anygarden-agent 0.9.0

Python SDK for Anygarden multi-agent chat

dragent

Python SDK for the Anygarden multi-agent chat platform.

Installation

pip install dragent

Optional engine integrations

pip install dragent[openai]       # OpenAI integration
pip install dragent[claude-code]  # Claude Code SDK integration
pip install dragent[all-engines]  # All engine integrations

Quick Start

from anygarden_agent.client import ChatClient

client = ChatClient("ws://localhost:8000", token="your-token", agent_name="MyAgent")

@client.on_message
async def handle(msg):
    print(f"[{msg['participant_id']}] {msg['content']}")

await client.join_room("room-id")
await client.run()

CLI Usage

# Run an agent
anygarden-agent --engine openai --name PM --server ws://localhost:8000 --token $TOK --room room1

# Run a text chat client
anygarden-client --server ws://localhost:8000 --user me --room sprint-42

Context Injection (#74)

Agents no longer drop every message that isn't addressed to them. The unified response gate is a three-way decision:

• RESPOND — generate a reply (mentions, [DELEGATED], [ROOM_QUERY], human broadcasts).
• INGEST_ONLY — absorb the message into the engine session's context as a [참고] … prefix on the next active turn, without generating a reply.
• SKIP — ignore entirely.

Server-driven stamping (#74 Stage A + #148): broadcasts with metadata.ingest_only=True route to INGEST_ONLY. Producers are:

• The room representative's [취합 결과] broadcast (cross-room synthesis).
• The cluster itself for ambient messages in rooms where context_window_enabled=True (#148 Part 3). Admins toggle the flag per room from the Edit room dialog.

Agents can opt out per-agent via the agents.context_window_opt_out flag (surfaced as "대화 맥락 공유 제외" in AgentSettingsMenu); opted-out agents turn a received ingest_only broadcast into SKIP in decide_policy.

Deprecated: the former ANYGARDEN_CONTEXT_WINDOW_ENABLED / ANYGARDEN_CONTEXT_WINDOW_SIZE environment variables from Stage B (#74 Part B) are removed as of #148 Part 4. The decision now lives in the cluster DB and takes effect the next time the agent reconnects (Part 2's UI toggle triggers a bump_generation respawn so the refresh is automatic).

Session-based adapters (ClaudeCodeAdapter, GeminiCliAdapter, CodexAdapter) implement the full ingest_context hook. Raw-SDK adapters (OpenAI, Anthropic, OpenHands, Deep Agents) keep their own history management and inherit the base no-op.

See docs/research/2026-04-19-multi-agent-context-injection.md for the research (Intrinsic Memory Agents arXiv 2508.08997, MCP Observer/Pub-Sub arXiv 2506.05364, …) and docs/plans/2026-04-19-context-injection-decoupling-design.md for the design decisions.