leagent 1.1.3

LeAgent — Intelligent Office Agent Platform

LeAgent Backend

The Python service powering LeAgent — an enterprise-grade intelligent office automation platform that combines LLM-powered agents, a visual workflow engine, a rule engine, a typed tool system and multi-channel delivery in a single FastAPI application.

For a product-level overview of the platform, see the root README. For deployment instructions, see ../deploy/README.md. For contributor conventions, see ../AGENTS.md.

---

Desktop & thin install

• Default database is SQLite under LEAGENT_HOME/leagent.db (or DB_SQLITE_PATH). Use DB_DRIVER=postgresql+asyncpg (and related DB_* fields) for Postgres.
• Heavy stacks ship as optional extras: leagent[browser], leagent[office], leagent[rag], leagent[ocr], leagent[paddle] (legacy OCR). Admins can also install packs from Settings → Plugins (/api/v1/extensions).
• After editing pyproject.toml, run uv lock to refresh uv.lock.

---

Architecture

The default deployment is a single-process async monolith (leagent/main.py). All modules — agent runtime, workflow engine, tool executor, LLM routing, session management — run in one FastAPI process backed by SQLite. This is suitable for single-machine and small-team use.

                 HTTP / WS / SSE
 clients ─────────────────────────▶  FastAPI (leagent.main:app)
                                       │  auth · CORS · uploads · SSE streaming
                                       │
           ┌───────────────────────────┼──────────────────────────────┐
           ▼                           ▼                              ▼
    QueryEngine                  Workflow Engine               LLM Service
    (agent runtime)              (YAML + visual)               (tier routing)
           │                           │                              │
           ▼                           ▼                              ▼
    ToolExecutor ◄── ToolRegistry ──► Tool.<name> workflow nodes     Providers
    (80+ tools)      (single source   (auto-generated palette)       (DeepSeek,
                      of truth)                                       OpenAI,
                                                                      Anthropic,
                                                                      DashScope,
                                                                      Ollama, vLLM)
           │
           ▼
    SQLite (default) + local filesystem under LEAGENT_HOME
    Optional: PostgreSQL, Milvus (vector memory), MinIO (object storage)

An optional gateway entrypoint exists at leagent/apps/gateway/ for teams that outgrow the monolith. See ../AGENTS.md for the modular-monolith discussion and ../docs/deployment/MODULAR_MONOLITH.md for the full topology guide.

---

Table of contents

• Highlights
• Tech stack
• Directory layout
• Requirements
• Installation
• Configuration
• Running the server
• CLI
• Database & migrations
• Testing
• Lint, format & type-check
• Extending the platform
• Local Desktop State
• Observability
• Troubleshooting

---

Highlights

• Async FastAPI application with 20+ versioned routers under leagent/api/v1/ (auth, chat, tasks, tools, rules, cron, skills, channels, templates, webhooks, MCP, files/folders/documents, metrics, health, ...). The workflow engine owns its own namespace at /api/v1/workflow/* (CRUD, prompts, WebSocket progress, node admin). Incubating v2 endpoints live under leagent/api/v2/.
• QueryEngine session orchestrator (leagent/agent/query_engine.py) — session-scoped state, streaming submit_message() async generator, DI-friendly QueryDeps seam, explicit Terminal / Continue transitions. The legacy AgentController is a compat shim over QueryEngine.
• Hybrid agent runtime in leagent/agent/ — a ReAct controller, a Plan-and-Execute planner, a tool executor, sub-agent spawning (fork_subagent, ScriptAgentTool), and abort-safe runs.
• Tool system (leagent/tools/) with 80+ tools across 15 categories, aliases, semantic input validation, concurrency classes (concurrent_safe / read_only / destructive), result-size budgets, live activity strings and interrupt behaviour.
• Layered prompt management (leagent/prompts/) — 8-layer PromptBuilder, file-system PromptRegistry, per-layer + global budget enforcement, SHA-256 fingerprinting, provider-aware rendering (Anthropic cache_control).
• Workflow engine (leagent/workflow/) — single canonical document schema, typed IO/Schema node contracts, per-node caches, in-memory prompt queue, and a dedicated FastAPI surface at /api/v1/workflow/* with WebSocket progress streams.
• Rule engine (leagent/rules/) supporting 8 primitives: compare, date_range, threshold, contains_all, date_diff, regex_match, cross_validate, llm_judge.
• Context assembly (leagent/context/) — source-driven pipeline with ContextManager, ContextRecipe, 12 ContextSource implementations, cost-function budget minimisation, and provider-specific strategies (e.g. strategies/deepseek.py).
• Service layer (leagent/services/) with database, auth (+ HMAC-signed file URLs), chat, session (SessionManager + TieredSessionStore), code execution (subprocess sandbox), coding projects (scaffolding + git init), canvas/gen-UI, file management, and TaskManager (lifecycle, abort, byte-offset output streaming).
• LLM abstraction (leagent/llm/) with tier-based routing and providers: DeepSeek (auto-aliased as tier1/tier2 when DEEPSEEK_API_KEY is set), OpenAI, Anthropic, DashScope, Ollama, vLLM.
• Cognitive agent memory (leagent/memory/) — AgentMemory facade with episodic, semantic, and procedural stores, RetrievalPipeline (hybrid search + reranking), WorkingScratchpad, RecallHandle, and micro/auto-compaction.
• Multi-channel delivery (leagent/channels/) — web, console, api, dingtalk, feishu, wechat_work, weixin.
• Agent Skills v1.0 — SKILL.md-first skill system (leagent/skills/) with progressive disclosure, cross-agent discovery, and a pluggable HTTP registry. See the Skills guide.
• Batteries included — cron scheduler (APScheduler + croniter), MCP client, structured JSON logging, Prometheus metrics, SSE streaming, WebSocket support, polite outbound HTTP (robots.txt, per-host locking).

---

Tech stack

Area	Libraries
Web framework	FastAPI, Uvicorn, Gunicorn, sse-starlette, websockets
Data layer	SQLModel, SQLAlchemy (async), Alembic, aiosqlite, sqlite-vec; optional asyncpg for PostgreSQL
Validation	Pydantic v2, pydantic-settings, jsonschema, email-validator
LLM	openai, anthropic, tiktoken, httpx, aiohttp
Docs & data	pymupdf, pdfplumber, python-docx, openpyxl, pandas, xlsxwriter, reportlab, pillow
Web automation	playwright (optional extra leagent[browser])
Vector & objects	pymilvus, minio (optional — only needed for RAG memory / object storage)
Scheduling	apscheduler, croniter, pytz
Security	python-jose[cryptography], passlib[bcrypt]
CLI & logging	click, rich, structlog
Serialization	orjson, pyyaml, jinja2, aiofiles, aiosmtplib
Tooling (dev)	pytest, pytest-asyncio, pytest-cov, ruff, mypy

See pyproject.toml for the authoritative dependency list.

---

Directory layout

backend/
├── alembic.ini                  Alembic configuration (DSN read from env)
├── pyproject.toml               Package metadata, deps, Ruff/Mypy/Pytest config
├── README.md                    (this file)
├── tests/                       Pytest suite (30+ modules)
│   ├── conftest.py
│   ├── fixtures/
│   └── test_*.py                agent, workflow, rules, tools, api, channels, …
└── leagent/
    ├── main.py                  ASGI app factory (FastAPI lifespan, routers, middleware)
    ├── server.py                Uvicorn / Gunicorn bootstrap
    ├── agent/                   QueryEngine (primary) + ReAct/Plan-Execute runtime
    │   ├── query_engine.py      Session orchestrator (submit_message, SDKMessage stream)
    │   ├── query.py · state.py  Per-turn loop + mutable state
    │   ├── transitions.py       Terminal / Continue explicit transition values
    │   ├── deps.py              QueryDeps DI protocol (call_model, micro/autocompact)
    │   ├── tool_use_context.py  Runtime context bundle for tool execution
    │   ├── controller.py        AgentController (compat shim → QueryEngine)
    │   ├── planner.py           Plan-and-Execute path
    │   ├── script_agent.py      Scoped compute / subprocess engine builder
    │   ├── subagent.py          Sub-agent delegation (fork_subagent)
    │   ├── coding_agent.py      Coding-project agent specialisation
    │   ├── runtime_profile.py   Agent runtime profiles
    │   └── hooks.py · recovery.py · executor.py · base.py
    ├── alembic/                 Migration scripts (inside leagent package)
    ├── api/
    │   ├── router.py            Top-level router aggregator
    │   ├── middleware.py        Auth, request IDs, rate limits, metrics
    │   ├── deps.py              FastAPI dependencies (DB session, current user, …)
    │   ├── v1/                  24+ routers (see list below)
    │   └── v2/                  Next-gen endpoints (incubating)
    ├── apps/                    Optional split-deploy entrypoints
    │   └── gateway/             API gateway (for modular-monolith topology)
    ├── bootstrap/               Canonical tool + workflow-node startup
    │   └── tools.py             bootstrap_tools, register_script_agent_tool
    ├── channels/                Delivery adapters + manager + registry + renderer
    │   ├── base.py · manager.py · registry.py · renderer.py
    │   └── web/ · console/ · api/ · dingtalk/ · feishu/ · wechat_work/ · weixin/
    ├── chat_workflow/           Chat-workflow orchestration
    ├── cli/                     Click command groups (23 modules, see CLI section)
    ├── config/                  Pydantic settings, defaults, env loading
    ├── console/                 Rich-based console helpers
    ├── context/                 Source-driven prompt assembly
    │   ├── manager.py · recipe.py · budget.py · types.py · ledger.py · cache.py
    │   ├── file_state.py        Session-scoped file read cache
    │   ├── working_set.py · compression.py · session_compression.py
    │   ├── sources/             12 ContextSource implementations
    │   └── strategies/          Provider-specific strategies (e.g. deepseek.py)
    ├── cron/                    APScheduler-based job manager
    ├── exceptions/              Domain error hierarchy
    ├── extensions/              Optional plugin extensions
    ├── initial_setup/           First-run seeding (admin, default channels, demo data)
    ├── llm/                     LLM service, tiered router, provider adapters
    │   ├── service.py · registry.py · router.py · base.py · provider_config.py
    │   └── providers/           DeepSeek, OpenAI, Anthropic, DashScope, Ollama
    ├── mcp/                     Model Context Protocol client
    ├── memory/                  Cognitive three-store agent memory
    │   ├── agent_memory.py      AgentMemory facade + RecallHandle
    │   ├── episodic.py · semantic.py · procedural.py   Three stores
    │   ├── recall.py            RetrievalPipeline (hybrid search + reranking)
    │   ├── embeddings.py · vector.py   Embedding + Milvus wrappers
    │   ├── working_scratchpad.py   Ephemeral tool history + scratchpad
    │   ├── compact.py · compaction.py   Micro/auto-compaction
    │   └── manager.py · short_term.py · working.py · long_term.py   Legacy paths
    ├── prompts/                 Layered prompt management
    │   ├── builder.py · registry.py · render.py · fingerprint.py · context.py · types.py
    │   └── templates/           .md files with YAML front-matter
    │       ├── default_agent.md · script_agent.md · coding_agent.md · subagent.md
    │       ├── rule_judge.md · compact_summariser.md
    │       └── policies/        Composable policy snippets
    ├── rules/                   Rule engine (8 primitives) + YAML definitions
    ├── schema/                  App-level shared Pydantic schemas
    ├── scripts/                 Ops scripts (e.g. run_migrations with advisory lock)
    ├── services/                Core service layer
    │   ├── database/            SQLModel ORM + repository helpers
    │   ├── auth/                JWT, RBAC, signed URLs
    │   ├── session/             SessionManager + TieredSessionStore
    │   ├── chat/                Chat service + daily greetings
    │   ├── cache/               Cache abstractions
    │   ├── canvas/              Gen-UI rendering + HTML bundling
    │   ├── code_execution/      Subprocess sandbox (workspace, runner, rlimits)
    │   ├── coding_projects/     Scaffold templates (vite-react, fastapi, vanilla-html)
    │   ├── compact/             Compaction service
    │   ├── diagnostics_parsers/ Shell output diagnostics parsing
    │   ├── event/               In-process event bus
    │   ├── execution/           Execution management
    │   ├── file_manager/        File operations
    │   ├── file_processing/     Document processing pipelines
    │   ├── file_store/          File persistence
    │   ├── gen_ui/              GenUI schema, print renderer, PPTX renderer
    │   ├── job_queue/           Background job management
    │   ├── python_env/          Python environment resolution
    │   └── variable/            Variable management
    ├── skills/                  Agent Skills v1.0 (SKILL.md loader, manager,
    │   │                          discovery, HTTP registry, bundle packaging)
    │   └── builtin/             Built-in skills (document-processor, data-analyzer, …)
    ├── tasks/                   Task system (handlers/)
    ├── tools/                   80+ tool implementations across 15 categories
    │   ├── base.py              BaseTool (aliases, validation, budgets, activity, sandbox)
    │   ├── registry.py          ToolRegistry (alias lookup, search, deny rules)
    │   ├── executor.py          ToolExecutor (concurrency, permissions, abort)
    │   ├── _data/               ArtifactRef, TabularSchema, streaming primitives
    │   ├── _sandbox/            PathSandbox (filesystem allow-list enforcement)
    │   ├── doc/ web/ data/ gen/ chart/ image/ code/ db/
    │   ├── canvas/ project/ coding_project/ skills/ workflow/
    │   ├── integration/ util/
    ├── utils/                   Shared helpers (CJK fonts, etc.)
    └── workflow/                Workflow engine + YAML parser + node implementations
        ├── engine/              Runner, executor, graph, caching, progress
        ├── io/                  Schema contracts, bridge, serializer
        ├── layout/              Visual layout engine
        ├── nodes/               Base + loader + tool_factory + builtin/
        ├── queue/               Prompt queue (in-memory)
        └── server/              Workflow-specific API + WebSocket

Shared primitives — leagent_core/

Cross-cutting primitives live in the sibling leagent_core/ package:

Module	Purpose
auth/	JWT helpers, request context
cache/	Cache abstractions
circuit/	Async circuit breaker
db/	Async engine factory + repository helpers
events/	In-process event bus
queue/	Queue abstractions
ratelimit/	Token-bucket limiter
schema/	Shared Pydantic schemas
telemetry/	OTel setup + structlog + traceparent

API routers (v1)

activities, auth, canvas, channels, chat, cron, documents, files, flows, folder_items, folders, health, mcp, metrics, models, rules, settings_mail, skills, stats, tasks, templates, tools, users, webhooks.

---

Requirements

• Python 3.11+ (3.12 recommended)
• SQLite (bundled with Python — the default, zero-config database)
• Playwright browsers — install via playwright install chromium to enable the tools/web/* tools (optional extra leagent[browser])

Optional (for advanced features):

• PostgreSQL 15+ — use DB_DRIVER=postgresql+asyncpg for production-grade persistence
• Milvus 2.4+ — required for vector-backed agent memory (episodic/semantic/procedural recall)
• MinIO (or any S3-compatible store) — for object storage
• NVIDIA GPU + NVIDIA Container Toolkit for local vLLM inference

---

Installation

cd backend

# Using uv (recommended)
uv sync --extra dev
uv run playwright install chromium   # if you use browser tools

# Or using pip
python -m venv .venv
source .venv/bin/activate            # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
playwright install chromium

This exposes the leagent console entry point declared in pyproject.toml:

[project.scripts]
leagent = "leagent.cli.main:cli"

---

Configuration

Configuration is loaded via pydantic-settings. The canonical template is ../deploy/.env.example; for local development you can place a .env file in backend/ or export variables in your shell.

Key variables:

Variable	Description
LEAGENT_SECRET_KEY	Required. JWT / signing secret. Generate with openssl rand -hex 32.
DB_SQLITE_PATH	SQLite file path (defaults to ~/.leagent/leagent.db).
DATABASE_URL	SQLAlchemy async DSN for PostgreSQL, e.g. postgresql+asyncpg://user:pass@localhost:5432/leagent.
DEEPSEEK_API_KEY	DeepSeek provider key (auto-aliases tier1 → v4-pro, tier2 → v4-flash).
OPENAI_API_KEY, ANTHROPIC_API_KEY	Other LLM provider keys.
LLM_TIER1_API_KEY / LLM_TIER1_ENDPOINT	Explicit primary (reasoning) tier.
LLM_TIER2_API_KEY / LLM_TIER2_ENDPOINT	Explicit secondary (fast / cheap) tier.
LEAGENT_LOG_LEVEL	DEBUG / INFO / WARNING.
LEAGENT_HOME	Root for data, knowledge, uploads, logs (default ~/.leagent).

Settings are loaded by leagent.config at startup and are also accessible via:

leagent config show
leagent env

---

Running the server

Hot-reload (dev)

uvicorn leagent.main:app --reload --host 0.0.0.0 --port 7860
# or equivalently:
leagent app --reload

Open:

• OpenAPI UI: http://localhost:7860/docs
• ReDoc: http://localhost:7860/redoc
• Health: http://localhost:7860/health
• Prometheus metrics: http://localhost:7860/metrics

Production

The Docker image runs Gunicorn with Uvicorn workers (see ../deploy/supervisord.conf and ../deploy/entrypoint.sh). To reproduce locally:

gunicorn leagent.main:app \
  -k uvicorn.workers.UvicornWorker \
  -w 4 \
  -b 0.0.0.0:7860 \
  --access-logfile - --error-logfile -

---

CLI

The leagent CLI groups operational commands. A selection:

leagent --help                          # top-level help

# Lifecycle
leagent init                            # initialize config, DB, default admin
leagent app [--reload] [--port 7860]    # run the API server
leagent daemon start|stop|status        # supervisor-style daemon mode
leagent clean                           # clean caches / temp artefacts

# Configuration & diagnostics
leagent config show
leagent env

# Providers, channels, chats
leagent providers list|add|remove
leagent channels   list|add|remove
leagent chats      list|show|delete

# Workflows, rules, templates
leagent workflows list|run|validate
leagent templates list|apply
leagent rules      list|test

# Tasks, cron, skills, webhooks
leagent tasks      list|show|kill|output
leagent cron       list|add|remove|run
leagent skills     list|show|init|validate|lint|migrate|install|uninstall|enable|disable|search
leagent webhooks   list|add|test

Implementations live under leagent/cli/ — one module per command group (e.g. tasks_cmd.py, workflows_cmd.py).

---

Database & migrations

Models are SQLModel subclasses under leagent/services/database/. The project uses Alembic for schema migrations.

# Apply all migrations (runs automatically on container start via entrypoint.sh)
cd backend
alembic upgrade head

# Generate a new migration after changing a model
alembic revision --autogenerate -m "describe the change"

# Downgrade one step
alembic downgrade -1

# Full history
alembic history --verbose

The Alembic config (alembic.ini) reads the DSN from DATABASE_URL. For local SQLite usage the DSN defaults to a file under LEAGENT_HOME.

---

Testing

# Full suite
cd backend
pytest

# Using uv
uv run pytest tests/ -v

# Single module / single test
pytest tests/test_query_engine.py
pytest tests/test_workflow.py::test_parallel_node

# Filter by expression
pytest -k "rule and not slow"

# Coverage
pytest --cov=leagent --cov-report=term-missing --cov-report=html

Pytest is configured in pyproject.toml:

• asyncio_mode = "auto" — no need to decorate async tests.
• testpaths = ["tests"].
• Deprecation warnings are filtered out.

Fixtures for DB sessions, temporary file stores, an in-memory tool registry and stubbed LLM providers live in tests/conftest.py and tests/fixtures/.

---

Lint, format & type-check

ruff check .               # lint
ruff check . --fix         # auto-fix where safe
ruff format .              # format (Black-compatible)

mypy leagent               # strict type-checking (see pyproject.toml)

Ruff is configured with line-length = 120, target py311, and rule groups E, W, F, I, N, UP, B, SIM, TCH. Mypy runs in strict mode with the Pydantic plugin.

---

Extending the platform

Add a new tool

1. Create leagent/tools/<category>/<name>.py.

2. Subclass BaseTool and implement:

   from leagent.tools.base import BaseTool, ToolResult
   
   class InvoiceSummarizer(BaseTool):
       name = "invoice_summarizer"
       aliases = ["invoice_sum", "summarize_invoice"]
       description = "Summarize an invoice PDF into structured fields."
       category = "doc"
       concurrency_class = "read_only"
       result_budget_bytes = 256_000
       activity = "Summarizing invoice..."
   
       path_params = ("file_path",)           # read-only access
       output_path_params = ("output_path",)  # allows file creation
   
       async def validate_input(self, payload: dict) -> None:
           if "file_id" not in payload:
               raise ValueError("file_id is required")
   
       async def run(self, payload: dict, ctx) -> ToolResult:
           ...
   

3. Export it from the category __init__.py so the registry picks it up on startup.

4. Declare sandbox path parameters. If your tool reads or writes files, set path_params (read-only) and/or output_path_params (allows creation) on the class. BaseTool.run() will validate them against PathSandbox before execution. For tools with nested path structures (e.g. arrays of file objects), override _enforce_path_sandbox(params, context).

5. Add a unit test under tests/test_tools/ (or a dedicated module) covering validate_input, the happy path and one failure path.

6. Verify: pytest -k invoice_summarizer && leagent tools list | grep invoice_summarizer.

Add a new API route

1. Create a router module under leagent/api/v1/<feature>.py exposing an APIRouter.
2. Register it in leagent/api/router.py.
3. Add request/response schemas in leagent/schema/ (or co-located) — always Pydantic v2.
4. Protect the route with Depends(get_current_user) / role dependencies from leagent/api/deps.py.

Add a new channel, rule type, LLM provider or workflow node

• Channel — subclass BaseChannel under channels/<name>/, register in channels/registry.py.
• Rule type — implement the evaluator in rules/ and register it in the rule registry.
• LLM provider — add an adapter in llm/providers/ and list it in cli/providers_cmd.py.
• Workflow node — add a node implementation under workflow/nodes/builtin/ and register it in the node loader.

See ../AGENTS.md for style conventions and coding guidelines.

---

Local Desktop State

LeAgent is a local-first desktop app. The backend keeps enough state to power chat, tool execution, files and background jobs on one machine; it is not a multi-tenant account system.

Persistent Data

Area	Purpose
chat_sessions	Conversation history, session metadata and authorized local paths
files	Session uploads, knowledge documents and signed preview/download metadata
coding_projects	Local coding project records backed by files under CODING_PROJECTS_ROOT
tasks	Background job lifecycle, abort state and file-backed output logs
agent memory tables	Episodic, semantic and procedural recall for local agent sessions

Default installs use SQLite under LEAGENT_HOME; optional PostgreSQL is for heavier single-machine or team deployments, not for workspace isolation.

Local Access Boundaries

New routes should be scoped around sessions, files, tools and local projects. For filesystem access, use PathSandbox and the session authorized-path APIs instead of adding account or workspace checks. File preview/download routes use short-lived HMAC-signed URLs so the desktop UI can open artifacts without exposing raw paths.

---

Observability

• Metrics -- GET /metrics emits Prometheus counters/histograms for HTTP latency, error rates, LLM token usage, tool execution durations and task queue depth.
• Logs -- structlog outputs JSON with request_id, task_id, tool, latency_ms, level, event. The log level is controlled by LEAGENT_LOG_LEVEL.
• Task output -- TaskManager streams handler output to file-backed logs; the API exposes GET /api/v1/tasks/{id}/output?offset=<bytes> for incremental polling.
• Health -- GET /health reports DB connectivity (and optional Redis/Milvus/MinIO when configured).

---

Troubleshooting

Symptom	Likely cause / fix
ModuleNotFoundError: leagent	Virtualenv not active, or uv sync --extra dev / pip install -e ".[dev]" was not run in backend/.
sqlalchemy.exc.OperationalError on startup	Check DATABASE_URL or DB_SQLITE_PATH. For PostgreSQL: verify the server is reachable.
playwright._impl._errors.Error: Executable doesn't exist	Run playwright install chromium inside the active venv.
401 on every endpoint	LEAGENT_SECRET_KEY changed between restarts — existing tokens are invalidated. Re-login or pin the secret.
LLM calls hang or 429	Check provider keys (DEEPSEEK_API_KEY, OPENAI_API_KEY, etc.), tier endpoint config, or provider rate limits.
DeepSeek balance errors	Run leagent providers list to check balance. Verify DEEPSEEK_API_KEY is valid.
Alembic Target database is not up to date	Run alembic upgrade head (the entrypoint does this automatically in Docker).

For anything else, check the structured logs:

tail -f ~/.leagent/logs/leagent.log            # local dev
docker compose logs -f leagent                 # Docker

---

License

This codebase is part of the LeAgent platform and is covered by the project's top-level license (see ../LICENSE).