Pass message author identity into the AI conversation history

The support bot cannot tell who is speaking in a thread. This is the direct cause of the second incident behind #172: a maintainer asked a community member a follow-up question, and the bot answered the maintainer, explaining in its reply that it "can't see other users' replies."

It was telling the truth. It can't.

Where

packages/outpost/queue/src/handlers/ai-response.ts, building conversationHistory:

const conversationHistory = ticket.messages
    .filter((m) => m.type !== 'SYSTEM' && m !== openingUserMessage)
    .map((m) => ({
        role: (m.type === 'USER' ? 'user' : 'assistant') as 'user' | 'assistant',
        content: m.content,
    }));

Every non-SYSTEM message collapses to a bare user or assistant role derived from Message.type alone. Message.author is discarded. So the reporter, a maintainer, and an unrelated third party all arrive at the model as indistinguishable user turns.

Why it's worth doing

Message.author is already populated on every row — InboundHandler writes "${platformUsername} (${platformUserId})" for every platform. The data exists and is simply never handed to the model. This is a small change.

It also matters more now, not less. #172 caps the bot at one response per ticket, so the single answer it does write has to be right. Handing it a conversation where it cannot distinguish "the maintainer already solved this" from "the customer is still asking" is exactly the input that produced the incident.

Considerations

• Deciding the shape matters more than the code. Options include prefixing content with the author, using the name field on message parts where the provider supports it, or passing a structured participant list alongside the history. Worth one deliberate decision rather than the first thing that works.
• Team membership is already computable (isTeamMember in InboundHandler) and is the distinction that matters most: "a maintainer has spoken here." That overlaps with the human-presence gate work, so the two are worth designing together.
• Author strings are user-controlled on most platforms. Whatever format is chosen should not let a display name masquerade as a role marker in the prompt.

Start in packages/outpost/queue/src/handlers/ai-response.ts at conversationHistory, then inspect InboundHandler where Message.author and isTeamMember are populated. Decide and document a safe way to preserve author identity and maintainer participation without allowing user-controlled names to act as role markers. Done means the model can distinguish participants and team presence in the conversation history.