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

Standalone binaries (no Python required)

Prebuilt phoson-cli binaries are attached to each GitHub release — no Python interpreter or uv/pip needed (issue #93):

Download the asset for your platform, make it executable (Unix), and run:

chmod +x phoson-cli
./phoson-cli --setup     # configure credentials, then just: phoson-cli

--self-update inside a binary points back to the Releases page (the binary is a one-file bundle with no package metadata to upgrade).

🛠️ 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.

Command-line flags (one-off overrides for this run; they never touch ~/.phoson/config.toml):

phoson-cli --version                 # print the version and exit
phoson-cli --model openai/gpt-4o     # override the model
phoson-cli --provider openai         # override the provider
phoson-cli --theme light             # override the theme (dark|light|ansi|no-color)
phoson-cli --max-turns 25            # override max_iterations for this run
phoson-cli --classic                 # use the classic line-by-line REPL
phoson-cli --no-fullscreen           # alias of --classic

The full-screen TUI is the default interactive front end. --classic launches the retained classic REPL (Rich scrollback, line-by-line streaming) — useful for debugging and on terminals without full-screen support. When TERM is unset or dumb on an interactive terminal, the classic REPL is selected automatically with a notice on stderr.

Available commands:

• /new — Start a new session
• /model <name> — Switch model
• /tree — Show conversation tree
• /sessions — List saved sessions
• /label <text> — Label current node
• /theme — Pick or set the color theme (live preview; list to list)
• /keys — List key bindings and how to remap them
• /undo — Undo the last turn (branch from before your last message)
• /skills — List available skills (/skills <name> shows one's instructions)
• /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).

Startup update check: at launch the CLI checks PyPI in the background — at most once every 24 h (cache in ~/.phoson/last_update_check; a failed check is retried on the next start). When a newer release exists it shows a dim one-line hint: ⬆ v0.8.1 available — /update — in the TUI header (full-screen) or the prompt line (classic). It never blocks first paint, input, or a run, and one-shot mode is untouched. /update or --self-update install it.

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).

Themes (light/dark aware): the first time you run phoson-cli without a saved theme, it asks your terminal for its default background color (COLORFGBG env when present, otherwise a ~150 ms OSC 11 probe that iTerm2, kitty, WezTerm, Alacritty, ghostty, VS Code and friends answer) and offers to save the matching tier — light or dark — as your default. If the terminal can't be classified it just doesn't ask. /theme opens a live-preview picker (the banner and every token rendered in the tier's own colors) in both front ends; /theme <tier> sets it directly and /theme list lists the four tiers. Switching applies immediately — no restart needed.

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).

Prompt caching: long conversations re-send the whole history on every turn; phoson-cli keeps that prefix cacheable so providers can bill it at the (much cheaper) cache-read rate instead of full input price. It is on by default, no configuration needed:

• Anthropic — each request carries ephemeral cache_control breakpoints on the three stable parts of the prompt: the system prompt, the tool list, and the end of the conversation history (which advances as it grows). Cached usage shows up as cache_creation / cache_read tokens.
• OpenRouter — the conversation's session id is sent as session_id (sticky routing) so OpenRouter pins you to one upstream provider and its cache stays warm from the first turn; anthropic/* models additionally opt into automatic caching. The adapter also identifies itself as phoson-cli in OpenRouter's app rankings.

The system prompt is deliberately a stable prefix (date + timezone, not a live clock) so it does not bust the cache between turns. Cached tokens accumulate in the session metrics and surface in /status (cache R read / W write) and /tokens (cache=Rr/Ww). Cached reads cost 10–50% of the base input price, so a warm cache typically cuts long-session prompt cost by 50–90%. See docs/api/phoson_llm.md for per-provider details.

Key bindings (customizable): the full-screen TUI's keys are remappable from the [keys] section of ~/.phoson/config.toml (IMPROVEMENTS.md E6) — one line per action, each a prompt_toolkit key sequence (a list means "try in order", and "" unbinds the action):

[keys]
toggle_reasoning = "c-x"          # Ctrl+X instead of Ctrl+T
line_up = ["s-up", "c-up"]        # list = precedence order
submit = ""                       # unbind (use mouse / another key)

/keys lists the effective map (defaults or your remaps) plus the config syntax. Sequences are validated at startup: an unparseable key, an unknown action, or a sequence bound to two actions is a clear error before the UI opens — never a silent fallback. Remaps apply on the next start. The classic REPL's single global key (Ctrl+T) is fixed.

Rewind (double-Esc, full-screen TUI): press Esc twice in quick succession while idle and a picker lists your earlier messages — select one to jump the conversation back to just before it, the same UX as Claude Code's double-Esc. The chat pane redraws up to that point, your composer is pre-filled with the selected message (edit it and press Enter to re-send), and Ctrl+Z undoes the jump, restoring the previous point (repeat it to undo several consecutive rewinds). The "undone" messages are not deleted — they remain as an abandoned branch in the conversation tree (still visible via /tree), and session cost/token totals stay cumulative (same contract as /undo). Precedence with the single-Esc run cancel is fixed: a lone Esc while a turn is running still cancels it immediately, and double-Esc is only interpreted when idle. The double-tap rides on whatever key escape is bound to — remapping escape in [keys] moves both the single-Esc cancel and the double-Esc rewind together (unbinding escape disables both); the jump undo undo_jump (default Ctrl+Z) is remappable on its own.

Selecting/copying chat text (full-screen TUI): the chat pane sets mouse_support=True so the scroll wheel is handled by the app — this is a terminal-level mouse-tracking switch (xterm's own DECSET 1000/1002/1006 modes), not something the app can opt out of selectively, and turning it on is what makes the terminal stop treating a plain click-drag as native text selection (every mouse-aware TUI — Claude Code, Pi, OpenCode — hits the same trade-off). Hold Shift while dragging to select text: this tells the terminal itself to ignore the app's mouse tracking for that gesture and fall back to its own native selection/copy, unaffected by whatever the app is doing (works in GNOME Terminal, iTerm2, Alacritty, WezTerm, Ghostty, kitty, Windows Terminal — check your terminal's docs if the modifier differs). The footer's [Shift+Drag] Select text hint is a reminder of this.

Clickable hyperlinks in responses: Markdown links in assistant answers render as real OSC 8 terminal hyperlinks (ESC ] 8 ; ; URL ESC \) in both front ends — the same escape sequence Neovim, tmux and modern editors use. In a terminal that supports it (kitty, iTerm2, WezTerm, GNOME Terminal, Ghostty, Alacritty, Windows Terminal, …) the link text becomes clickable, typically with Ctrl+click (check your terminal's docs — the exact modifier follows the same convention as the Shift+Drag text-selection bypass above: it's the terminal that intercepts the gesture, not the app, so it isn't affected by mouse_support=True capturing the rest of the mouse for the scroll wheel).

@file mentions: type @ in the message and the composer offers repo paths (fuzzy-filtered as you type, with a size hint per file); selecting one inserts the path. On send, each @mention is expanded into the file's content so the model sees the actual file — text files are inlined, and images/audio/video/pdf become their native media blocks (same as /attach). Works in both the full-screen TUI and the classic REPL. user@domain emails and bare @user handles in prose are left alone.

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

Skills (on-demand instructions): a skill is a directory with a SKILL.md file — YAML frontmatter (name + description) followed by Markdown instructions, optionally next to bundled scripts/ and references/. Unlike AGENTS.md (always in the prompt) or a tool (schema in every request), a skill costs one line while dormant: only its name: description is indexed in the system prompt, and the agent pulls the full body in with the skill tool when it decides the skill matches the task. On this repo's own skill that is 157 tokens indexed vs 2399 loaded — 15× cheaper until it is actually needed, and because the body arrives as a tool result (not in the system prompt), loading a skill mid-session never invalidates the prompt cache.

.phoson/skills/code-reviewer/SKILL.md     # project skill (git-versionable)
~/.phoson/skills/my-workflow/SKILL.md     # available in every repo

---
name: code-reviewer
description: Use when the user asks to review a diff, a PR or a file for
  bugs, security issues or style violations.
---

# Code reviewer

1. Run `git diff` to see the change.
2. Check for N+1 queries, missing error handling, unvalidated input.
3. Report findings grouped by severity.

Project skills shadow same-named global ones. .agents/skills/ and .claude/skills/ are also read, so a repo already set up for another agent harness works unchanged (same rationale as the CLAUDE.md alias). /skills lists what was discovered and /skills <name> prints a skill's full instructions. The skill tool only joins the registry when at least one skill exists — no skills, no added schema on any request.

Models file: ~/.phoson/models.json (optional) holds model overrides (context window, labels — user-defined models appear in /model) and non-sensitive provider settings (default_model, base_url for self-hosted/proxied endpoints). Model listings are always fetched live — a bare /model shows one unified picker of every configured provider (OpenRouter ordered by agentic_index), and a provider whose fetch fails is marked unavailable instead of silently degrading. API keys never live there; see docs/api/phoson_cli.md.

Context management (long sessions): when a session grows past a fraction of the model's context window, phoson compacts it automatically — older turns are replaced by a structured handoff summary (goal, completed work, key decisions, a distillation of the model's reasoning, open questions, next steps, constraints) so continuity survives long tasks. Captured reasoning from the summarized turns is folded into that summary, not dropped. You control it:

• /compact previews what would be summarized and asks before applying it; /compact aggressive previews a deeper cut.
• /compact on|off toggles automatic compaction at runtime (persisted).
• ~/.phoson/config.toml [defaults] knobs: compact_mode (balanced|aggressive|off), compact_threshold (fraction of the window that triggers auto-compact), compact_min_keep_messages (recent turns kept verbatim), and offload_tool_outputs / offload_max_chars (large tool results — default >24 KB — are written to ~/.phoson/compacted/ with only a head/tail preview kept in context).

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. While idle, press Esc twice to rewind the conversation to an earlier message (a picker lists your previous messages; the pane redraws to the chosen point and your composer is pre-filled with that message — Ctrl+Z undoes the jump); see Key bindings (customizable) below. 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