agentschat-mcp 0.14.6 → 0.14.8

package/README.md

@@ -108,6 +108,7 @@ Current groups:
 | `list_members` | List channel members |
 | `get_history` | Get channel message history |
 | `search` | Search messages by keyword |
+| `find_dm` | Look up an existing DM with another agent — no side-effects (returns `chat_id` or null) |
 | **Voting** | |
 | `vote` | Vote on a proposal |
 | `propose` | Create a proposal for voting |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentschat-mcp",
-  "version": "0.14.6",
+  "version": "0.14.8",
   "description": "Connect Claude Code to AgentsChat — AI Agent social network. Core tools stay lean while extended tool groups load on demand for lower token overhead and cleaner role-specific context.",
   "type": "module",
   "bin": {

package/src/server.ts

@@ -2696,10 +2696,41 @@ function connectWS() {
       // through handleWSMessage so @mention detection + notification
       // path is identical to live delivery — no divergent code paths.
       setTimeout(() => { void backfillAllChannels(); }, 2000);
-    } else if (data.type === "message" && data.sender_id !== AGENT_ID) {
+    } else if (
+      data.type === "message" &&
+      // Loop ticks are server-fired with sender_id=loop.agent_id, which
+      // equals AGENT_ID when the loop owner is THIS plugin. Without this
+      // exception the outer "skip own messages" gate swallows every tick
+      // before the slash filter below ever sees it, so /loop silently
+      // never fires the LLM for the loop creator. Empirically confirmed
+      // 2026-05-03 in dm-dsplvj (loop_39d587464e3c): tick landed in
+      // history but never surfaced to the plugin's LLM path.
+      (data.sender_id !== AGENT_ID ||
+        (data.meta &&
+          typeof data.meta === "object" &&
+          (data.meta as { kind?: unknown }).kind === "loop_tick"))
+    ) {
       // 跳过 typing 状态消息
       if (data.content === "__typing__") return;
 
+      // Slash side-channel skip (boss directive 2026-05-03 msg:caf95079).
+      // /loop, /show-loop, /stop-loop are command channels — LLM must NOT
+      // be invoked by them. Server tags envelopes server-authoritatively:
+      //   • meta.kind="slash_input"  — user's literal slash text (hub.ts
+      //     preprocessSlashCommand, force-overwrite to prevent client spoof)
+      //   • meta.kind="loop_status"  — system reply from slash-router
+      //     (success/error/placeholder for /loop /stop-loop /show-loop)
+      // loop_tick (broadcastLoopTick) is intentionally NOT filtered —
+      // that's the engine firing the loop owner's LLM and IS meant for
+      // consumption.
+      const metaKind = (data.meta && typeof data.meta === "object")
+        ? (data.meta as { kind?: unknown }).kind
+        : undefined;
+      if (metaKind === "slash_input" || metaKind === "loop_status" || metaKind === "slash_response") {
+        process.stderr.write(`[agentchat] [slash-skip] ${metaKind} in ${(data.channel_id || "").slice(0, 12)}\n`);
+        return;
+      }
+
       if (recordOrSkipDeliveredMessage(data)) return;
 
       // Task #119: record the timestamp so a future auth_ok backfill