phoson-engine-minimal

Minimal Python runtime for the Phoson autonomous-agent platform

🔥 Open Source — Built for developers who want full control over their AI agents.

📋 Table of Contents

• What this project is
• Why Phoson?
• Features
• High-level architecture
• Repository map
• Core modules
  • phoson_llm — LLM normalization layer
  • phoson_agent — Agent orchestration
  • phoson_agent.sessions — Conversation persistence
  • phoson_cli — Interactive REPL
• 🚀 Quick Start
• Installation
• Development setup
• Run checks locally
• Environment variables
• Usage examples
  • Minimal agent usage
  • Define a tool
  • Interactive CLI
• CI and security workflows
• Commit message format
• Roadmap
• Contributing
• License
• Support

🤔 What this project is

phoson-engine-minimal is the core runtime behind the Phoson autonomous-agent platform. It's a lightweight, framework-free Python implementation that gives you complete control over your AI agents without the bloat of heavy frameworks.

Unlike other agent frameworks (LangChain, LangGraph, etc.), Phoson is built from scratch using provider SDKs directly, with a custom ReAct loop designed for:

• 🔄 Streaming behavior — Token-by-token events for real-time UIs
• 🔧 Tool-call orchestration — Full control over tool execution
• 💰 Cost accounting — Track spend per run with built-in pricing
• 👁️ Observability — RunStep events and typed event streams
• 🌳 Session trees — Branchable conversation history (not linear!)
• ⌨️ Interactive REPL — Debug and iterate on agents interactively

🎯 Why Phoson?

✨ Features

🏗️ High-level architecture

flowchart LR
    U[App / CLI / API] --> AE[AgentEngine\nphoson_agent]
    AE --> MW[Middleware Hooks]
    AE --> T[Registered Tools]
    AE --> S[ConversationTree + Storage]
    AE --> C[BaseLLMChat Contract]
    C --> OA[OpenAIChat]
    C --> AN[AnthropicChat]
    OA --> P1[OpenAI / OpenRouter / Ollama]
    AN --> P2[Anthropic]
    OA --> E[Typed LLM Events]
    AN --> E
    E --> AE
    AE --> R[Agent Events + RunResult]

Runtime loop (tool call cycle)

sequenceDiagram
    participant Client
    participant Engine as AgentEngine
    participant LLM as LLM Adapter
    participant Tool as Tool Handler

    Client->>Engine: run(messages, config)
    Engine->>LLM: stream(history, config, tools)
    LLM-->>Engine: TokenEvent / ReasoningTokenEvent
    LLM-->>Engine: ToolCallEvent
    Engine->>Tool: execute(args)
    Tool-->>Engine: result/error
    Engine->>LLM: continue with ToolResultBlock
    LLM-->>Engine: UsageEvent + LLMDoneEvent
    Engine-->>Client: AgentRunResult

🗺️ Repository map

phoson-engine-minimal/
├── phoson_llm/           # LLM normalization layer (adapters + schemas + pricing)
├── phoson_agent/         # ReAct agent loop, tools, middleware, sessions
├── phoson_cli/           # Interactive CLI (REPL) for agent sessions
├── phoson_plugin_*/      # Official plugins (checkpoint, mcp, memory)
├── tests/                # Unit/integration tests for all layers
├── docs/api/             # Per-package API documentation
├── .github/workflows/    # CI and security automation
├── ROADMAP.md            # Project roadmap
└── pyproject.toml        # Project metadata, dependencies, tooling config

📦 Core modules

phoson_llm — LLM normalization layer

Provider adapters return a single typed event stream (LLMEvent subclasses):

Supported providers:

All of them are available via the build_chat() factory, e.g. build_chat("openrouter"), and expose the same stream() event contract.

Pricing module (phoson_llm.pricing) provides calculate_cost() for provider-level USD usage.

phoson_agent — Agent orchestration

Stateless-by-run orchestration over message history with tool execution:

• AgentEngine — Main entry point for running agents (async and sync)
• @tool decorator — Transform Python functions into AgentTool definitions with JSON Schema
• AgentMiddleware — Hooks for pre/post processing (LLM calls, tool execution)
• AgentContext — Shared state across middleware and tools

phoson_agent.sessions — Conversation persistence

• ConversationTree — Branchable conversation structure (not linear)
• ConversationNode — Individual node with messages, children, label
• JsonlStorage — JSONL-backed session storage (local file)
• SessionMeta — Session metadata (id, message_count, created_at, updated_at)

phoson_cli — Interactive REPL

Command-line interface for interactive agent sessions:

• PhosonRepl — Interactive read-eval-print loop
• Commands: /exit, /quit, /clear, /new, /model, /tree, /sessions, /label, /help
• Real-time streaming responses
• Session persistence and labeling
• Multiple model switching

🚀 Quick Start

from phoson_agent import AgentEngine
from phoson_llm.chats.openai import OpenAIChat
from phoson_llm.schemas import Message, ModelConfig

engine = AgentEngine(
    chat=OpenAIChat(),
    tools=[],
    phoson_weight=1.2,
)

result = engine.run_sync(
    messages=[Message(role="user", content="Summarize this project in one line")],
    config=ModelConfig(model="openai/gpt-4o-mini", max_tokens=128),
)

print(result.final_content)
print(result.total_cost_usd, result.total_credits)

Or run the interactive CLI:

uv run phoson-cli

Run the setup wizard to configure provider credentials and defaults:

uv run phoson-cli --setup

📥 Installation

# Clone the repository
git clone https://github.com/phoson-lat/phoson-engine-minimal.git
cd phoson-engine-minimal

# Install dependencies
uv sync --dev --locked

# Install git hooks
uv run pre-commit install --install-hooks
uv run pre-commit install --hook-type commit-msg
uv run pre-commit install --hook-type pre-push

🛠️ Development setup

Install dependencies

uv sync --dev --locked

Install git hooks

uv run pre-commit install --install-hooks
uv run pre-commit install --hook-type commit-msg
uv run pre-commit install --hook-type pre-push

✅ Run checks locally

uv sync --dev --all-extras   # --all-extras is needed for pyright (provider SDK stubs)
uv run ruff format --check .
uv run ruff check .
uv run pyright
uv run python -m compileall phoson_llm phoson_agent phoson_cli
uv run pytest -q

🔐 Environment variables

Set the variables for the providers you use (the adapter reads the default when no api_key is passed):

# Cloud providers
OPENAI_API_KEY=
ANTHROPIC_API_KEY=
OPENROUTER_API_KEY=
GEMINI_API_KEY=
MISTRAL_API_KEY=
GROQ_API_KEY=
XAI_API_KEY=
DEEPSEEK_API_KEY=
TOGETHER_API_KEY=
PERPLEXITY_API_KEY=
NVIDIA_API_KEY=
FIREWORKS_API_KEY=
COHERE_API_KEY=
GITHUB_TOKEN=

# Azure OpenAI
AZURE_OPENAI_ENDPOINT=
AZURE_OPENAI_API_KEY=
AZURE_OPENAI_DEPLOYMENT=

# AWS Bedrock (plus standard AWS credentials: AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY)
AWS_DEFAULT_REGION=us-east-1

Local servers (Ollama, LM Studio, vLLM) need no API key — just the right base_url.

💻 Usage examples

Minimal agent usage

from phoson_agent import AgentEngine
from phoson_llm.chats.openai import OpenAIChat
from phoson_llm.schemas import Message, ModelConfig

engine = AgentEngine(
    chat=OpenAIChat(),
    tools=[],
    phoson_weight=1.2,
)

result = engine.run_sync(
    messages=[Message(role="user", content="Summarize this project in one line")],
    config=ModelConfig(model="openai/gpt-4o-mini", max_tokens=128),
)

print(result.final_content)
print(result.total_cost_usd, result.total_credits)

Define a tool

import ast
import operator

from phoson_agent import tool


@tool
def calculate(expression: str) -> str:
    """Safely evaluate a basic arithmetic expression like "2 + 2 * 10"."""

    def _eval(node: ast.AST):
        match node:
            case ast.Expression(body=value):
                return _eval(value)
            case ast.Constant():
                return node.value
            case ast.BinOp(left=left, op=op, right=right):
                ops = {
                    ast.Add: operator.add,
                    ast.Sub: operator.sub,
                    ast.Mult: operator.mul,
                    ast.Div: operator.truediv,
                }
                if type(op) not in ops:
                    raise ValueError(f"Unsupported operator: {type(op).__name__}")
                return ops[type(op)](_eval(left), _eval(right))
            case ast.UnaryOp(op=op, operand=value) if isinstance(op, ast.USub):
                return -_eval(value)
        raise ValueError(f"Unsupported expression: {expression!r}")

    return str(_eval(ast.parse(expression, mode="eval")))

⚠️ Never use eval()/exec() on model-generated input — treat LLM output as untrusted and validate or sandbox every tool argument.

Interactive CLI

uv run phoson-cli

One-shot mode (no REPL, no session — for scripts and CI):

phoson-cli "fix the failing tests"     # positional task
phoson-cli -p "summarize this repo"    # --print flag
echo "explain the CI failure" | phoson-cli   # piped stdin

The final answer is printed to stdout; the exit code is 0 on success and 1 on agent error.

Available commands:

• /new — Start a new session
• /model <name> — Switch model
• /tree — Show conversation tree
• /sessions — List saved sessions
• /label <text> — Label current node
• /undo — Undo the last turn (branch from before your last message)
• /update — Check for and install CLI updates
• /help — Show all commands

Self-update: phoson-cli --self-update performs the same check/upgrade flow from outside the REPL (e.g. from a script).

Appearance: PHOSON_THEME=light|ansi|no-color (or theme = "..." in ~/.phoson/config.toml) switches the color tier; NO_COLOR / CLICOLOR=0 always produce plain output (scripts, CI).

Reasoning: press Ctrl+T to toggle the live "thinking" view while a run is streaming, or to expand the full reasoning of the last turn after it finishes (persisted with the session, so it survives resume).

Permissions: control what each tool may do via ~/.phoson/permissions.json:

{
  "levels": { "bash": "ask", "web_search": "deny" },
  "allow_patterns": { "bash": ["git status", "pytest*", "uv *"] }
}

Levels: allow (run freely), ask (confirm every call), deny. A matching allow-pattern runs without asking even under ask/deny — handy for safe subcommands. Inspect or change levels at runtime with /permissions bash ask (persisted immediately). Non-interactive contexts (one-shot mode, scripts) fail closed: an ask-level tool is refused instead of hanging.

Project memory: drop an AGENTS.md in the repository root (or any directory between the root and your working directory) and its contents are injected into the agent's system prompt on every turn — no plugin or database needed. A global ~/.phoson/AGENTS.md applies everywhere; CLAUDE.md is supported as an alias; @path/to/file.md lines import other files; content is capped at ~2000 tokens with a visible truncation marker and re-read every turn. /agents-md lists what was loaded.

# AGENTS.md

- Use ruff for lint/format and pytest for tests — never black.
- Commit messages follow Conventional Commits.
- Public APIs need type hints and docstrings.
@docs/style-guide.md

Models file: ~/.phoson/models.json (optional) holds model overrides (context window, labels — user-defined models appear in /model), non-sensitive provider settings (default_model, base_url for self-hosted/proxied endpoints) and an automatic 24 h model-list cache that makes /model instant and works offline. API keys never live there; see docs/api/phoson_cli.md.

UI: the full-screen prompt_toolkit front end is the default interactive experience; it offers a persistent scrollable chat pane, multiline input (Ctrl+J inserts a newline, Enter sends), persistent input history (~/.phoson/history.txt, shared with the retained classic REPL), and /model//provider//sessions pickers and bash confirmation as overlay floats. The multiline composer wraps long pasted lines, takes only the height it needs (up to five lines), and scrolls internally after that cap. If a turn is already running, Enter keeps the draft and shows a warning; press Esc to cancel the active turn before sending it. The chat also shows a transient animated activity line immediately after sending (Thinking… with rotating phrases, then Streaming… / Running tool… as applicable), which vanishes when the turn settles. One-shot mode (phoson-cli "task") is always stdout-only.

🔒 CI and security workflows

• .github/workflows/ci.yml: Format check, lint, smoke compile, and tests on PRs and pushes to main.
• .github/workflows/security.yml: Dependency audit and secret scan on PRs, pushes to main, and weekly schedule.

📝 Commit message format

Conventional Commits are enforced through a commit-msg hook.

Examples:

feat: add streaming chat abstraction
fix: handle unknown model pricing fallback
chore: update pre-commit hook versions

Common types: feat, fix, docs, refactor, test, chore, ci

🗓️ Roadmap

For the project roadmap see ROADMAP.md, and per-package API documentation under docs/api/.

🤝 Contributing

Contributions are welcome! Here's how you can help:

1. Fork the repository
2. Create a feature branch: git checkout -b feature/amazing-feature
3. Commit your changes: git commit -m 'feat: add amazing feature'
4. Push to the branch: git push origin feature/amazing-feature
5. Open a Pull Request

🌐 Language policy: everything in this repository must be in English — documentation, docstrings, code comments, commit messages, issue titles and bodies, and PR descriptions. This keeps the project accessible to contributors worldwide. If you're more comfortable writing in another language, draft your changes in a branch and maintainers will help polish the English before merge.

Please read CONTRIBUTING.md for details on our code of conduct and development process.

Ideas for contributions

• 🆕 Add new LLM providers (20+ already supported — see the table above)
• 🔧 Improve tool execution (batching, retries, caching)
• 📊 Add observability integrations (OpenTelemetry, Langfuse)
• 🖥️ Build a web-based REPL or playground
• 📚 Improve documentation and examples

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

MIT License

Copyright (c) 2024 Phoson

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

💬 Support

• Issues: GitHub Issues for bug reports
• Discussions: GitHub Discussions for questions
• Documentation: See docs/api/ for per-package API notes
• Website: https://phoson.lat
• SDK Docs: https://phoson.lat/docs

⭐ Show your support

Give us a ⭐️ if this project helped you build better AI agents!

Built with 🔥 by phoson.lat