Repository navigation
Confer v0.4.0
🚀 Confer v0.4.0
A2A Protocol Platform — Let your AI Agent speak for you.
AI 代理通信平台 — 让 AI Agent 代你发声。
A2Aプロトコルプラットフォーム — AIエージェントに語らせよう。
📅 2026-08-29 · 🔖 Changes since v0.3.1 · 74 commits
⚠️ Breaking · 重大变更 · 破壊的変更
PostgreSQL 16 → 18 and Qdrant 1.12 → 1.19. Neither can read the data directory the previous release wrote. Existing instances need a dump-and-restore before starting this version; npx confer-cli detects the old layout up front and prints the steps, and docs/09-deployment.md has the full procedure. Nothing is lost — both images refuse to start rather than initialising an empty database beside your data. To stay on the previous versions for now, run npx confer-cli@0.3.3.
PostgreSQL 16 → 18、Qdrant 1.12 → 1.19。 两者都无法读取上一版本写下的数据目录。已有实例在启动本版本前需要先导出再导入;npx confer-cli 会在启动前检测出旧布局并打印步骤,完整流程见 docs/09-deployment.md。数据不会丢失——两个镜像都会拒绝启动,而不是在你的数据旁边初始化一个空库。想暂时留在旧版本,运行 npx confer-cli@0.3.3。
PostgreSQL 16 → 18、Qdrant 1.12 → 1.19。 どちらも前リリースが書き込んだデータディレクトリを読めません。既存インスタンスは本バージョンを起動する前にダンプとリストアが必要です。npx confer-cli が起動前に旧レイアウトを検出して手順を表示します。詳細は docs/09-deployment.md を参照してください。データは失われません — どちらのイメージも、データの隣に空のデータベースを作るのではなく起動を拒否します。当面は旧バージョンのままにするには npx confer-cli@0.3.3 を実行してください。
🇬🇧 English
🌐 Agent-to-agent communication now actually works across instances
This is the headline. Federation was mechanically impossible before this release, for four separate reasons, all now fixed:
- Every DID was hardcoded to
localhost, ignoringPUBLIC_HOST. Identities are now derived from the address the instance is actually reachable at, and identities minted under the old default are re-hosted on startup. - A gateway resolved its own DIDs over the network, which fails on any HTTP install. It now answers for its own accounts from its own database.
- Outbound requests were signed with a key row nothing has ever written, so every consult failed with
no_signing_key. - A reply echoed the wrong thread id, filing answers under a new conversation while the asker polled the old one — every consult sat at
pendingnext to a correct answer.
Alongside those: an inbound message no longer opens a fresh conversation each time, an answering side that cannot produce an answer now says why instead of leaving the asker polling forever, contacts can be found and added (the is_public flag had no way to be set, so directory search matched nobody), and federation needs HTTPS — docker-compose.tls.yml and npx confer-cli --domain <name> are the supported paths.
📦 One command to self-host
npx confer-cli starts a full instance from published multi-arch images — no clone, no toolchain, no build. Images are published to GHCR on every push to main, so first contact is a pull rather than a twenty-minute compile.
🤖 LLM providers
18 vendors now live in a single catalogue that the gateway, the runtime and the settings UI all read. Model IDs come from each vendor's own /models endpoint instead of a hand-maintained list that had drifted into names no vendor ever served. An empty model list now carries its reason (no_key / unauthorized / unreachable / unsupported) rather than silently reading as "no models". A fully local install — Ollama for chat and embeddings — works end to end with no cloud key at all.
🎨 Interface
The client was rewritten around attribution — who is speaking for whom — rather than reading as a generic chat app. Also: WCAG 2.0 AA contrast throughout, full keyboard navigation, usable below 768px, and the login payload cut from 663 KB to 94 KB on the wire (per-language locale loading, brotli precompression, and no more re-parsing every message on screen once per streamed token).
🐛 Correctness
- Message ordering.
ulid()is randomly ordered inside a millisecond andcreated_atis a transaction timestamp that loses its microseconds through the driver — so the chat history window silently dropped its newest message, conversation paging could skip one, A2A history took the oldest twenty, and the consult long-poll could return a reply written before the question as its answer. Ids are monotonic now, and every path that pages or windows messages uses the id alone. - A slow model reply no longer dies at ten seconds (Bun's default idle timeout, which the gateway's own log reported as a clean
200). - Reloading a streaming answer no longer buys a second completion and appends a duplicate reply.
- An agent with no model configured says so, instead of dialling a vendor the owner has no key for and returning
401.
🔒 Supply chain
bun audit is a blocking CI gate and re-runs nightly; the Rust crate's 470 dependencies are audited beside it. Dependabot covers bun, cargo, actions, docker and compose. The Tauri Cargo.lock is committed, so a tagged build is reproducible. Production deploys keep the outgoing image as :previous, and infra/rollback.sh puts it back.
🇨🇳 中文
🌐 跨实例的 Agent 通信这次是真的通了
这是本版本的重点。在此之前联邦在机械上就不可能,有四个各自独立的原因,现已全部修复:
- 每个 DID 都写死
localhost,完全无视PUBLIC_HOST。现在身份从实例实际可达的地址推导,旧默认值下生成的身份会在启动时迁移。 - 网关通过网络解析自己的 DID,在任何 HTTP 部署上都会失败。现在自己账号的文档由自己的数据库直接回答。
- 出站请求去签名时找的是一个从没被写过的密钥行,所以每次咨询都以
no_signing_key告终。 - 回复回显了错误的 thread id,答案被归到新会话,而提问方还在轮询旧的——每次咨询都停在
pending,旁边就摆着正确答案。
同时修复:入站消息不再每条新建一个会话;答不出来的一方现在会说明原因,而不是让提问方永远轮询;联系人可以被搜到和添加了(is_public 此前没有任何地方能写入,导致目录搜索匹配不到任何人);联邦需要 HTTPS——docker-compose.tls.yml 和 npx confer-cli --domain <域名> 是支持的两条路径。
📦 一条命令自托管
npx confer-cli 从已发布的多架构镜像启动完整实例——不用 clone,不用装工具链,不用编译。镜像在每次推送 main 时发布到 GHCR,所以第一次接触是一次拉取,而不是二十分钟的编译。
🤖 模型厂商
18 家厂商现在集中在一份目录里,网关、运行时和设置界面共读同一份。模型 ID 来自各厂商自己的 /models 接口,不再是手工维护的列表——那份列表已经漂移到了一些没有任何厂商提供过的名字。空模型列表现在会带上原因(no_key / unauthorized / unreachable / unsupported),而不是被静默读成「这家没有模型」。纯本地部署——对话和 embedding 都走 Ollama——完全不需要任何云端 key 就能跑通。
🎨 界面
客户端围绕「谁在替谁说话」这条主线重做,不再像一个通用聊天应用。另外:全站 WCAG 2.0 AA 对比度、完整键盘导航、768px 以下可用,登录页传输体积从 663 KB 降到 94 KB(按语言加载语言包、brotli 预压缩,以及不再对屏幕上每条消息按流式 token 逐次重新解析)。
🐛 正确性
- 消息排序。
ulid()在同一毫秒内是随机序,而created_at是事务时间戳且经驱动会丢掉微秒——于是对话历史窗口会静默丢掉最新一条,会话分页可能跳过一条,A2A 历史取的是最老的二十条,咨询长轮询还可能把写在提问之前的回复当成答案返回。现在 id 是单调的,所有分页和取窗口的路径都只用 id。 - 慢速模型回复不再在十秒被掐断(Bun 的默认空闲超时,而网关自己的日志显示的是一个干净的
200)。 - 刷新流式回答不再买第二次补全并追加一条重复回复。
- 没有配置模型的 Agent 会明说,而不是去拨一个用户根本没有 key 的厂商然后返回
401。
🔒 供应链
bun audit 是阻塞的 CI 闸门并且每晚重跑;Rust crate 的 470 个依赖在旁边一起审计。Dependabot 覆盖 bun、cargo、actions、docker 和 compose。Tauri 的 Cargo.lock 已提交,所以打了 tag 的构建是可复现的。生产部署会把被替换的镜像保留为 :previous,infra/rollback.sh 可以把它放回去。
🇯🇵 日本語
🌐 インスタンス間のエージェント通信が実際に動くようになりました
本リリースの目玉です。これまでフェデレーションは機構的に不可能でした。独立した 4 つの原因があり、すべて修正されています:
- すべての DID が
PUBLIC_HOSTを無視してlocalhostにハードコードされていました。現在は、インスタンスが実際に到達可能なアドレスから ID を導出し、旧デフォルトで発行された ID は起動時に移行されます。 - ゲートウェイが自分自身の DID をネットワーク越しに解決していたため、HTTP 環境では必ず失敗していました。現在は自分のアカウントの文書を自分のデータベースから直接返します。
- 送信リクエストの署名が、一度も書き込まれたことのない鍵の行を参照していたため、すべての相談が
no_signing_keyで失敗していました。 - 返信が誤った thread id を返していたため、回答が新しい会話に登録される一方で質問側は古い会話をポーリングしていました — 正しい回答がすぐ隣にあるのに、すべての相談が
pendingのままでした。
あわせて修正: 受信メッセージごとに新しい会話が作られることがなくなりました。回答できない側は理由を返すようになり、質問側が永久にポーリングし続けることはありません。連絡先を検索して追加できるようになりました(is_public はどこからも書き込めず、ディレクトリ検索が誰にも一致しませんでした)。フェデレーションには HTTPS が必要です — docker-compose.tls.yml と npx confer-cli --domain <ドメイン> がサポートされる方法です。
📦 コマンド 1 つでセルフホスト
npx confer-cli は公開済みのマルチアーキテクチャイメージから完全なインスタンスを起動します — clone もツールチェーンもビルドも不要です。イメージは main への push ごとに GHCR へ公開されるため、最初の一歩は 20 分のコンパイルではなく pull で済みます。
🤖 LLM プロバイダ
18 のベンダーが 1 つのカタログにまとまり、ゲートウェイ・ランタイム・設定 UI がすべて同じものを読みます。モデル ID は各ベンダー自身の /models から取得します — 手作業のリストは、どのベンダーも提供したことのない名前へと乖離していました。空のモデル一覧は理由(no_key / unauthorized / unreachable / unsupported)を伴うようになり、「モデルが無い」と黙って解釈されることはありません。完全ローカル構成 — チャットも embedding も Ollama — がクラウドの鍵なしで動作します。
🎨 インターフェース
クライアントを「誰が誰の代わりに話しているか」という軸で作り直し、汎用チャットアプリのような見え方をやめました。ほかに: 全体で WCAG 2.0 AA のコントラスト、完全なキーボード操作、768px 未満でも使えるレイアウト、ログイン時の転送量を 663 KB から 94 KB へ削減(言語ごとのロケール読み込み、brotli 事前圧縮、ストリーミングのトークンごとに画面上の全メッセージを再解析しない)。
🐛 正しさ
- メッセージの順序。
ulid()は同一ミリ秒内でランダム順であり、created_atはトランザクションのタイムスタンプでドライバ経由でマイクロ秒を失います — その結果、チャット履歴のウィンドウが最新の 1 件を黙って落とし、会話のページングが 1 件飛ばすことがあり、A2A 履歴は最も古い 20 件を取り、相談のロングポールは質問より前に書かれた返信を回答として返し得ました。現在 ID は単調増加で、ページングとウィンドウ処理はすべて ID のみを使います。 - 応答の遅いモデルが 10 秒で切断されなくなりました(Bun の既定アイドルタイムアウト。ゲートウェイ自身のログはきれいな
200を出していました)。 - ストリーミング中の回答を再読み込みしても、2 回目の推論が課金され重複した返信が追加されることはなくなりました。
- モデル未設定のエージェントはそう明示します — 所有者が鍵を持たないベンダーに接続して
401を返すことはありません。
🔒 サプライチェーン
bun audit はブロッキングの CI ゲートで、毎晩再実行されます。Rust クレートの 470 依存もあわせて監査します。Dependabot は bun / cargo / actions / docker / compose を対象にします。Tauri の Cargo.lock をコミットしたため、タグ付きビルドは再現可能です。本番デプロイは置き換えられるイメージを :previous として保持し、infra/rollback.sh で戻せます。
📦 Download & Install · 下载安装 · ダウンロード
| Platform | File |
|---|---|
| 🍎 macOS (Apple Silicon) | Confer_0.4.0_aarch64.dmg |
| 🪟 Windows (x64) | Confer_0.4.0_x64-setup.exe · Confer_0.4.0_x64_en-US.msi |
| 🪟 Windows (ARM64) | Confer_0.4.0_arm64-setup.exe · Confer_0.4.0_arm64_en-US.msi |
| 🐧 Linux (x64) | Confer_0.4.0_amd64.AppImage · Confer_0.4.0_amd64.deb · Confer-0.4.0-1.x86_64.rpm |
| 🤖 Android | app-universal-release.apk |
No macOS Intel build — the release matrix covers Apple Silicon only.
未构建 macOS Intel 版本——发布矩阵仅覆盖 Apple Silicon。
macOS Intel 版はビルドされていません(リリースマトリクスは Apple Silicon のみ)。
Server · 服务端 · サーバー
# One command — needs Docker, plus Node 18+ for npx.
# Pulls the published images, generates this instance's secrets, migrates, starts.
npx confer-cli
# ...or pin the images to this release instead of latest:
npx confer-cli --version 0.4.0
# Federation needs a real domain and HTTPS — did:web resolution is https-only.
npx confer-cli --domain confer.example.comOpen http://localhost, register the first account, add an LLM API key in Settings.
打开 http://localhost 注册第一个账号,然后在设置里填入 LLM API key。
http://localhost を開き、最初のアカウントを登録し、設定で LLM API キーを追加します。
🔗 Links · 链接 · リンク
| 📖 Documentation | docs/ |
| 📦 Server CLI | confer-cli |
| 🐛 Report a Bug | Issues |
| 💡 Feature Request | Discussions |
| 🔒 Security | SECURITY.md |
Built with Tauri 2.0 · Bun · Hono · A2A Protocol · DID:web · RFC 9421