opencode-collaboration 0.5.0 → 0.5.2

package/README.md

@@ -27,7 +27,7 @@ Run several opencode terminals in parallel (different repos, worktrees, or tasks
 - **Explicit TUI controls**: palette actions use host dialogs for selection and confirmation; slash wrappers remain available for automation and compatibility
 - Local only: everything stays on your machine (Unix-domain sockets on macOS/Linux, loopback TCP on Windows, plus a compatibility loopback listener for v1 peers)
 - The peer name is appended to this process's root session titles, so every opencode window shows who it is at a glance
-- Peer-triggered turns run under the receiving session's selected agent/mode (e.g. a custom `teacher`) instead of opencode's default `build`, so that agent's model, tools and permission rules apply. The mode is taken from the session's last human input (message or command) and its server-persisted agent; if none is known yet, the default agent is used
+- Peer-triggered turns run under the receiving session's selected agent/mode (e.g. a custom `teacher`) instead of opencode's default `build`, so that agent's model, tools and permission rules apply. The mode is the agent recorded server-side at session creation and by prompts (messages/commands); if none is known yet, the default agent is used. If the recorded agent is later removed, the next peer message falls back to the default agent
 
 ## Install
 
@@ -96,7 +96,7 @@ Other Opencode sessions (2):
 Use send_message to tell "backend" that the login form now posts to /v2/login.
 ```
 
-The receiving session gets the text immediately as a synthetic user message, including the sender's exact endpoint ID and how to reply. `send_message` returns a tracking ID; use `peer_message_status` or `/peers-outbox` to distinguish transport receipt from final delivery.
+The receiving session gets the text immediately as a synthetic user message, including the sender's exact endpoint ID and how to reply. The receiver is told to reply via `send_message` only when a response is actually expected, and is explicitly told not to send acknowledgement-only messages — so two agents do not loop politely forever. `send_message` returns a tracking ID; use `peer_message_status` or `/peers-outbox` to distinguish transport receipt from final delivery.
 
 **Review held messages** (when `inboundPolicy` is `"hold"`):
 

package/README.zh-CN.md

@@ -27,7 +27,7 @@
 - **明确的 TUI 控制**：面板操作使用宿主对话框进行选择确认；斜杠命令封装仍可用于自动化和兼容性场景
 - 纯本地运行：一切都在你的机器上（macOS/Linux 使用 Unix 域套接字，Windows 使用回环 TCP，另有一个兼容 v1 对端的回环监听器）
 - 本进程的 peer 名字会追加到根会话标题后面，一眼就能看出每个 opencode 窗口是谁
-- peer 触发的回合会沿用接收方会话当前选中的 agent/模式（如自定义的 `teacher`），而不是 opencode 默认的 `build`，因此该 agent 的模型、工具与权限规则都会生效。模式取自该会话最近一条人类输入（消息或命令）及其服务端持久化的 agent；若尚无所知则使用默认 agent
+- peer 触发的回合会沿用接收方会话当前选中的 agent/模式（如自定义的 `teacher`），而不是 opencode 默认的 `build`，因此该 agent 的模型、工具与权限规则都会生效。模式取自服务端记录的 agent（会话创建时写入，并随消息/命令的 prompt 更新）；若尚无所知则使用默认 agent。若记录的 agent 之后被删除，下一条 peer 消息会回退到默认 agent
 
 ## 安装
 
@@ -96,7 +96,7 @@ Other Opencode sessions (2):
 Use send_message to tell "backend" that the login form now posts to /v2/login.
 ```
 
-接收方会话会立即收到这条文本，以一条合成的用户消息形式出现，其中包含发送方精确的端点 ID 以及回复方式。`send_message` 返回一个追踪 ID；用 `peer_message_status` 或 `/peers-outbox` 来区分"传输层已收到"和"最终已投递"。
+接收方会话会立即收到这条文本，以一条合成的用户消息形式出现，其中包含发送方精确的端点 ID 以及回复方式。接收方被要求**只在确实需要回应时**才通过 `send_message` 回复，并明确禁止仅确认/致谢类消息——避免两个 agent 无限礼貌往返。`send_message` 返回一个追踪 ID；用 `peer_message_status` 或 `/peers-outbox` 来区分"传输层已收到"和"最终已投递"。
 
 **审阅待审消息**（当 `inboundPolicy` 为 `"hold"` 时）：
 

package/dist/delivery.js

@@ -15,10 +15,14 @@ const FOOTER = "---\n" +
 // appends this to the agent's own system prompt (LLMRequestPrep.prepare), so it
 // never replaces it and does not leak into later turns or other sessions.
 const REPLY_DIRECTIVE = "This turn was triggered by a plain-text message from another opencode session (a peer), " +
-    "not by your user. Your user cannot see it, and your normal chat reply is NOT delivered to " +
-    "the peer. To respond — including to ask a clarifying question or request missing details — " +
-    "you MUST call the send_message tool with `to` set to the sender's name. Never answer a peer " +
-    "by typing a normal chat reply.";
+    "not by your user. Your user cannot see it, and your normal chat reply is NOT delivered to the peer. " +
+    "Reply only when a reply is actually expected:\n" +
+    "- If the peer asked a question, requested work, or is waiting for your response, call the " +
+    "send_message tool with `to` set to the sender's name.\n" +
+    "- If the message is only an acknowledgement, a thank-you, an agreement, a status update that " +
+    "needs no action, or a confirmation that the task is done, do NOT reply — briefly tell your user and stop.\n" +
+    "Never send acknowledgement-only, thanks, or agreement messages: they cause an endless " +
+    "back-and-forth between agents. Once you and the peer have reached a conclusion, stop.";
 const NOTICE_FOOTER = "---\n" +
     "This is an automated notification from the opencode-collaboration plugin. " +
     "Show it to the user verbatim, then stop. Do not take further action.";
@@ -29,8 +33,11 @@ export function formatMessages(messages) {
         ? `to="${senders[0].name}" (exact endpoint ID "${senders[0].instanceId}")`
         : `to the sender name shown in each message header (endpoint IDs ${senders.map((sender) => `"${sender.instanceId}"`).join(", ")})`;
     const header = "[incoming message from another opencode session — this is NOT from your user. " +
-        "Reply by calling the send_message tool; a normal chat reply is never delivered to the sender.]";
-    return header + "\n\n" + blocks.join("\n\n") + "\n\n" + FOOTER + ` To reply, call the send_message tool with ${replyTarget}.`;
+        "To reply to the sender, call the send_message tool; a normal chat reply is never delivered. " +
+        "Reply only if the peer expects a response — do not acknowledge just to be polite.]";
+    return header + "\n\n" + blocks.join("\n\n") + "\n\n" + FOOTER +
+        ` To reply, call the send_message tool with ${replyTarget}. ` +
+        "Reply only if a response is expected; do not send acknowledgement-only messages.";
 }
 export function deterministicPeerMessageId(sessionId, message) {
     const digest = createHash("sha256")
@@ -89,18 +96,21 @@ export function Delivery(opts) {
         }));
         assertPromptSucceeded(result);
     }
-    function isAgentNotFoundError(err) {
-        const source = err;
-        const text = [
-            err instanceof Error ? err.message : String(err),
-            source?.data?.message,
-            source?.responseBody,
-            source?.body,
-        ]
-            .filter((value) => typeof value === "string" && value.length > 0)
-            .join(" ")
-            .toLowerCase();
-        return text.includes("agent not found");
+    async function agentStillExists(agent) {
+        // opencode reports an unknown agent as a generic 500 whose body has no
+        // "Agent not found" text, so decide by asking the server for the live list.
+        // If the list is unavailable, assume the agent is fine and let the caller
+        // requeue as before (only a definite miss triggers the downgrade).
+        try {
+            const response = await opts.client.app.agents();
+            const agents = response.data;
+            if (!Array.isArray(agents))
+                return true;
+            return agents.some((entry) => entry?.name === agent);
+        }
+        catch {
+            return true;
+        }
     }
     async function inject(sessionId, text, message) {
         const messageID = message ? deterministicPeerMessageId(sessionId, message) : undefined;
@@ -129,9 +139,10 @@ export function Delivery(opts) {
             await sendPrompt(sessionId, buildBody(agent));
         }
         catch (err) {
-            // The recorded agent no longer exists (e.g. removed from config). Drop it
-            // and retry once under opencode's default agent instead of requeue-looping.
-            if (!agent || !isAgentNotFoundError(err))
+            // The recorded agent may no longer exist (removed from config). Verify
+            // against the live agent list, and only then drop the record and retry
+            // once under opencode's default agent instead of requeue-looping.
+            if (!agent || (await agentStillExists(agent)))
                 throw err;
             opts.onAgentRejected?.();
             await sendPrompt(sessionId, buildBody(undefined));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-collaboration",
-  "version": "0.5.0",
+  "version": "0.5.2",
   "description": "Cross-session messaging for opencode — let independent sessions discover and text each other, modeled after Claude Code's cross-session messaging",
   "type": "module",
   "main": "./dist/index.js",