TeleFlock v0.7.0

TeleFlock 0.7.0 is built for agents that work while you sleep: they pick up again after usage limits, survive herdr restarts, crashes and reboots, and can answer their own questions when you say so. It also adds a dozen new tools around the board, the session and your hosts.

Updating from 0.6.x or older: apps from v0.2.0 to v0.6.1 never checked for updates by themselves (a bug fixed in this release). Open Settings → Updates → Check for updates once. From 0.7.0 on, automatic updates work again: the app downloads a new version in the background and installs it once TeleFlock has been in the background for 20 seconds (hidden, minimized or unfocused, and Quick capture not in use); while you are using it you get "TeleFlock vX is ready" with Restart now. The host daemon (tf-hostd) changed, so the first device that updates restarts it once; your agents keep running. Agents that were already running keep the previous TeleFlock bridge loaded until their omp process restarts (for example Mark done and then Resume; F5 alone does not reload it), so auto-resume after usage limits and the other overnight features cover them from then on. Some new features need a 0.7.0 host: older hosts tell you to update them.

Overnight runs

• Auto-resume after usage limits. When an agent or its subagents stop on a provider usage or rate limit after omp's own retries, the host schedules a resume for when a limit window reopens (counting every account omp can switch to), then continues the subagents and the agent where they stopped, without a new prompt. Provider overloads and server errors are resumed too, with a growing pause (15, 30, then 60 minutes) and the words "Overloaded" or "Provider error" instead of "Rate limit". The limit errors are kept out of what the model sees afterwards. Works with no app connected; the card shows when it resumes, with Resume now and Don't auto-resume (Settings → Auto-resume to turn it off per host).
• While omp still retries, an agent shows "Rate limit · retrying in 4:54 (1/10)" instead of looking quiet or asking you to "answer in the terminal". Repeated limit errors fold into one row.
• Decide for me. If a question waits longer than you choose, the host picks omp's ★ recommended option — per agent, per Dispatch message, per automation or as a host default, optionally only while no TeleFlock app is connected. Approvals and plan reviews are never answered for you.
• Resume (or F5) continues an agent you stopped, or one whose turn failed, without a new prompt, like omp's own F5. Failed subagents have Retry.

Nothing lost when something crashes

• herdr health. The host notices within seconds when herdr stops or crashes, tells you instead of showing stale "running" cards, and starts it again with a growing pause. A herdr that keeps dying right after it starts is left alone after three tries and the host tells you; Restart herdr is always there.
• Agents come back. When herdr stops or crashes, an omp ends or the host restarts, agents no longer vanish: each stays on the board as an ended card saying what happened and what it was doing, and is relaunched under its name, in its folder, with its model and settings, even with no app connected. A turn that was cut off continues when its last step is safe to repeat (a reply being written is written again); a command that may have run already is never run again by itself: the agent is told what happened and checks first. An ended card shows when it comes back; Continue brings it back at once.
• What doesn't come back by itself: an omp you stopped yourself (for example with kill), an agent that ended more than three times in ten minutes ("keeps crashing"), and an agent that ended before omp saved its first reply. They stay as ended cards until you press Continue; the last kind keeps its prompt and images on the card, and Continue starts it again with them.
• After a reboot the host brings agents back through a small boot service, installed once there are agents to bring back. It needs systemd (Linux, as root or with lingering enabled for your user); the setting says when a host can't have one, as on macOS and Windows hosts, where agents come back once the host daemon runs again. Turn it off per host in Settings → Auto-resume → Bring agents back after a reboot.
• Sturdier saved state: the host's card, ticket and recovery files are flushed to disk on every save, and a file the host can't read is moved aside (.bad-…) for inspection before it starts over. If the auto-resume settings file is damaged, auto-resume and bringing agents back after a reboot stay off on that host until you turn them on again.
• herdr updates without restarts. A newly installed herdr is loaded into the running server by live handoff once the host has been idle for a while (20 minutes by default, and on Linux only while no herdr window is open), so agents and terminals keep running. When only a restart would load it, the host shows "herdr update ready" and waits for you to press Restart herdr.

In the session

• Errors that say what happened: failed requests show their message and status, provider refusals (for example a cyber safety refusal) their category and explanation, and omp's warnings and fallback notices appear in the transcript. omp's xd:// mount notice no longer floods the chat.
• Large pastes become attachments, with omp's choice of an attachment block, a file the agent reads on demand, or inline text (up to 16 pastes per draft). A pasted log or path that starts with / is sent as text, not run as a command.
• Find (Ctrl+F) across the transcript, tool output included; Past sessions searches the text of every session.
• Inspector: the btw side question moved up under Session; background jobs with output and Stop; the session's documents (plans, reports, handoffs, subagent outputs; "Docs" on the phone); Context X-ray with compact-with-focus (dropping images asks first).
• Subagents that finished collapse into a group and show green or red by their real outcome (they no longer turn blue). One that a crash cut off shows grey Stopped, not Done; the label replaces "Aborted", and the group reads "N done · N stopped".
• Model roles for one agent: set every role's model, and fallback chains, for a single agent at launch or while it runs, without touching the host's settings (needs a 0.7.0 host). The omp settings page has a guided editor for Retry Fallback Chains (Model → Retry & Fallback).

Board, Dispatch and capture

• Review and Done cards show omp's recap of where the agent stands.
• Make it green runs omp's /green on a failing pull request; Rebase and resolve hands conflicts to the agent.
• Dispatch Auto also weighs how much of the model's subscription window each host has left.
• Quick capture sends a task, a screenshot or an answer from any app. It opens from the tray, from teleflock://capture or from a global shortcut that is off until you set one in Settings (Ctrl+Alt+Space, or Cmd+Shift+Space on macOS, is suggested). On Android, share text, links, screenshots or GitHub issues to TeleFlock.

Hosts

• Move an agent to another host, with its worktree changes, session, subagents and documents. A step that fails rolls back what the move did, and when it can't yet tell whether the agent started on the other host, it keeps the agent parked here and waits until it knows instead of guessing. A move is refused while background jobs run or another live agent works in the target checkout and files would change, and "Try again" after a rollback reopens on the agent that was resumed.
• Make this host like that one: copy omp settings, model roles and MCP servers after a preview of every difference (omp ≥ 18.13.3); sign-ins optionally, encrypted between the hosts.
• Clean up: worktrees with pull request state, size and age, bulk removal of merged and stale ones with the usual safety checks, omp storage cleanup without touching sessions, and a low-disk warning. Folders that are missing are never removed in bulk (one at a time, with a confirm), symlinked paths are recognised, the confirms say that files git ignores are deleted too, and a merged or closed pull request only counts when it ended at the worktree's HEAD.
• Skills and plugins page next to MCP servers; GitHub CLI sign-in from the app; Host setup checks git's name and email and offers updating herdr's omp integration.
• Notifications: native on Linux and macOS; the host notifies the device you are using first and the others if nobody answers; with no app connected it can push to ntfy, Pushover or a webhook.

Usage and fun

• The Usage page reads omp's own usage index, now kept up to date by the host: complete history in your time zone by model, provider, project and agent type, cache hit rate, speed, top tools, streaks and records. The sync keeps going as long as it makes progress; the background warm-up is off on Windows and skipped on Linux laptops running on battery.
• Flock Wrapped: your week on one shareable card.
• Wall (2×2 or 3×3 live transcripts) and Ambient mode (one orb per agent) for spare screens.

Fixed

• Automatic updates never ran (since v0.2.0): the background check sent a malformed settings request and silently counted updates as off.
• A session's transcript could freeze while the agent kept working when omp cut an emoji in half in subagent progress; such text is now repaired, and a stalled feed reloads itself.
• WSL and SSH hosts whose login shell is nushell (or fish, tcsh and others) failed to connect with "Syntax error: Unterminated quoted string"; commands now reach only a POSIX shell, and deterministic failures stop retrying and offer Copy diagnostics. (With csh or tcsh a PATH set only in .login is not picked up; use .cshrc or .tcshrc.)
• The tray animation ignored the saved motion setting until it was changed.
• On a phone, a transcript shorter than the screen opened with its first rows missing.

Good to know

• Questions still waiting when a host's daemon restarts are announced again; ntfy replaces its copy, a webhook or Pushover receives a second one.
• Without herdr's omp integration (Host setup offers it), herdr reports an agent as idle while its background jobs run, so the "Waiting for background jobs" line does not show.
• Windows does not install an update when the app quits (its installer relaunches the app); it installs while the window is away or when you click Restart now. A .deb update always waits for your password.

Downloads

Platform	File
Windows 10/11	TeleFlock_0.7.0_x64-setup.exe
Android 7+	TeleFlock_0.7.0_universal.apk
Debian 12+ / Ubuntu 22.04+	TeleFlock_0.7.0_amd64.deb
Other Linux (glibc 2.35+)	TeleFlock_0.7.0_amd64.AppImage
Arch Linux	teleflock-bin-0.7.0-1-x86_64.pkg.tar.zst
macOS 11+	TeleFlock_0.7.0_universal.dmg (attached by CI shortly after release)

Every file has a .sha256 next to it.