fix(messages): confirm a short history page before calling it the start

[leemour]

Task

TASK-370:

• Part B: the personal history list (client.messages.list, the backward branch) used messages.length >= limit, so a short page counted as the start of the chat.
• Part A: the MAX side of the unread-coverage fix, fix(inbox): look at every chat for unread, new and review cli-messaging#548.

Why

• There is no paging signal from MAX. The chats.history spec (src/spec/operations/chats.ts) lists only messages in the answer. The 2026-09-25 web-tab captures record the requests, but no answer bodies. That is an absence of evidence, not proof. Nothing we hold says whether older messages exist.
• A short page may sit mid-chat. Whether MAX ever does this is unverified. If it does, store fetch marks the chat complete and messages list drops the hint for older messages.
• #chatsCut was set and never read (#readRestOfChats). When the one CHATS_LIST after the login was refused, or came back with 26 or more chats, the chat list still said hasMore: false.

What

• A short, nonempty page back is confirmed with one more CHAT_HISTORY: { chatId, from: oldest - 1, forward: 0, backward: 1, getMessages: true }. These are exactly web.max.ru's fields, with no interactive, so it marks nothing read.
  • It goes straight to the wire, so it adds no reactions request (opcode 180) and no contact-name lookups.
  • One message back means hasMore: true; none means the start of the chat. A full page still counts as more, and an empty page as the end.
  • Cost: one extra request for each short page. That includes the per-chat reads of inbox and inbox --new for a chat with fewer messages than --limit.
• The forward (after) branch is unchanged. It already asks for one more than the page.
• chats.list reports hasMore: true on its last nonempty page when the list was cut. An empty page past the end still says false, so a script that pages ends.
• Coverage limit, stated plainly: MAX gives the first ~41 chats (15 from the login plus one CHATS_LIST page). On a bigger account, the rest is now reported (partial), but still not read. Reading the rest means following the CHATS_LIST marker. That is new protocol behaviour on the real account, so it needs its own plan, and it is not built here.
• Until max-cli bumps to the cli-messaging release with #548, inbox on a cut account prints the old "only the 100 newest chats were looked at". The number is wrong, but it does say the list was incomplete.
• Tests: the history mocks now page by from/backward as MAX does (pagedHistory in src/testing/mock-max.ts), and the download fixture's time now matches its id. Counts that now include the confirming request are updated: inbox, review, store.
• Docs: docs/usage.md (a user page) is rewritten. docs/dev/ARCHITECTURE.md is corrected in place, with a correction mark.

Testing

• pnpm lint && pnpm typecheck && pnpm test:coverage && pnpm docs:check && pnpm parity:check all pass, on main b5d3691.

• New tests in src/client.test.ts:

  • a short page followed by one older message gives hasMore: true, with the exact confirming request and no extra reactions request;
  • a short page followed by an empty answer gives false;
  • a cut chat list says hasMore on its last nonempty page only.

  All three fail on main.

• The adapter contract was run against fix(inbox): look at every chat for unread, new and review cli-messaging#548's kit: 22 passed, 2 skipped, the same as on main.

• Noticed, not fixed: on main, src/store.test.ts "asks MAX for --page-size messages a page" fails when run alone (-t) and passes in a full-file run, so it depends on test order.

• Not verified: anything against live MAX. Fakes only.

🤖 Generated with Claude Code