[Feature]: Graph-blackboard shared state plugin: persistent cross-agent knowledge base
メンテナーはふだん 1 日以内に返信
まだ誰も着手していません。
評価
- 難易度
- 5/5
- 見積もり時間
- 1週間以上
- 初心者へのやさしさ
- 35/100
- issue の種類
- 機能追加
- 明瞭さ
- おおむね明確
- 活発さ
- 活発
- 技術スタック
- typescript
- 領域
- ai-infra-agents, cli, databases
調査の方向性
Start with examples/hello-mcode-mcp and the MiniMax-Code-Plugins contract, especially plugin.json, optional mcp.json, and the stdio MCP server shape. Define the plugin boundaries and storage approach before implementing. Done means parallel agents can write and query shared entries, data survives reopening a session, and uninstalling the plugin leaves core functionality intact.
索引モデルが issue の本文から書いたものです。
説明
Product or interface
CLI - interactive TUI
Use case and problem
MiniMax Code already supports subagent dispatch (the task tool, with per-dispatch model and effort — confirmed by maintainers in #197). But each subagent's context is isolated; there is no unified, queryable, persistent shared state:
- Parallel subagents rediscover the same information independently; results cannot be reused across agents.
- The main agent acts as a manual relay, aggregating subagent outputs one by one — large context overhead on long tasks.
- Knowledge accumulated during collaboration is discarded when the session ends; the next task starts from zero.
This proposal targets that gap: cross-agent shared state and knowledge reuse. #197 requested multi-agent collaboration; this issue focuses only on the shared-blackboard part.
Desired behavior
Provide a "graph blackboard" shared state as a plugin, exposing read/write via an MCP server, without touching the core pipeline:
- After installing via /plugins, both the main agent and task-dispatched subagents can write structured entries (findings, conclusions, todos) to the shared blackboard through MCP tools, and query entries written by other agents by keyword/tag.
- The blackboard is scoped per project/workspace; data lives in a local single-file graph database — no network, no upload.
- Data persists across sessions, so later tasks keep accumulating into a project-level knowledge base.
Example acceptance criteria:
- In one session, dispatch two parallel subagents; after A writes entry X, B can query X without the main agent relaying it.
- X stays queryable after reopening the session.
- Uninstalling the plugin does not affect core functionality.
Implementation vehicle: a plugin conforming to the MiniMax-Code-Plugins contract (plugin.json + optional mcp.json, stdio MCP server); see examples/hello-mcode-mcp.
Platform
Multiple platforms
Alternatives and additional context
Prior art — shared state / blackboard is already first-class in mainstream multi-agent frameworks:
- MetaGPT (70k stars, MIT, ICLR 2024): global message pool = blackboard + publish/subscribe; roles write structured outputs to a shared pool and subscribe by type.
- CrewAI (59k stars, MIT): official shared memory (short-term / long-term / entity), read and written by all agents in the crew.
- AutoGen Magentic-One (61k stars, Microsoft): the Orchestrator keeps a shared Task/Progress Ledger as the team's state core and dispatches/re-plans from it.
- Letta / MemGPT (25k stars): official multi-agent shared memory — memory blocks attach to multiple agents with a live synchronized view.
- Graphiti / Zep (31k stars, Apache-2.0): real-time knowledge graphs as agent memory — evidence that graphs are a production-proven carrier for agent memory.
Proposer's own reference implementation: https://github.com/wufufu770/d2d — my personal project, work-in-progress; it validated the feasibility of "a shared graph-database blackboard as the single source of shared state" in parallel multi-agent scenarios. Included for mechanism reference only, not as a maturity signal.
Storage choice: Kuzu (MIT embedded property graph database; single-file, no server, built-in vector/full-text search). Disclosure: its upstream repository was archived 2025-10; as a local embedded library the risk is acceptable. Also open to abstracting the storage layer behind a replaceable interface — FalkorDB (actively maintained, used by Graphiti) is a candidate backend.
Alternative shape: could also be an experimental core feature behind a config flag, but the plugin shape is least invasive to the core and independently iterable — the plugin route is preferred.
- 主要言語
- TypeScript
- スター
- 2.1k
- フォーク
- 271
- 平均マージ
- 9時間 55分
- マージ済み PR（30日）
- 52
環境構築
- Dockerfile・Docker Compose ファイルなし
- プルリクエストのテンプレートあり
- コントリビューションガイドを読む
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
MiniMax-AI/minimax-code のほかの issue
-
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
MiniMax-AI/minimax-code#507 ·
メンテナーはふだん 1 日以内に返信
-
bug needs-triage tui
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
MiniMax-AI/minimax-code#481 · コメント 1 件 ·
メンテナーはふだん 1 日以内に返信
-
bug cli documentation needs-triage
難易度 2/5 1〜3時間 初心者へのやさしさ 84/100
MiniMax-AI/minimax-code#299 · コメント 3 件 ·
メンテナーはふだん 1 日以内に返信
-
desktop
難易度 5/5 1週間以上 初心者へのやさしさ 15/100
MiniMax-AI/minimax-code#552 ·
メンテナーはふだん 1 日以内に返信
-
難易度 4/5 3〜5日 初心者へのやさしさ 42/100
MiniMax-AI/minimax-code#551 ·
メンテナーはふだん 1 日以内に返信
MiniMax-AI/minimax-code の issue をすべて見る
似ている issue
-
bug
難易度 2/5 1〜3時間 初心者へのやさしさ 82/100
scaleway/ultraviolet#7089 ·
メンテナーはふだん 2 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 67/100
tencent-connect/dsh-qqbot#81 ·
メンテナーはふだん 1 日以内に返信
-
[Bug]: Pen taps on sidebar rows start a drag and lose their click対応中かも @makhnatkin が今日担当しました。 オープン
難易度 2/5 1〜3時間 初心者へのやさしさ 62/100
メンテナーはふだん 1 日以内に返信
-
bug
難易度 2/5 1〜3時間 初心者へのやさしさ 66/100
clangd/vscode-clangd#1009 ·