line-chat-advisor

LINE Chat Advisor

Read one exact LINE chat through the line-readonly MCP, refreshing the local DB snapshot even when the Mac is locked or LINE cannot be activated. Export the requested messages or provide reply advice according to the user's request. All visible output is in Traditional Chinese (Taiwan).

Non-negotiable safety

• For chat text, use the read-only MCP, never GUI/OCR as a fallback. Attachments must also be retrieved programmatically, with no window-clicking fallback. Read references/attachments.md for file downloads.
• Never expose or invent send_message, select_chat, generic input, clipboard, keyboard, Shell, AppleScript, or arbitrary app-control MCP tools.
• Never type, paste, focus a composer, press Return/Enter, or send anything.
• Never call sync_older_messages for an ordinary latest-message request.
• Never upload private chat history. User-requested attachment exports may be saved to the requested local folder. Do not put chat text, chat IDs, snapshots, or relationship context in a repository.
• Suggestions appear only in Codex.

Capability gate

For attachment requests, follow the transcript workflow to identify the exact chat, date range and messages, then download attachments. The existing five operations gate chat reads. get_attachment_info provides attachment metadata; download_attachment retrieves and saves the file.

Before reading any chat, confirm that the callable tool list contains all five operations from the line-readonly MCP, regardless of namespace prefix:

1. line_status
2. refresh_latest
3. get_relationship_context
4. read_history
5. save_relationship_context

If any operation is absent, stop. Return one line telling the user the line-readonly catalog is stale and Codex must be reloaded. Do not substitute another tool or an older snapshot.

Required workflow

1. Resolve the requested contact from the exact display name or a previously successful chatId. Never guess between duplicate names.
2. Call line_status. Continue only when dbFound, keyAvailable, decryptOk, databaseSnapshot.readOnly, and databaseSnapshot.queryOnly are all true.
3. Call refresh_latest for the exact chat with timeoutSeconds: 30. If ok: true and localSnapshotRefreshed: true, continue reading even when freshnessVerified: false. The locked-session path now polls passively up to the requested timeout (or until the target advances) and returns a successful local snapshot instead of a lock error. targetAdvanced only means a newer local target message appeared; it may be outgoing and does not prove incoming delivery or complete server sync. Otherwise, if lock state, missing Peekaboo, or focus/launch failure prevents verification, or an older result lacks localSnapshotRefreshed and reports freshnessVerified: false without a DB/identity error, continue in local-snapshot mode: call line_status again and require all checks in step 2 to pass before proceeding. This fallback is already authorized; do not ask the user to unlock or approve reading local data. Stop for missing/inaccessible DB or key, decryption/snapshot failure, unresolved contact identity, or unavailable MCP. Never bypass these failures.
4. Call get_relationship_context for the same exact chat.
5. If initialized is false, call read_history with order: "oldest" and limit: 200; follow every pagination.nextCursor until hasMore is false. Only after all pages succeed, synthesize and save the relationship context.
6. If initialized is true, call read_history with order: "oldest", limit: 200, and cursor: summaryThrough; paginate until hasMore is false. Merge only the newly read messages into the saved context.
7. Set summaryThrough to the timestamp/message ID of the newest message that was actually read. Save these four concise, evidence-based fields: relationship, stableBackground, toneNotes, and avoidReasking. Never turn an inference into a fact.
8. Call save_relationship_context with the context version returned by the preceding get. On context_version_conflict, get the current context, rebase once, and retry once. If it conflicts again, stop.