5dive

Added

• codex agents now support --channels=telegram. 5dive agent create --type=codex --channels=telegram --telegram-token=… [--telegram-allowed-users=…] wires the full telegram-codex bridge the same
  one-flag way claude does: it writes the bot token to
  ~/.codex/channels/telegram/.env, seeds access.json from the allowlist,
  and at first boot appends the [mcp_servers.telegram] block plus the
  Stop / PreToolUse / Notification / PermissionRequest lifecycle hooks
  to the agent's config.toml. codex's first-run "Hooks need review" TUI
  prompt is auto-accepted once on first boot — codex then persists the trust to
  [hooks.state] so restarts never re-prompt. (codex's
  --dangerously-bypass-hook-trust flag only suppresses the gate for
  non-interactive codex exec, not the TUI, so it isn't used.) The plugin is a
  single shared checkout — resolved from $TELEGRAM_CODEX_PLUGIN_DIR,
  /usr/local/lib/5dive/telegram-codex, or the 5dive-plugins checkout, in
  that order — and server.ts resolves each agent's own state dir from $HOME,
  so one copy serves every codex agent. telegram only; no discord build for
  codex yet. Note: customer VMs need the telegram-codex plugin deployed to
  /usr/local/lib/5dive/telegram-codex (install.sh staging is a follow-up); on
  the control-plane host the 5dive-plugins checkout satisfies the resolver.

Added

• install.sh now stages the 5dive-cli skill under
  /usr/local/lib/5dive/skills/5dive-cli/ (whole-directory: SKILL.md plus
  references/). Pulled via tarball from 5dive-com/skills, mirroring how
  notify-user is staged. Pairs with the 5dive-api update.sh change that
  refreshes every agent's installed copy from this stage on the daily 03:00
  cron — so docs improvements (e.g. the new task/org reference sections)
  reach existing agents instead of being frozen at agent-create time.

Fixed

• 5dive-agent-start now dispatches grok and antigravity, fixing a
  crash-loop regression (unknown AGENT_TYPE: grok|antigravity). Both types
  were already first-class everywhere else in the CLI (TYPE_BIN, installer,
  auth, agent create), but the systemd launcher's case statement never got
  the matching branches — so agent create --type=grok succeeded, then the
  unit exited 2 on every spawn, racking up thousands of restarts. The
  per-type credential scrub also covers them now (same posture as
  hermes/openclaw — OAuth-via-file, no provider env vars).

Added

• Inter-agent mirror can post into a forum topic: when the group entry in
  access.json carries a message_thread_id, mirrored agent send/ask
  traffic lands in that thread (e.g. a dedicated "#5dive" topic) instead of
  the supergroup's General channel.

Fixed

• Inter-agent mirror now survives a group→supergroup migration. Upgrading a
  paired group to a supergroup (also how it gains forum topics) changes its
  chat id, and Telegram rejects sends to the old id with
  migrate_to_chat_id. The mirror now follows that, rewrites the stored group
  id in access.json (preserving owner/mode + the thread id), and retries —
  instead of silently posting nothing.

See CHANGELOG.md for the full breakdown.