@pandanpc/mcp-server 0.1.58 → 0.1.59

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pandanpc/mcp-server",
-  "version": "0.1.58",
+  "version": "0.1.59",
   "description": "PandaNpc MCP Server — PandaNote and Agent connection management for Codex, Claude Code, and other MCP clients",
   "keywords": [
     "mcp",

package/skills/pandanpc/references/agents.md

@@ -1,6 +1,8 @@
-# Cross-engine workers (agents)
+# Cross-connection and cross-engine workers
 
-Use these tools when the user names an engine or model (Codex, GPT, Astra, Sol, gpt-6-astra, gpt-5.6-sol, PandaCode, Kimi, remote Claude Code) or asks to delegate, dispatch, hand off, review with, or spawn a worker/subagent. No project skill is required.
+Keep ordinary delegation on the current connection and engine in that engine's native subagent tools. This preserves native child creation, messaging, waiting and follow-up, including children restored with an existing session.
+
+Use these PandaNpc tools only when the user names or the task requires a different connection or engine/model (Codex, GPT, Astra, Sol, gpt-6-astra, gpt-5.6-sol, PandaCode, Kimi, remote Claude Code). No project skill is required.
 
 ## Model -> engine
 
@@ -12,8 +14,9 @@ Use these tools when the user names an engine or model (Codex, GPT, Astra, Sol,
 
 ## Steps
 
-1. `list_engine_models(engine?, cwd?)` -> choose an online `connection_id` whose engine serves the model.
-2. `spawn_engine_agent(name, connection_id, task, model?, reasoning_effort?)`.
+1. Confirm that the target differs from the current connection or engine. For same-connection, same-engine work, stop here and use native subagent tools.
+2. `list_engine_models(engine?, cwd?)` -> choose an online `connection_id` whose engine serves the model.
+3. `spawn_engine_agent(name, connection_id, task, model?, reasoning_effort?)`.
    - Omit `parent_session_id`: the worker is attached to the calling session, which receives the reply. Check the echoed `parent` (id, title, engine, connection) in the result.
    - Attach to a different session only on purpose: pass its PandaNpc `parent_session_id` with `parent_override: true`. A mismatch without the override is rejected. Never guess the id from a URL or a native thread id.
    - `parent_link_warning` in a result means the sidebar link failed and the worker is listed on its own.
@@ -21,8 +24,8 @@ Use these tools when the user names an engine or model (Codex, GPT, Astra, Sol,
    - Workers always run with full access and never prompt (Codex: approval never + danger-full-access; Claude: bypassPermissions), regardless of the connection's permission setting.
    - The session is filed under the connection's project whose directory matches the parent session's cwd; pass `project` (folder_id, name or absolute directory) to choose another. The response's `project` field says which one was used.
    - `request_id` is optional; pass a stable value to retry safely.
-3. Wait for the plaintext teammate reply in the parent session. If none arrives, call `engine_agent_result(agent_id)`.
-4. Follow-ups: `send_engine_agent(agent_id, task)`. Interrupt: `stop_engine_agent(agent_id)` (Esc semantics; background subagents and other sessions keep running).
+4. Wait for the plaintext teammate reply in the parent session. If none arrives, call `engine_agent_result(agent_id)`.
+5. Follow-ups to that PandaNpc worker only: `send_engine_agent(agent_id, task)`. Native children continue through their native follow-up/message tools. Interrupt a PandaNpc worker with `stop_engine_agent(agent_id)` (Esc semantics; background subagents and other sessions keep running).
 
 ## Verify