English only.

OpenManus

A conversation-first agent runtime with persistent sandbox computers, live observability, and practical team workflows.

This fork focuses on real daily usage: long-running conversations, isolated workspaces, runtime controls, and a UI that makes the agent’s work visible and manageable.

Why This Repo

OpenManus here is built for people who want an agent they can actually run, inspect, and improve:

• Persistent conversations with replayable history
• Per-conversation sandbox + workspace isolation
• Live tool traces, terminal output, browser screenshots, and token counts
• Runtime controls for processes, ports, and spawned containers
• Optional RL policy scaffold compatible with OpenManus-RL workflows

What’s Implemented

• Auth + admin bootstrap (first signup becomes admin)
• Session-based web auth and admin settings UI/API
• Conversation model with grouped tasks and follow-up continuity
• Per-conversation files under workspace/conversations/<conversation_id>/
• Reused sandbox computer per conversation
• Mid-task user messages to running agents
• Live SSE lifecycle stream (thoughts, tools, terminal, browser, usage)
• Process/container visibility and kill controls (with protection rules)
• Better completion artifacts and warnings (including missing PDF hints)
• OpenHands-style skill loading support
• RL integration scaffold and policy export helper

Latest Additions

• apply_patch_editor as the primary code-editing tool for safer, atomic patch updates
• Live and final GitHub-style change summary (files changed, +added, -deleted)
• Richer tool execution cards with better status and file-change visibility
• Improved conversation reliability around long runs and follow-up continuity
• Runtime context observability (requested window, received window, usage ratio, auto-compress status)
• Dynamic UI Connection Profiles: Active model and provider connection settings configured directly in the Admin UI dynamically override static config/config.toml settings across backend worker tasks, agent execution loops, and semantic memory indexing (AgentMemory).
• LM Studio API v1 Integration: Native /api/v1 model management (/models/load, /models/unload, /models) with resilient HTTP 404 fallback to /api/v0 and automatic requested context window slot synchronization (128k default).
• Obsidian Graph & Wikilink Synchronization: Automatic bidirectional syncing between workspace markdown files and Obsidian note graphs (auto_sync_obsidian_notes), featuring path-qualified [[wikilink]] resolution ([[projects/Overview]]), duplicate title ambiguity detection, and diff-based edge preservation.

Agent Loop Architecture & Reliability Refactoring (vs. Upstream)

• Structural Termination: Replaced fragile regex finish detection (_INCOMPLETE_RE, _STRONG_FINAL_RE) with strict structural termination where the LLM must call the terminate tool (status, summary).
• Typed Error Handling & Auto-Retry: Replaced string-sniffing (result.lower().startswith("error")) with typed ToolResult.is_error. When a tool fails, the loop automatically retries with concrete _error_context injected into the tool input.
• State Machine & Lifecycle Tracking: Introduced explicit AgentPhase lifecycle tracking (PLAN -> ACT -> OBSERVE -> VERIFY -> DONE) and real-time lifecycle event emissions (agent:lifecycle:phase, reason, observe).
• Robust Stuck Detection: Enhanced is_stuck() with MD5 content hashing across whitespace-normalized turns and repeated exact tool-call batch detection.
• Smart Context Compression: Replaced naive 220-character truncation with structured summarization that preserves pinned artifacts (pinned_context such as file paths and diffs).
• Pydantic V2 & Modernization: Migrated core schema, tools, and agent models to clean Pydantic v2 ConfigDict(...) and resolved silent exception swallowing.

Architecture

• Backend API: server/api.py (FastAPI + SSE)
• Worker runtime: server/tasks.py (Celery)
• Agent core: app/agent/manus.py, app/agent/toolcall.py
• LLM compatibility/runtime: app/llm.py
• Sandbox lifecycle: app/sandbox/conversation.py
• Frontend: frontend/ (React + Vite)
• Persistence: PostgreSQL
• Event transport: Redis Streams

Quick Start (Docker)

1. Copy config:

cp config/config.example.toml config/config.toml

2. Set your model endpoint in config/config.toml.

3. Start:

docker compose up --build

Or use the project helper:

make build

make build runs a clean Docker builder prune and then starts the stack with docker compose up -d --build.

4. Open:

• Frontend: http://localhost:3000
• API: http://localhost:8000

Local Python Start

uv venv --python 3.12
source .venv/bin/activate
uv pip install -r requirements.txt
python main.py

Configuration

Main file: config/config.toml

Important sections:

• [llm] model, endpoint, key, token limits (http://127.0.0.1:1234/v1 by default for local LM Studio)
• [sandbox] runtime limits and network/socket access
• [agent] max steps and tool-call behavior
• [rl] optional policy integration

[!NOTE] Dynamic Profile Overrides: While config/config.toml sets the initial static defaults on startup, any connection profile changes (style, base_url, model, api_key) saved in the Admin Settings UI (/admin) are persisted to PostgreSQL (app_settings) and serve as the authoritative active connection for all running tasks, background workers, and local FAISS/keyword memory embeddings.

Example RL toggle:

[rl]
enabled = true
policy_mode = "rl"
policy_path = "research/openmanus-rl/artifacts/policy/latest/policy.md"
metadata_path = "research/openmanus-rl/artifacts/policy/latest/metadata.json"

RL Scaffold

This repo includes a clean path for OpenManus-RL style integration:

• research/openmanus-rl/
• app/agent/policy_loader.py
• scripts/export_policy.py

Example:

python scripts/export_policy.py \
  --policy-file /path/to/policy.md \
  --model qwen3.6-35b-a3b \
  --benchmark gaia \
  --run-id exp-2026-05-14

Semantic Memory with FAISS

OpenManus features local, zero-dependency SQLite-based memory persistence (AgentMemory). You can optionally enable local FAISS vector semantic search to enhance memory recall when query wording differs from saved memories:

[agentmemory]
enabled = true
vector_backend = "faiss" # "none" or "faiss"
embedding_provider = "openai_compatible"
embedding_model = "text-embedding-nomic-embed-text-v1.5"
hybrid_search = true
vector_weight = 0.65
keyword_weight = 0.35

When enabled, memories are indexed locally in FAISS alongside SQLite FTS5/BM25 keyword indices, returning weighted hybrid search results. If embedding generation or FAISS lookup fails, recall gracefully falls back to SQLite keyword search.

DeepSpec Research Path

OpenManus includes an opt-in research integration scaffold for speculative decoding experiments using DeepSeek's DeepSpec framework:

• See research/deepspec/README.md for workflow and evaluation instructions.
• Run scripts/prepare_deepspec_research.sh to clone or update the upstream repository.

Note: DeepSpec is strictly experimental and separate from standard runtime execution or default make build container generation. Target cache preparation requires high-capacity storage (e.g. ~38 TB for Qwen3-4B).

Contributing

Contributions are very welcome.

• Open an issue for bugs or feature proposals
• Keep PRs focused and easy to review
• Add verification steps in your PR description

Maintainer Commitment

I actively monitor this repo and I will respond to pull requests quickly. If your PR is blocked, tag me in the thread and I’ll help unblock it fast.

Suggested Team Flow

1. Create feature branches from main
2. Open PRs early (draft PRs are welcome)
3. Merge after review + verification

Runtime Notes

• Sandboxes are scoped to conversation lifecycle.
• Protected system processes/containers are not killable by runtime controls.
• Deleting a conversation clears associated task records, streams, and sandbox state.

License

MIT (same as project root license).