pi-brainstorm 0.3.0 → 0.4.0

package/README.md

@@ -10,7 +10,7 @@ The plugin stores each participant's full contribution in a local meeting blackb
 
 ## Features
 
-- Multi-model brainstorming with configurable participants (default: GPT, DeepSeek, MiniMax).
+- Multi-model brainstorming with configurable participants (default: GPT, DeepSeek, MiniMax, GLM).
 - Debate / battle mode where agents prosecute, stress-test, and challenge positions.
 - Configuration-driven: add, remove, or customize participants via YAML.
 - Round-by-round summaries focused on consensus, disagreement, and next questions.
@@ -30,7 +30,7 @@ pi install npm:pi-brainstorm
 From GitHub:
 
 ```bash
-pi install git:github.com/Jarcis-cy/pi-brainstorm@v0.3.0
+pi install git:github.com/Jarcis-cy/pi-brainstorm@v0.4.0
 ```
 
 For local development:
@@ -88,7 +88,7 @@ Agent files generated by pi-brainstorm contain `<!-- managed-by: pi-brainstorm -
 /debate <topic>
 ```
 
-`/brainstorm` starts a three-round multi-model brainstorming session.
+`/brainstorm` starts an interactive three-round multi-model brainstorming session. After Round 1 and Round 2, the facilitator should stop, summarize the round, and wait for your feedback or permission before continuing.
 
 `/debate` starts an open-ended multi-agent battle that should continue until convergence or user intervention.
 
@@ -98,7 +98,7 @@ Agent files generated by pi-brainstorm contain `<!-- managed-by: pi-brainstorm -
 participant model -> meeting_append_entry -> .pi-meetings/... files
 participant model -> short WROTE_ENTRY summary -> facilitator context
 file watcher -> compact visible card -> main chat
-facilitator -> consensus / disagreement / final synthesis
+facilitator -> round checkpoint / consensus / disagreement / final synthesis
 ```
 
 The blackboard files are the source of truth for the session. The facilitator can read the index or specific entries when producing summaries.

package/README.zh-CN.md

@@ -10,7 +10,7 @@ English README: [README.md](./README.md).
 
 ## 功能
 
-- 多模型头脑风暴：参与者可配置（默认：GPT、DeepSeek、MiniMax）。
+- 多模型头脑风暴：参与者可配置（默认：GPT、DeepSeek、MiniMax、GLM）。
 - 辩论 / battle 模式：Agent 会攻击、审视、反驳彼此的观点。
 - 配置驱动：通过 YAML 添加、删除或自定义参与者。
 - 每轮输出聚焦共识、分歧、关键问题和下一步方向。
@@ -30,7 +30,7 @@ pi install npm:pi-brainstorm
 通过 GitHub 安装：
 
 ```bash
-pi install git:github.com/Jarcis-cy/pi-brainstorm@v0.3.0
+pi install git:github.com/Jarcis-cy/pi-brainstorm@v0.4.0
 ```
 
 本地开发安装：
@@ -83,7 +83,7 @@ participants:
 
 ## 前置条件
 
-该扩展依赖 pi 中已有的 `subagent` 工具。命令处理器会先创建本地会议记录，然后让主 Agent 调用配置中定义的参与者（默认：`gpt-brainstormer`、`deepseek-brainstormer`、`minimax-brainstormer`）。
+该扩展依赖 pi 中已有的 `subagent` 工具。命令处理器会先创建本地会议记录，然后让主 Agent 调用配置中定义的参与者（默认：`gpt-brainstormer`、`deepseek-brainstormer`、`minimax-brainstormer`、`glm-brainstormer`）。
 
 第一次使用时，如果这些用户级 agent 不存在，扩展会询问是否写入内置默认定义。已有同名且不含托管标记的文件不会被覆盖。
 
@@ -94,7 +94,7 @@ participants:
 /debate <主题>
 ```
 
-`/brainstorm` 启动三轮多模型头脑风暴。
+`/brainstorm` 启动交互式三轮多模型头脑风暴。Round 1 和 Round 2 结束后，主持人应停止、总结本轮，并等待你的反馈或继续许可，再进入下一轮。
 
 `/debate` 启动开放式多 Agent battle，直到收敛或用户介入为止。
 
@@ -104,7 +104,7 @@ participants:
 参与者模型 -> meeting_append_entry -> .pi-meetings/... 文件
 参与者模型 -> 短 WROTE_ENTRY 摘要 -> 主持人上下文
 文件 watcher -> 短卡片 -> 主会话可见区域
-主持人 -> 共识 / 分歧 / 最终综合
+主持人 -> 每轮 checkpoint / 共识 / 分歧 / 最终综合
 ```
 
 黑板文件是会话事实源。主持人可以按需读取索引或具体条目，然后生成结构化总结。

package/config/default.yaml

@@ -26,7 +26,7 @@ participants:
   - displayName: DeepSeek
     agentName: deepseek-brainstormer
     description: DeepSeek brainstorming consultant. Meticulous systems thinker for multi-model discussion sessions.
-    model: deepseek/deepseek-v4-pro:xhigh
+    model: opencode-go/deepseek-v4-pro:xhigh
     roleTitle: Meticulous Systems Thinker
     rolePrompt: |
       你是多模型头脑风暴中的系统思考者。分析结构、依赖、扩展上限和失败模式。用中文回答。
@@ -56,3 +56,21 @@ participants:
       prompt: |
         DEBATE MODE. You are THE CONTRARIAN. Take the opposite position from the dominant view. Expose groupthink. Propose radical alternatives. Challenge the fundamental premise if needed. Use Chinese.
     brainstormRole: Creative lateral thinker
+
+  - displayName: GLM
+    agentName: glm-brainstormer
+    description: GLM-5.2 brainstorming consultant. Pragmatic implementer who grounds ideas in feasibility, cost, and real-world constraints.
+    model: opencode-go/glm-5.2:xhigh
+    roleTitle: Pragmatic Implementer
+    rolePrompt: |
+      你是多模型头脑风暴中的务实执行者。当别人谈愿景、讲架构、玩创意时，你负责问：「这到底能不能做出来？」「需要多少资源？」「有什么更简单的方案？」你的价值在于把讨论拉回现实。用中文回答。
+    whatYouDo:
+      - 从可行性、成本、时间和资源角度审视每个提案
+      - 指出不切实际的想法——诚实、直接
+      - 提出更简单的替代方案，避免过度设计
+      - 关注执行路径和最小可行版本
+    debatePersona:
+      label: THE PRAGMATIST
+      prompt: |
+        DEBATE MODE. You are THE PRAGMATIST. Challenge every proposal on feasibility, cost, and real-world constraints. Expose impractical ideas. Push for simpler, cheaper, faster solutions. Use Chinese.
+    brainstormRole: Pragmatic implementer

package/extensions/brainstorm.ts

@@ -525,6 +525,7 @@ function buildBrainstormPrompt(
     '   - phase: "Round 1", "Round 2", or "Round 3"',
     "   - summary: a ONE-SENTENCE summary of your contribution",
     "   - content: your FULL analysis in Chinese (中文)",
+    "   - content must contain only the participant's analysis. Do not include wrapper tags, hidden thinking markers, tool-call text, or WROTE_ENTRY text inside content.",
     "2. After writing, your FINAL ANSWER must be ONLY:",
     "   `WROTE_ENTRY: <your one-sentence summary>`",
     "3. DO NOT paste your full analysis into the chat. The main agent and user will read it from the blackboard.",
@@ -538,10 +539,42 @@ function buildBrainstormPrompt(
     "",
     "## Protocol",
     "Round 1: Each consultant gives initial analysis on the topic. Run all in parallel.",
-    "Round 2: Feed prior discussion back to each. Ask each to challenge the others and propose improvements.",
-    "Round 3: Each gives FINAL recommendation, synthesizing the best ideas.",
+    "After Round 1: read the index, present summaries plus a structural overview, then STOP. Ask the user for feedback or permission to continue. Do NOT start Round 2 in the same assistant turn.",
+    "Round 2: only after the user replies, feed Round 1 plus the user's VERBATIM feedback back to each consultant. Ask each to challenge the others and propose improvements.",
+    "After Round 2: read the index, present summaries plus an updated structural overview, then STOP. Ask the user for feedback or permission to continue. Do NOT start Round 3 in the same assistant turn.",
+    "Round 3: only after the user replies, feed all prior rounds plus the user's VERBATIM feedback back to each consultant. Each gives FINAL recommendation, synthesizing the best ideas.",
     "",
-    "After Round 3, present the complete structural overview and a synthesized conclusion.",
+    "After Round 3, present the complete structural overview and ask whether to write the final conclusion. Only write conclusion.md after the user confirms.",
+    "",
+    "## FAILURE HANDLING",
+    "Subagent calls can fail (timeout, API error, model unavailable, crash). You MUST detect and handle failures:",
+    "",
+    "### Detecting failure",
+    "A subagent call has FAILED if:",
+    "- The tool result contains error text (\"error\", \"failed\", \"timeout\", \"unavailable\", \"connection\", etc.)",
+    "- The subagent returns no output or empty output",
+    "- The subagent output does NOT contain \"WROTE_ENTRY:\" (the expected success marker)",
+    "- The subagent produced no meeting_append_entry call (check with meeting_read_index after the parallel batch)",
+    "",
+    "### Retry strategy",
+    "When you detect a failure:",
+    "1. Wait 5-10 seconds before retrying (API rate limits may have triggered the failure)",
+    "2. Retry the SAME consultant with the SAME task up to 2 more times (3 attempts total)",
+    "3. On retry, add to the task: \"PREVIOUS ATTEMPT FAILED. This is retry N of 2. Focus only on completing your contribution and writing WROTE_ENTRY.\"",
+    "4. If the consultant succeeds on retry, proceed normally",
+    "5. If ALL retries are exhausted, treat this consultant as UNAVAILABLE for this round",
+    "",
+    "### When a consultant is UNAVAILABLE",
+    "- Do NOT wait indefinitely or freeze the session",
+    "- Proceed with the remaining consultants who succeeded",
+    "- In your structural overview, clearly mark the failed participant as: \"[displayName] ([agentName]): 本轮未响应 / did not respond\"",
+    "- Continue to the next round normally — do NOT skip the consultant in future rounds (they may recover)",
+    "- If the same consultant fails in 2 consecutive rounds, warn the user with the real display name, e.g.: \"DeepSeek 连续两轮失败，建议检查该模型是否可用\"",
+    "",
+    "### NEVER do this",
+    "- Do NOT fabricate or simulate a consultant's response",
+    "- Do NOT wait more than 3 minutes total per consultant (including retries)",
+    "- Do NOT abort the entire session because one consultant failed",
     "",
     "## IMPORTANT",
     "- All responses in Chinese (中文).",
@@ -631,6 +664,37 @@ function buildDebatePrompt(
     "- If NOT converged: run another cycle. Keep going.",
     "- If converged: present synthesis to me.",
     "",
+    "## FAILURE HANDLING",
+    "Subagent calls can fail (timeout, API error, model unavailable, crash). You MUST detect and handle failures:",
+    "",
+    "### Detecting failure",
+    "A subagent call has FAILED if:",
+    "- The tool result contains error text (\"error\", \"failed\", \"timeout\", \"unavailable\", \"connection\", etc.)",
+    "- The subagent returns no output or empty output",
+    "- The subagent output does NOT contain \"WROTE_ENTRY:\" (the expected success marker)",
+    "- The subagent produced no meeting_append_entry call (check with meeting_read_index after the chain step)",
+    "",
+    "### Retry strategy",
+    "When you detect a failure in a chain step:",
+    "1. Wait 5-10 seconds before retrying (API rate limits may have triggered the failure)",
+    "2. Retry the SAME debater with the SAME task up to 2 more times (3 attempts total)",
+    "3. On retry, add to the task: \"PREVIOUS ATTEMPT FAILED. This is retry N of 2. Focus only on completing your argument and writing WROTE_ENTRY.\"",
+    "4. If the debater succeeds on retry, proceed to the next debater in the chain",
+    "5. If ALL retries are exhausted, SKIP this debater for this cycle and move to the next debater",
+    "",
+    "### When a debater is UNAVAILABLE",
+    "- Do NOT wait indefinitely or freeze the session",
+    "- Skip to the next debater in the cycle",
+    "- In your structural overview, clearly mark the failed participant as: \"[displayName] ([agentName]): 本轮未响应 / did not respond\"",
+    "- Continue the debate with remaining debaters — do NOT abort the whole debate",
+    "- If the same debater fails in 2 consecutive cycles, warn the user with the real display name, e.g.: \"DeepSeek 连续两轮失败，建议检查该模型是否可用\"",
+    "",
+    "### NEVER do this",
+    "- Do NOT fabricate or simulate a debater's argument",
+    "- Do NOT wait more than 3 minutes total per debater (including retries)",
+    "- Do NOT abort the entire debate because one debater failed",
+    "- Do NOT get stuck in an infinite retry loop — 3 attempts max per debater per cycle",
+    "",
     "## Rules",
     "- NEVER stop at a predetermined count. Only convergence or user intervention ends this debate.",
     "- All responses in Chinese (中文).",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-brainstorm",
-  "version": "0.3.0",
+  "version": "0.4.0",
   "description": "Multi-model brainstorming and debate sessions for pi subagents.",
   "type": "module",
   "keywords": [