feat(telegram): add inline keyboard to /session_recent, accept plain_dir in topic adoption - #3904

Reviewed as independent architect review. CI is green and the head contains current dev, but there are three things to fix before this should land.

1. Blocking — /session_recent silently lost disable_notification.

The old replyHtml helper carried telegramDisableNotification(this.opts.sound, "finalized"). The new inline sendMessage loop does not:

await this.botApi.call("sendMessage", {
  chat_id: this.opts.chatId,
  ...(threadId !== undefined ? { message_thread_id: threadId } : {}),
  text: chunks[i]!,
  parse_mode: TELEGRAM_PARSE_MODE,
  ...(i === chunks.length - 1 ? { reply_markup: { inline_keyboard } } : {}),
});

So a user who configured silent finalized notifications now gets a push for every /session_recent. The sibling reply helper right above still honours it, which makes the inconsistency easy to miss in review. Restore the flag (or keep replyHtml and give it an optional reply_markup parameter — that also removes the duplicated chunk loop).

2. Blocking — CHANGELOG.

This is user-visible twice over (new inline keyboard on /session_recent, and topic adoption now accepts a non-existent directory). packages/coding-agent/CHANGELOG.md is untouched; the PR checklist item is unticked.

3. Please split — the plain_dir change is a policy change wearing a UI change's PR.

existing_path → plain_dir in topic adoption means a Telegram message now causes fs.mkdirSync(cwd, { recursive: true }) in lifecycle-control-runtime.ts. The test being flipped was an explicit assertion (pending topics reject dir and worktree creates instead of spawning duplicate topics), so this is deliberately reversing a prior decision — that deserves its own PR with its own reasoning, not a bullet under a keyboard feature. Concretely, please state in that PR: what bounds the created path (the chat is full-trust, but normalizeLifecyclePath accepts any absolute path, so /session_create path /etc/whatever now creates a directory), and what happens to the directory if session creation subsequently fails — is it left behind?

Non-blocking notes:

• e.sessionId.slice(0, 12) in the button payload is fine — validateLifecycleTarget routes resume through sessionIdOrPrefix — but the test fixtures use 5-char ids (s-019), so truncation is never actually exercised. Worth one case with a full-length session id.
• Telegram caps inline_keyboard rows in practice; recent.entries is unbounded here. Cap the button list (the text list can stay long since it already chunks).
• Only the last chunk carries reply_markup, which is correct, but the chunk loop no longer awaits ordering failures — a rejected middle chunk is swallowed by .catch(() => undefined) and the keyboard still posts. Pre-existing pattern, just noting it now applies to a multi-message body.

gajae.pr-review-verdict.v1 merge-blocked sha256:c19e8ecf2e895399636e22399923c4f12841773f reviewer:architect evidence:https://github.com/Yeachan-Heo/gajae-code/pull/3904/files