English | 中文

Lumen — A Model-Agnostic AI Coding Agent SDK

One command turns any LLM into your code architect.

Lumen is a model-agnostic, production-grade AI coding agent framework (Python 3.11+). Give it an API key and a model name, and you get an AI assistant that can deeply read, write, and analyze code. Works with OpenAI, Anthropic, DeepSeek, Ollama, and any OpenAI-compatible LLM.

---

Examples

Pomodoro timer — generated end-to-end in Review Mode

A pure HTML/CSS/JS Pomodoro timer with an animated SVG countdown ring, auto-switching work/break sessions, a WebAudio beep, and a localStorage-backed task list — built end-to-end by Lumen running gpt-4.1 under Review Mode (design → pipeline → implementation, gated by request_approval between phases).

Source: examples/pomodoro/

---

Features at a glance

• Model-agnostic — OpenAI / Anthropic / DeepSeek / Ollama / any OpenAI-compatible API, swap with a single line
• 13 built-in tools — file read/write, code search, shell execution, web search/fetch, LSP code intelligence, sub-agents
• Extended Thinking — Anthropic thinking blocks / OpenAI reasoning_effort / generic CoT, with adaptive budgeting
• Prompt Cache — native Anthropic cache_control plus a generic hash-based cache; automatically reduces API cost
• Structured output — JSON Schema-enforced output for OpenAI / Anthropic / Gemini, validated via Pydantic
• Smart permission control — feature-scored command classifier plus regex dual engine, five risk levels, fully explainable
• Dynamic session memory — LLM auto-extracts key facts, TF-IDF retrieval, no vector database required
• Automatic context compaction — compresses conversation history as the window fills up, continues seamlessly
• Skill system — reusable task templates, defined in JSON/YAML with fuzzy search
• Persistent retry — tiered model escalation plus webhook alerts, designed for CI / batch workloads
• Lifecycle hooks — pre/post tool hooks for interception, mutation, and remediation
• Auto-persisted transcripts — JSONL append-only, crash-safe, resume past sessions with /resume
• Sub-agent system — parent agents can spawn sub-agents for independent tasks, sync/async modes, linked aborts

---

Quick start

# Clone, create an env, and run
git clone https://github.com/boots-coder/lumen.git
cd lumen
conda create -n lumen python=3.11 -y && conda activate lumen
pip install httpx pydantic tiktoken rich prompt_toolkit
# Optional: web search tool, YAML-defined skills
pip install duckduckgo_search pyyaml
python chat.py

Or skip the wizard and launch directly:

# OpenAI
python chat.py --model gpt-4o --api-key sk-proj-...

# Anthropic
python chat.py --model claude-sonnet-4-6 --api-key sk-ant-...

# Local Ollama (no key needed)
python chat.py --model llama3.1 --base-url http://localhost:11434/v1

---

Core capabilities

Tool suite (invoked autonomously by the model)

Tool	Purpose
read_file	Read file content by line, with offset + limit pagination
write_file	Create a new file
edit_file	Edit an existing file (line-level replacement)
tree	Render the project directory (auto-filters node_modules etc.)
definitions	Extract every class / function / method from a file with line numbers
glob	Match file paths by pattern, sorted by modification time
grep	ripgrep-powered regex search with three output modes
bash	Execute shell commands
web_search	DuckDuckGo search (no API key required)
web_fetch	Fetch a web page and extract readable text
lsp	LSP code intelligence: go-to-definition / find-references / hover / symbol search
sub_agent	Spawn a sub-agent for an independent task (requires enable_subagents())

Extended Thinking

Three strategies automatically adapt to different models:

Model family	Mechanism
Anthropic	thinking: {"type": "enabled", "budget_tokens": N}
OpenAI o-series	reasoning_effort: "low"/"medium"/"high"
Other models	CoT instructions injected via the system prompt

The thinking budget adapts to context usage — as the window fills up, it shrinks automatically to leave room for the actual response.

Structured output

from pydantic import BaseModel

class CodeAnalysis(BaseModel):
    complexity: str
    issues: list[str]
    suggestions: list[str]

result = await agent.query("Analyze this code", schema=CodeAnalysis)
# result is a validated CodeAnalysis instance

Supports OpenAI json_schema / Anthropic tool-use trick / Gemini / generic JSON mode.

Command safety classifier

Replaces pure regex with feature scoring — a weighted analysis across four dimensions:

• Executable (weight 0.4) — rm 0.7, sudo 0.9, ls 0.0
• Flags (weight 0.2) — -rf 0.8, --force 0.3, --help 0.0
• Argument paths (weight 0.1) — / 0.9, /etc 0.4, ./src 0.0
• Command composition (weight 0.3) — curl | bash 0.9, subshell 0.3

Context-aware: rm -f is much riskier than -f on its own.

Auto-persisted transcripts

Every message is appended to a JSONL file — crash-safe:

~/.lumen/projects/{project-path}/{session-id}.jsonl

• Append-only writes — 100 ms buffer, async flush
• Fast recovery — reads metadata from the tail 64 KB, /resume restores in one step
• Session ID — UUID auto-generated, shown at startup

Sub-agent system

A parent agent can spawn an isolated sub-agent to parallelize work:

agent.enable_subagents()  # Register the sub_agent tool

# Or use it programmatically
from lumen import SubAgentConfig
result = await agent.subagent_manager.spawn(SubAgentConfig(
    prompt="Audit the security of the auth module",
    description="Security audit",
    run_in_background=True,  # Run asynchronously in the background
))

Isolation policy:

• File cache — cloned (no cross-contamination)
• Abort — parent/child linked (parent abort cascades to child; child abort does not affect parent)
• Session — independent conversation history
• Tools — inherited or customized

Dynamic session memory

• Every three turns, key facts (preferences / project / patterns / corrections) are auto-extracted
• TF-IDF keyword retrieval — no vector database required
• Jaccard similarity deduplication, persisted to a JSON file
• Relevant memories are automatically injected on the first turn

Context management

• Automatic token counting: precise per-turn counts with a live progress bar
• Automatic compaction: when the window is nearly full, history is compacted into a structured summary
• Manual compaction /compact: trigger any time, keeping the most recent N messages

Memory system (ENGRAM.md)

Four priority levels, auto-discovered and loaded:

/etc/lumen/ENGRAM.md      # System level (lowest priority)
~/.engram/ENGRAM.md       # User level
./ENGRAM.md               # Project level (team-shared)
./ENGRAM.local.md         # Local private (gitignored)

---

SDK usage

Basic usage

import asyncio
from lumen import Agent
from lumen.tools import (
    FileReadTool, FileWriteTool, FileEditTool,
    GlobTool, GrepTool, TreeTool, DefinitionsTool, BashTool,
)

async def main():
    agent = Agent(
        api_key="sk-...",
        model="gpt-4o",
        tools=[
            TreeTool(), DefinitionsTool(), FileReadTool(),
            GlobTool(), GrepTool(), BashTool(),
            FileWriteTool(), FileEditTool(),
        ],
        auto_compact=True,
        inject_git_state=True,
    )

    response = await agent.chat("Explain the overall architecture of this project")
    print(response.content)

asyncio.run(main())

Advanced: structured output + Thinking

from lumen import Agent, ThinkingConfig, ThinkingMode

agent = Agent(
    api_key="sk-ant-...",
    model="claude-sonnet-4-6",
    tools=[...],
    thinking=ThinkingConfig(
        enabled=True,
        budget_tokens=10000,
        mode=ThinkingMode.AUTO,
    ),
)

# Structured query
result = await agent.query("List every API endpoint", schema=APIEndpoints)

Advanced: persistent retry (CI scenario)

from lumen import Agent, PersistentRetryConfig

agent = Agent(
    api_key="sk-...",
    model="gpt-4o",
    tools=[...],
    persistent_retry=PersistentRetryConfig(
        enabled=True,
        escalation_threshold=5,
        fallback_models=["gpt-4o-mini", "claude-sonnet-4-6"],
        total_timeout_seconds=3600,
    ),
)

---

Slash commands

Command	Description
/help	Show help
/status	Token usage and session info
/mode code	Enter deep code-reading mode
/mode general	Return to general mode
/compact	Manually compact the context
/reset	Clear conversation history
/save	Save the session to a JSON file (transcripts are already auto-persisted)
/load	Restore a session from a JSON file
/resume	List recent sessions and pick one to resume
/config	Reconfigure the model and API key
/quit	Exit (or Ctrl+D)

---

Supported models

Provider	Example models	Notes
OpenAI	gpt-4o, gpt-4o-mini, o3-mini, o1	Auto-handles API differences for reasoning models
Anthropic	claude-sonnet-4-6, claude-opus-4-6, claude-haiku-4-5	Native API + thinking blocks
DeepSeek	deepseek-chat, deepseek-reasoner	OpenAI-compatible protocol
Ollama	llama3.1, qwen2.5, mistral, deepseek-r1	Run locally, no key needed
Other	Any OpenAI-compatible API	Custom base_url

---

Requirements

• Python 3.11+
• ripgrep (optional, used by the grep tool; brew install ripgrep)
• An API key for OpenAI / Anthropic / DeepSeek, or a local Ollama

Optional dependencies:

• duckduckgo_search — required for the web search tool
• pyyaml — required for YAML-formatted skill definitions

---

License

MIT