support-thread

Extract actionable structure from a support/incident thread, and separate what was actually promised from what merely sounded like a promise. "客戶" below reads as whoever reported the problem — an external customer, another team, or the owner themselves filing against their own infra.

Input

A chat thread from any enabled channel (pull with mcp__discord__fetch_messages, or mcp__slack__fetch_messages when the external Slack adapter is wired in). Optionally the customer's prior message / ticket if linked.

Output template

支援案件：<one-line problem>
客戶：<name / org if known>
來源：<thread permalink>

問題描述：
- <what the customer reports — facts, no speculation>

客戶環境：
- 產品/版本：<X>
- 系統：<OS / browser / cluster>
- 規模：<users / data / load if relevant>
- (or: 未知)

已嘗試處理：
- <step 1, who did it, outcome>
- <step 2, ...>

目前判斷：
- <best current hypothesis + 信心強度>

客戶期待：
- <what would 'solved' look like from their POV>
- <timeline they expect>

下一步：
- <action> · owner：<@who> · 預計：<when>

風險：
- <risk: SLA / churn / commitment leakage / blast radius>

關聯：
- 相關 issue / PR / past support thread: <links>

Commitment classification

For every statement aimed at the customer in this thread, classify into:

• confirmed commitment — agreed deliverable, with date + owner. Should be reflected in memory/commitments.md.
• proposed — suggestion not yet ratified by both sides. Don't write to commitments yet; flag for owner to follow up.
• internal-only — said inside the team thread, not yet relayed externally.
• unclear — ambiguous wording (e.g. "we'll look into it") that may be misread as a commitment. Surface to owner with the exact quote.

End the output with a 承諾辨識： section listing each classified statement.

When to write memory

After producing the structured triage, suggest (but don't auto-write):

• memory/commitments.md — add any confirmed commitment
• memory/open_questions.md — add open questions raised by the customer that the team can't answer yet
• memory/current_understanding.md — if this thread changes how we understand the product / behaviour

Owner uses approve-proposal flow (contributors) or directly commits (owner).

What NOT to do

• Don't invent customer environment details — say "未知"
• Don't escalate a proposed to confirmed commitment without explicit owner OK
• Don't write to commitments.md inline — go through propose / commit flow