduct spawns a Claude subprocess, which is blocked inside an active Claude Code session. If you see a "Heads Up" warning, either open a separate terminal, use --no-agent, or use the Agent tool to spawn the conductor directly (see AGENTS.md).

Configuration

You can configure SpecSoloist via environment variables or a .env file:

# Gemini
export SPECSOLOIST_LLM_PROVIDER="gemini"
export GEMINI_API_KEY="your_key_here"
export SPECSOLOIST_LLM_MODEL="gemini-2.0-flash"  # optional

# Anthropic
export SPECSOLOIST_LLM_PROVIDER="anthropic"
export ANTHROPIC_API_KEY="your_key_here"
export SPECSOLOIST_LLM_MODEL="claude-sonnet-4-20250514"  # optional

# Pydantic AI (OpenAI, OpenRouter, Ollama, and more)
export SPECSOLOIST_LLM_PROVIDER="pydantic-ai"
export SPECSOLOIST_LLM_MODEL="openai:gpt-4o"  # or openrouter:..., ollama:...

Arrangement Files

An Arrangement is SpecSoloist's makefile — it bridges language-agnostic specs to a concrete build environment by specifying the target language, output paths, build commands, and constraints.

sp init --template <name> copies a validated starter arrangement into your project. Run sp init --list-templates to see what's available. You can also start from one of the examples in arrangements/:

See arrangements/arrangement.python.yaml for a complete example:

target_language: python
output_paths:
  implementation: src/mymodule.py
  tests: tests/test_mymodule.py
environment:
  tools: [uv, ruff, pytest]
  setup_commands: [uv sync]
build_commands:
  lint: uv run ruff check .
  test: uv run pytest
constraints:
  - Must use type hints for all public function signatures

Usage:

# Explicit path
sp compile myspec --arrangement arrangement.yaml
sp build --arrangement arrangement.yaml
sp conduct --no-agent --arrangement arrangement.yaml

# Auto-discovery: place arrangement.yaml in your project root
# and it will be picked up automatically
sp compile myspec

External Dependencies

Specs describe your code. External libraries are inputs to the build. There are three patterns depending on how well the LLM knows the library:

1. Well-known libraries — constraints only

For React, pytest, lodash, etc. the soloist already knows the API. Just mention it in the arrangement or spec constraints:

constraints:
  - Use React hooks (useState, useEffect) for state management
  - Use Tailwind CSS for styling

2. Obscure or new libraries — reference spec

For newer libraries (e.g. FastHTML) where LLMs may hallucinate the API, write a type: reference spec documenting the subset you actually use. No code is generated — the spec body is injected as context into every dependent soloist's prompt:

---
name: fasthtml_interface
type: reference
---
# Overview
FastHTML is a Python web framework. Package: `python-fasthtml`. Import: `from fasthtml.common import *`.

# API
## Components
- `Div(**attrs, *children)` — renders a div
- `Form(hx_post, hx_swap, *children)` — HTMX-enabled form
- `Input(name, type, placeholder)` — form input

## App
- `@rt(path)` — route decorator
- `serve()` — start dev server

# Verification
```python
from fasthtml.common import Div, Form, Input
assert callable(Div)

### 3. Complex SDKs — adapter spec

For SDKs with many moving parts (e.g. Vercel AI SDK), write a thin adapter spec that wraps the SDK. Everything else in your project depends on *your adapter*, not the SDK directly. If you swap the underlying SDK, only the adapter spec changes:

```markdown
---
name: ai_client
type: bundle
dependencies: []
---
# AI Client

Wraps the Vercel AI SDK for this project.

## `streamChat(messages, options)`
Streams a chat completion using `streamText()` from the `ai` package.
Returns an AI SDK `StreamingTextResponse`.

See examples/ for concrete interface and adapter spec examples.

Sandboxed Execution (Docker)

For safety, SpecSoloist can run generated code and tests inside an isolated Docker container.

1. Build the sandbox image:

   docker build -t specsoloist-sandbox -f docker/sandbox.Dockerfile .
   

2. Enable sandboxing:

   export SPECSOLOIST_SANDBOX=true
   # Optional: override the image (default: specsoloist-sandbox)
   # export SPECSOLOIST_SANDBOX_IMAGE="my-custom-image"
   

3. Run tests: sp test my_module will now wrap execution in docker run.

Native Subagents (Claude & Gemini)

For the full agentic experience, SpecSoloist provides native subagent definitions for Claude Code and Gemini CLI. These allow the AI to delegate tasks to specialized agents:

Agent	Purpose
compose	Draft architecture and specs from natural language
conductor	Orchestrate builds — resolves dependencies, spawns soloists
soloist	Compile a single spec — reads spec, writes code directly
respec	Extract requirements from code into specs
fix	Analyze failures, patch code, and re-test

Usage with Claude Code:

> conduct score/
> respec src/specsoloist/parser.py to score/parser.spec.md

Usage with Gemini CLI:

> compose a todo app with user auth

The subagent definitions are in .claude/agents/ and .gemini/agents/.