Start Discussion

Start Discussion — Non-blocking Discussion Initiation

Agent 识别到需要深入讨论的话题后，创建飞书讨论群、通过 push_to_agent 注入初始化上下文、记录映射，立即返回（不阻塞当前工作）。

适用于: 发起讨论、离线提问、非阻塞交互 | 不适用于: 解散群（用户驱动）、PR 审查群（用 PR Scanner）

When to Use

• Agent 在工作中发现需要与用户深入讨论的话题（用户重复指令、多步反复修正、隐式抱怨、花费较大的工作存疑等）
• Agent 需要用户输入但不想阻塞当前任务
• Agent 想要委派一个问题让用户在独立群中讨论
• 定时任务分析发现需要讨论的问题

Single Responsibility

• ✅ Create a Feishu discussion group via lark-cli
• ✅ Inject initialization context via push_to_agent MCP tool
• ✅ Record mapping in workspace/bot-chat-mapping.json
• ✅ Return immediately — non-blocking by design
• ❌ DO NOT wait for user response in the discussion group
• ❌ DO NOT dissolve groups — let users handle lifecycle
• ❌ DO NOT use IPC Channel for group operations — use lark-cli via Bash
• ❌ DO NOT create PR review groups (use PR Scanner skill)

Context Variables

When invoked, you receive:

• Chat ID: Source chat where the discussion topic was identified
• Message ID: The triggering message ID
• Sender Open ID: The user who triggered the discussion

Workflow

Step 1: Determine Discussion Topic and Context

Analyze the current conversation and determine:

1. Topic: A concise title for the discussion (max 64 chars for group name)
2. Question: The specific question or issue to discuss
3. Background: Relevant context — chat history excerpts, file references, error logs, etc.
4. Participants: Open IDs of additional users to include (optional — sender is always included automatically)

Step 2: Create Discussion Group

Use lark-cli to create a new Feishu group. Always include the triggering user (Sender Open ID) in the group:

# Create group with the triggering user
lark-cli im +chat-create --name "讨论: {topic}" --description "Agent 发起的讨论: {topic}" --users "{sender_open_id}"

If additional participants need to be included, merge them into the same --users list:

lark-cli im +chat-create --name "讨论: {topic}" --description "Agent 发起的讨论: {topic}" --users "{sender_open_id},ou_xxx,ou_yyy"

Parse the response to extract the new group's chatId (format: oc_xxx).

If lark-cli is not available, report the error and stop:

lark-cli --version || echo "ERROR: lark-cli not found in PATH"

Step 3: Inject Context via push_to_agent

Use the push_to_agent MCP tool to send an initialization instruction to the new group's agent. This triggers lazy agent creation and injects a system instruction.

push_to_agent(chatId: "{new group chatId}", message: "{initialization prompt}")

The initialization prompt should include:

• The discussion topic and purpose
• The source chat ID (for traceability)
• Background materials and the specific question to discuss
• Instructions for the agent (e.g., "引导用户讨论以上问题")

Note: push_to_agent handles agent creation automatically. The agent will then manage the conversation in the new group using its standard messaging capabilities (send_text, send_interactive, etc.).

Step 4: Record Mapping

Append the new group to workspace/bot-chat-mapping.json:

# Read current mapping
cat workspace/bot-chat-mapping.json 2>/dev/null || echo "{}"

Add an entry with key discussion-{short-uuid}:

{
  "discussion-{uuid}": {
    "chatId": "oc_xxx",
    "createdAt": "{ISO timestamp}",
    "purpose": "discussion"
  }
}

Write the updated mapping atomically (write to temp file, then move):

echo '{ ... updated JSON ... }' > workspace/bot-chat-mapping.json.tmp \
  && mv workspace/bot-chat-mapping.json.tmp workspace/bot-chat-mapping.json

Step 5: Confirm and Return

Report to the source chat that the discussion has been initiated:

已创建讨论群「{topic}」，上下文已发送。我将继续当前工作，讨论结果稍后会处理。

Do NOT wait for any response — return immediately.

lark-cli Command Reference

Error Handling

Design Principles

1. Non-blocking: Return to source chat immediately after sending context
2. push_to_agent for initialization: Use MCP tool for agent creation + context injection, NOT send_text/send_interactive
3. Idempotent: Check mapping table before creating (avoid duplicates)
4. Cache is rebuildable: bot-chat-mapping.json can be reconstructed from Feishu API
5. No IPC for group ops: Direct lark-cli calls via Bash — no MCP/IPC indirection

Feedback Origin — This Group Is Not a Feedback Channel

Issue #4017. When this group is the execution chat of a longer-running delegated task driven by scheduled runs, it carries progress updates and delivery only. User feedback — corrections, intent changes, scope or source-preference adjustments — originates in the initial conversation with the user, not here.

If the task maintains a shared state file (e.g. RESEARCH.md), feedback flows only through that file's dedicated ## User Feedback section (append-only, timestamped, one point per entry). If it does not, tell the user in the initialization prompt to give feedback in the source chat rather than in this group.

Concretely, when composing the initialization prompt (Step 3):

• If a state file with a feedback section exists, tell the execution agent to re-read the ## User Feedback section at the start of every run and carry it forward verbatim when writing state back.
• Either way, do not tell the execution agent to read the initial conversation's messages — the state file is the single feedback channel, and cross-conversation message reading is out of scope (the earlier approach was reviewed and rejected; see closed PR #4030).

Integration with Other Skills