Changelog
What changed in each release.
Newest first. Update with npm install -g trackline@latest. Follow it by RSS.
Run trackline remote enable once on your computer to turn these on.
Confirm with Google. Signed in with Google, you confirm a message from your phone by tapping Continue with Google. No passkey, nothing to scan. Google signs a token that names that one message, and your computer checks Google's signature itself, so our server still cannot write a message.
Or a code to your email. For any account.
trackline remote enableasks once whether to allow it. Only our server can check the code, so your computer takes its word for it: allowing email codes means someone in your inbox, or someone who took over our server, could send tasks there. One code covers your messages for the next 14 minutes.Adding a passkey on a new phone no longer goes round in a loop. It asked you to confirm it was you, with a passkey the phone did not have. It now offers Google, an email code, a passkey or an authenticator code.
An update says so. The app tells you when a newer trackline is out, and when a project is missing hooks a release added, with the one command that fixes it. Nothing is changed on your computer without you.
The phone runner picks up an update by itself. After
npm install -g trackline@latest, the runner started at login restarts on the new release once no agent it started is still working. Before, it kept the old one until you next logged in.trackline statusnames any agent in the project set up by an older release.No more dead end when a passkey is not on this device. Sending used to open the browser's passkey prompt and stop at "no passkey found". Now a "Confirm it's you" sheet offers every way that works: a passkey (Face ID, Touch ID, your device PIN, or your phone by scanning a code), your send password, or setting this device up. A device that lacked a passkey once is asked this way first after that.
A send password, for a device without a passkey. Run
trackline remote passwordon your computer, and any device signed in to your account can send by typing it. Your computer checks it, as it checks a passkey, so our server still cannot write a message. It needs 12 or more characters, turns off after 5 wrong tries, and is off unless you set it. One limit, said plainly: someone who stole our server's data could try guessing it offline, which a passkey does not allow.trackline remote password offremoves it.
Run trackline init in each project to pick up the new hooks, then
trackline remote enable there. Codex asks you to trust its hooks again.
- Your message stays yours. When an agent's memory filled up, the summary it writes for itself replaced your last message in the chat, and the checks read it as your request. They now ignore it, and old chats hide it.
- A chat keeps up with the computer. A message you send in Claude Code or Codex shows as soon as you send it, and you can read what the agent is saying while it works, not only when it finishes.
- Steps say what they do. A command now shows the agent's own line about it
where Claude Code writes one, instead of "Bash". The full command stays on your
computer unless you run
trackline share commands onin a project. - Waiting for you. When an agent stops on a permission prompt, the chat says so and your phone tells you.
- Quiet while you are looking. A chat you have open does not also send a notification.
- A phone message can reach a Claude Code chat that is open. Run
trackline remote enable --open-chats. The message waits for the agent's turn to end, or wakes the chat if it is idle, and runs with that chat's own permissions. It cannot reach a chat that has not finished a turn since this was installed. - Long sessions stopped slowing every step. The hook read the whole session log before each action, about a second on a 178 MB log. It now reads the end.
- Agents are individuals. Every reply and chat shows its agent's name in bold, with the model that answered. Tap a name to rename it, add it to your team, or move a chat to someone already on it.
- Chats, Team, Settings. Home and Sessions are now one list, with what needs you first. Team holds the agents you chose to keep, each with a profile of what it has actually done. Helpers show as one line in the chat.
- Positions. Give a team member a title and a description, such as CTO, and it holds that job across a resumed chat in Claude Code and Codex. It reaches the agent with messages sent from trackline, and it never changes what the agent may do.
- One agent can ask another. An agent you started from trackline can ask a team member to do something. You read the whole request and approve it with your passkey, and the result can go back to the asker once. A handoff that changes the files telling agents how to behave fails, however it did so.
- A reply never lands on a chat open on your computer. A Claude Code chat
that is open stays yours: the phone says so and sends nothing. A Codex chat
open in a terminal can take the message if you run
trackline remote enable --open-chats. It then runs with that chat's own permissions, not the narrower ones of a phone run.
- Reply from your phone in every project. Pair your phone once, during
trackline connect, and every project you connect takes messages, with no command per project. Every message still needs your passkey, and your laptop still checks it. On a laptop already paired, runtrackline remote enableonce in a terminal after updating; you get one email listing the projects.trackline remote disableturns a project off, and it stays off. - The reply box is never locked. You can always type, and a line underneath says what would let it send.
- One conversation. A reply from your phone continues on the same screen, with the history above it, and a notification opens there. Your messages stay in place, including ones an agent answered in words alone. Long chats load earlier messages as you scroll up, and coming back opens at a "New messages" line. Any running task has a Stop button.
- Sessions have names. A session takes the agent's own title, or its first real request, and keeps it. Rename, pin and archive from the three dots.
- Delete and clear. Swipe left or use the three dots, with six seconds to undo. Hold a row to choose several. Clearing a chat keeps its name. Both remove the words from trackline only, never from your laptop, and your on-track numbers do not change.
- Search names and replies, filter by project, and see every session rather than the latest 50.
- Feels like an app. A phone opens on your chats, with what needs you first and unread ones marked. Tabs change instantly, the app follows your phone's dark mode, the keyboard stays out of the way, and the icon shows a count. A turn's work reads as one line, like "Edited 4 files, Ran 2 commands", and opens to the steps. With no signal the app still opens on the last list.
- Live. The app updates the moment something changes, and shows what the agent is saying while it works. It uses less of the server, too.
- A sleeping laptop no longer loses your message. A message sent while the laptop is asleep runs when it wakes, for up to 12 hours, once, and Stop cancels it. A message with images still needs the laptop awake. While a task runs, a Mac stays awake unless its lid is closed.
- Write while the agent works. Your message waits in the chat and Face ID sends it when the agent finishes.
- Longer replies. The limit goes from 8,000 to 32,000 bytes, and a reply cut at the limit says so.
- A project in use at your desk no longer pauses after 30 days.
trackline remote enableneeds a real terminal. It accepts changed settings and lifts a stop made from the site, so only a person may run it. A stop made from the site now also survives the runner restarting at login.trackline connectnever shares without you. Run by an agent with no terminal, it keeps the answer you gave before, or does not share.
- Send images from your phone. Attach up to four screenshots or photos to
a task or a reply; the agent on your laptop looks at them first. Each
image's hash is signed with your passkey, and the laptop takes only images
that match, so the server that carries them cannot swap one. They are saved
in the project under
.trackline/attachments/, out of git, and the server keeps them only until the laptop has them.
- Cursor from your phone. Tasks, replies, and Allow once or Keep blocked
now work with Cursor as they do with Claude Code and Codex. It runs headless
with its sandbox on and never with
--force. - A task from your phone cannot write outside its project, whatever the agent. Measured: Cursor's file editor, unlike its shell, is not confined by its sandbox, and a mistyped path wrote a file elsewhere on the laptop.
- Fix: an approval now matches a project reached through a symbolic link,
such as
/tmpon macOS.
- Deployed agents on the dashboard.
trackline serve --connectsends each checked conversation to your account, so a production agent appears under Production beside your coding agents, and a broken policy or an incident alerts your phone. What is sent is what trackline concluded: tool names, findings, incidents and counts. Never what a customer or the model said.trackline traces --connectsends a recorded stream the same way.
- A new check:
workaround. After trackline blocks a file, the same file reached another way: a script that writes it, a build step, a command. The finding shows the blocked call and the route taken after it, side by side. test-weakenedshows what changed.trackline showlists the test's lines as they were and as they are, so a test that really needed updating can be told from one that was made to stop failing.
- A new check:
test-weakened. A test skipped, or its assertions removed or loosened, instead of the code being fixed. It reads the change to the test itself, so it catches the two-line edit that scope and diff size miss. - Remote, hardened after a security review. Pairing a phone can no longer be attacked by a compromised server, which could have found the code in the answer it relays. The laptop now records trackline's settings and hook wiring when you turn remote on, and refuses a task from your phone if they have changed since; an agent started from your phone cannot edit them, and approvals from your phone are kept outside the project. A conversation can only be continued in the project it happened in. Turning remote off stops any agent it started.
- After updating, run
trackline remote enableagain in each project that takes tasks from your phone, and pair again if asked. Earlier versions cannot pair with the site any more.
- Multi-turn evals.
trackline-gategains a fourth suite: conversations scored turn by turn, withmemorySafetyfor turns that are only correct because of an earlier constraint or refusal. Optional: a project withoutdatasets/multi-turn.jsonlreports it as skipped. - A one-row regression now fails the gate in fixture mode. Live runs and the judge keep the 0.05 tolerance.
- Any OpenAI-compatible judge.
EVAL_JUDGE_BASE_URLandEVAL_JUDGE_API_KEY, with retries on rate limits and dropped connections. - Faster live runs.
--concurrency=N, default 4. trackline-gate initstarts empty. The fintech sample istrackline-gate init --example=fintech-support.
- Allow once, or keep it blocked, from your phone. When trackline stops something in a session started from your phone, the app offers Allow once and Keep blocked. Allow once lets that one action through, in that session, once, and the agent retries it; Keep blocked tells the agent not to try again. Both are signed with your passkey.
- Fix: turning remote off with a test or development config folder no longer stops the real runner.
- Reply from your phone. A task or a session can be continued from the app: the reply goes to the same Claude Code or Codex session on your laptop, with the same permissions and sandbox as a new one. Works for sessions you started in the terminal too.
- macOS permission prompts. When an agent started from your phone seems to be held by a Files and Folders prompt on the Mac (Documents, Desktop, Downloads, iCloud Drive), the app says so, so a task does not look frozen.
- Remote prompts. Send a task from your phone to Claude Code or Codex on
your own laptop, and watch it run.
trackline remote enableturns it on for one project and pairs your passkey by a code shown on the laptop. Every message is signed by your passkey and checked on the laptop, so the server that carries it cannot write one. Remote sessions block writes to secrets and new dependencies, and run only where trackline is watching.trackline remote disable --allturns it off everywhere and forgets the passkeys. - An optional account.
trackline connectlinks a project to a trackline account and a live dashboard: what the agent is doing, its replies, and alerts.trackline accountandtrackline disconnect. Nothing uploads without it. - Setup.
trackline initsets up every installed agent when not told which, andtrackline statussays which agent is set up but has never reported, and why.
- Cursor.
trackline init --host cursorwires the hook into Cursor. It sees the agent once the agent is started inside the project. - Any MCP client.
trackline mcpruns an MCP server withcheck_actionandget_rules, so an agent can ask before it acts. It advises. It cannot block. - Production traces.
trackline traceschecks OTLP exports, protobuf or JSON.trackline servereceives them live and can post an alert. Both check tool policy (tools that must never be called, and approvals that must come first), outages, loops and drift. In production it alerts and never stops anything. - The judge, per action. The optional judge now rules on each action rather
than the work as a whole.
trackline review --jsongives machine-readable output, and the judge runs through Claude, Codex or Cursor's CLI. - Codex. Reads Codex transcripts, including codex-cli 0.155's Bash tool and headless requests, instead of reading an unknown format as silence.
- Fixes. Commands that can write anywhere are no longer reported as reads. An
unreadable rules file is reported instead of read as no rules. Every
subcommand answers
--helpinstead of running.
The first release, for Claude Code and Codex.
- The hook. Runs before every action the agent takes, in about 14 ms.
- The checks. Off-limits paths, new dependencies, scope, diff size and repetition.
- Modes. Warn, ask or auto, set per check, with remembered approvals:
trackline allow,allowedandrevoke. - Commands.
init,status,doctor,showandreview. - The judge. Optional and off by default, for the one question counting cannot answer: whether the work serves the request.
- Install.
npm install -g trackline, with a prebuilt binary for each platform.