🔧 update: inject sender userId into LLM context so owner is correctly identified via web dashboard

[Copilot]

When the owner interacts through the web dashboard, the LLM was misidentifying them as a "friend" and refusing owner-only actions (e.g., changing the companion's name). The system prompt declared cli:owner as the owner userId, but no userId was ever passed alongside incoming messages — leaving the LLM to guess, and incorrectly infer that a web session couldn't be cli:owner.

Change

packages/core/src/loop.ts

• Added sanitizeUserIdForPrompt() to strip backticks, square brackets, and newlines from userId before embedding it in any prompt, preventing a crafted userId from breaking the marker format or injecting additional instructions.
• Sender identity is now appended directly to the system prompt for the turn rather than injected as a separate system message immediately before the user message. This ensures the identity covers the entire turn — including all tool-follow-up user messages — without needing per-message injection.

// Before
const messages: Message[] = [
  { role: 'system', content: systemPrompt },
  ...history,
  { role: 'system', content: `[Current message sender: userId = \`${userId}\`]` },
  { role: 'user', content: sanitizedMessage },
];

// After
const senderIdentityPrompt = `\n\n[Current message sender: userId = \`${sanitizeUserIdForPrompt(userId)}\`]`;
const messages: Message[] = [
  { role: 'system', content: systemPrompt + senderIdentityPrompt },
  ...history,
  { role: 'user', content: sanitizedMessage },
];

packages/core/tests/loop.test.ts

• Added assertions that the [Current message sender: userId = ...] marker is present in the system prompt content and that the immediately following message is the user turn (no separate system entry in between).
• Existing tool-summary at(-2) / at(-1) assertions remain correct and unchanged.

The userId passed here is resolved server-side from configManager.get('owner.ownerId'), so the LLM now has a ground-truth sender identity to match against the ownerId in the system prompt — regardless of which channel the message arrived through.

[Copilot]

Pull request overview

Updates the core agent loop prompt construction to explicitly provide the LLM with the current sender’s userId, so owner-vs-friend authority rules can be applied correctly when chatting via the web dashboard.

Changes:

• Adds a system message marker ([Current message sender: userId = ...]) immediately before the incoming user message in the LLM message array.
• Documents the rationale inline to reduce future regressions around owner identification.