Repository navigation
Releases
v0.5.167: Pull requests from your agents stop colliding, and Helme tells you the order to merge them in
Pull requests from your agents stop colliding, and when several sit on one another Helme tells you the order to merge them in.
Includes v0.5.165 – v0.5.167
New
- A merge order on your list. When pull requests sit on one another, Waiting for you shows one item, "Merge in order: 1) #12 then 2) #13 then 3) #14", and each merge item says "step 2 of 3, after #12". Helme reads GitHub every minute and only reads: it merges nothing and changes no base.
- Traps are named before you press Merge. If the pull request under another one was already merged or closed, the one on top would go into a branch that no longer leads to
main; an item says so and shows how to point it atmainwith one press on GitHub. Pull requests that nobody has touched for more than 7 days are listed together in one item, so you can close them or keep them. - A conflict is handed back to the agent. When a pull request from one of your agents conflicts, or the one under it was merged, that project's agent gets one task to bring
mainin. This happens no more than once an hour for each pull request, and the item of the pull request says so. - Agents avoid the collision in the first place. Every agent now starts a new branch from the newest
main, bringsmainin with a merge commit instead of stacking pull requests, and says in its report whether it checked that its branch has no conflict.
Good to know
- Nothing here merges for you. Merging stays your press on GitHub, or the switch "Merge by Helme" when you turn it on.
Viết bởi: helme-web project agent · tài khoản claude-main · model claude-sonnet-5-5 · phiên 3b14d68b-d5d4-4506-af1a-7be97f77d54c
v0.5.164: The Chief of Staff can now set reminders that survive a restart
The Chief of Staff can now set reminders that survive a restart, so a reminder for next week still arrives after Helme is closed, restarted or installed again.
New
helme remind.helme remind add '2026-10-31 08:00' 'Check the Pro price with the owner'keeps the reminder in a plain Markdown file,reminders.md, that you can read. The time is a date and hour of this computer (a date alone means 08:00) or how long from now:30m,2h,1d,1h30m.- It arrives once, at the time. Helme looks at the file every minute and types one line into the Chief's chat, marked as a notice from Helme and not as your words, so a reminder can never be taken for an answer from you or approve a ticket. If Helme was off at the time, the reminder comes when it is back and is marked late.
- A reminder can be a task.
helme remind add 1d 'Run the tests again' --project shop-web --taskgives the text to that project's agent as a task. If the task does not reach the agent, the Chief is told with the full text, so nothing is lost without a word. - A Reminders box on the System tab shows what is waiting, with its time, and the last ones delivered.
helme remind listandhelme remind done <code>show and withdraw them.
Improved
- The Chief of Staff no longer relies on its own timer for a reminder that must come: its instructions now point it to
helme remind.
Viết bởi: helme-web project agent · tài khoản claude-main · model claude-sonnet-5-5 · phiên 3b14d68b-d5d4-4506-af1a-7be97f77d54c
v0.5.163: A new lane or night worktree can now reuse the libraries your project already has installed
A new lane or night worktree can now reuse the libraries your project already has installed, so it is ready in a fraction of a second instead of waiting for a full install.
Includes v0.5.162 – v0.5.163
New
- Shared libraries for lanes. When a lane or a night worktree opens, its
node_modulesfolder is made out of links to your project's own, so nothing is copied and no extra disk space is used. On a large project this takes about a tenth of a second; copying the same files with hard links took 14 to 29 seconds, and a freshnpm citakes much longer and adds hundreds of megabytes. - Safe by design. Helme only does it when the lockfile and the dependencies in
package.jsonare the same as the project's, both folders are on the same drive, and the worktree has nonode_modulesyet. In every other case nothing is linked and the agent installs as before. The reason is written in.internal/shared-modules.json. - Removing a lane never touches your libraries. Helme takes the links away before it asks git to remove the worktree. If a package changes inside the worktree, npm writes a new folder there and your project's copy stays as it was.
Good to know
- It is off until you turn it on. Add
- Share modules: shop-web(project ids, separated by commas) tosettings.md, or set"shareModules": trueon the project's entry. - A tool that writes inside a package folder in place (a postinstall script,
npm rebuild) changes the project's copy too, because the folder is the project's. - Windows only for now. A worktree on another drive than the project is not served.
Viết bởi: helme-web project agent · tài khoản claude-main · model claude-sonnet-5-5 · phiên 3b14d68b-d5d4-4506-af1a-7be97f77d54c
v0.5.161: New agents never start on an account with no quota left
Helme now looks at how much quota each Claude account has left every time it opens or moves an agent, and never puts one on an account that has none.
New
- No agent on an account with no quota. An account with 0% left, or in a usage limit, never takes a new pane, a move, a restart or a background job, even when a project pins it or a ticket names it. Before, a pane opened while the other accounts were busy could end up on the old account and stop on a limit until the next morning.
- Fresh numbers. The quota of each account is measured again before a pane is placed when the last numbers are older than 15 minutes or a window has just reset.
- A pane waits instead of starting on the wrong account. When no account has room, the pane opens with a shell and a line that says so, keeps the work it was given, and Helme starts the agent in it by itself as soon as an account has room.
- Slow starts are waited for. Claude Code gets 90 seconds on an idle machine and up to 3 minutes while many agents run; a start that fails gets a second try, and only then the account is skipped, for 4 minutes instead of 30.
- The System tab says why. Each account in the Accounts section says whether it takes new panes now, why not, and when it can again; panes that wait are listed with the reason.
Good to know
- The accounts' numbers are now asked every 15 minutes instead of 25.
v0.5.160: Heavy commands take turns, and the agents' browsers stay few
Twenty or thirty projects can now run on one PC without all of them building at the same moment.
New
- Heavy commands take turns. A project agent's full test run, build, type check, browser audit or deploy tool asks Helme for a place first. Three run at once (the number follows the CPU and the free memory); the next one waits in line, and if the wait is long the agent is told what is running and tries again a minute later.
- The Machine box shows it. On the System tab: how many heavy commands run, how many wait, and which ones.
- A ceiling for the agents' browsers. At most three of the agents' own hidden Chrome browsers stay open; the idle ones over the ceiling are closed, oldest first, and an agent that needs its browser gets a new one by itself.
- Leftovers are cleaned. Browser and helper processes that Helme started and that no agent owns any more are closed. Your own Chrome, Zoom and every other program are never touched.
Good to know
- Change the number with
- Heavy jobs at once: 3(oroff) and- Agent browsers at most: 3insettings.md.
v0.5.159: Idle agents sleep, and wake up where they left off
Agents that have nothing to do now go to sleep, so twenty or thirty projects can stay open without keeping twenty or thirty programs running.
New
- Idle agents sleep. A Claude agent that has rested for 15 minutes is closed. Its pane stays, shows the shell, and the Status tab says "Sleeping".
- It wakes up where it left off. A task for it, your click in the pane, a key you type, or
helme restart-agentstarts it again on the same conversation, so it still knows what it was told. What you type while it starts waits and lands in the agent's box. - Only when it is safe. An agent does not sleep while a task is queued or in hand, a ticket of its project is approved or in work, it waits for you or for a usage limit, files in its folder are not committed, you have just used the pane, or your own message is in its box. The Chief of Staff, the Reviewer, lanes and Cursor and Gemini agents never sleep.
- The Chief gets one line when a round put agents to sleep.
Good to know
- Turn it off with
- Agent sleep: offinsettings.md; change the rest time with- Agent sleep minutes: 15. - Waking takes a few seconds, the time Claude Code needs to start.
v0.5.158: See what every project costs your PC, and agents stay out of your way
A new Load box on the System tab shows what each project's agents cost your PC, and agents now run at a low priority so your own screen, Zoom and browser stay smooth.
New
- Load box. CPU and memory of the machine and of the agents, then one row per project (the Chief of Staff and the Reviewer too): number of processes, CPU, memory, and what kind they are. The busiest project is first, so when the PC is slow you see who it is.
- The last hours as a line. The machine's CPU and the agents' CPU over the last six hours, kept on disk.
- A warning when the CPU stays high. Three minutes at 85% or more turns the heading red, names the heaviest project, and puts the reason on the red dot of the System tab.
- Low priority for agents. Every agent pane and every background run starts at Below normal priority, and what grows from it follows. Your browser, Zoom and screen get the CPU first.
Good to know
- Helme only lowers processes that belong to its own agent panes. It never touches your browser, Zoom or a plain shell tab, and it closes nothing.
- To turn the low priority off, put
- Low priority agents: offinsettings.md; it applies to panes started after that. - The reading takes about a second every 30 seconds.
v0.5.157: Your unsent message no longer holds your agents back
Helme no longer waits for the message you are writing: your agents keep working while you type, and your words are still there afterwards.
Fixed
- An agent no longer stands still for a message you are not writing. Helme used to guess from your keys that you were still writing in an agent's chat box, and the guess could stay wrong for hours (after Esc Esc, after you removed a file path the paperclip had typed, after an agent restarted). Three approved tasks of one project waited three hours for it. Helme now also reads the box on the screen: a box that is empty there ends the guess.
New
- A message you did not send is set aside while a task is typed, then given back. Claude Code keeps it with its own stash (Ctrl+S): the task, and the commands before it (
/clear,/model,/effort), are typed and sent, and your message is in the box again as you wrote it, with the images you attached. Cursor and Gemini have the text read from the screen and typed again after the task. - A task only waits while you are pressing keys. A second and a half after your last key it goes in.
helme restart-agentis no longer refused for your message. It is set aside for the old agent's exit and typed again in the new agent.- Pictures on the phone page. When you are away from the computer, an item that needs your approval now shows its before and after pictures on your phone, under its words and its how-to. Tap a picture to enlarge it, then move between them with the arrows, a swipe or the arrow keys; a counter such as "2 / 4" shows where you are.
- Small copies, uploaded once. Helme makes a small copy of each picture (long side at most 1600 px, usually well under 600 KB) and sends it together with the normal update of the page, so no extra run starts. A picture that did not change is never sent twice, and a picture that could not be sent is tried again later.
- Only pictures that belong to the item. Helme shows only picture files saved with that item. Links to other places, files that are not pictures and files above 8 MB are left out, at most 8 pictures for one item.
Good to know
- Your keys are held back for the second or two it takes, so nothing you type lands inside the task.
- Helme never presses Enter for your message: it only puts the words back.
- The phone page has to be published once by the owner of the page before pictures appear; until then it says the picture is still loading.
- When the box cannot be read (the agent is not at its prompt), a message Helme saw you type still makes the task wait, as before. The Chief of Staff's own chat still waits for you to finish a message.
Viết bởi: Chief of Staff · tài khoản claude-main · model claude-sonnet-5-5 · phiên 82e840cb-8528-4bc3-86e0-a6d55b81f926
v0.5.156: Tonight's plan on one list, and your design rules reach every agent
Helme now lays out the night shift for you to look over before it starts, and every agent that builds a public website follows your design, search and brand-voice rules.
Includes v0.5.155 – v0.5.156
New
- A plan for tonight, at 22:30. One item on your waiting list says which tickets will run overnight, which are held and why, and how much quota each Claude account has left. Approve it, or ask the Chief of Staff to drop a line you do not want. If you do nothing, the night runs exactly as the plan says. A ticket that is dropped is not started and not approved by the night autopilot; it stays as it is.
- The morning report compares the plan with the night. A new section of the 06:30 report sets each line of the plan against what the night actually did.
- Search, answer-engine and brand-voice rules are the default for websites with a public page. Project agents, lanes and departments are told to give every page its own title and description, a real FAQ, structured data and a written voice for the product's words, and the Reviewer asks for them too. A project with no public page is not forced.
- Your design standard comes first. An agent that changes a screen is told to design for a technical, international audience, to avoid anything that looks machine-made, and to show pictures for your approval before the real build.
- The exact model name sits beside the alias. Identity files and the Agents tab show the full model name next to a short name such as "sonnet", and the real context size with where the number came from and the day it was measured. A size nobody has measured says so.
- A set of outside skills, pinned to one version. Helme can install a set of third-party skills for Claude agents (market research, a "what next" triage). It uses one fixed version, runs nothing from the set, keeps it read-only for agents, and holds back the skills that ask for broad access until you allow them.
Fixed
- No more false "could not read its rules file" notices. Helme wrote an agent's rules file in place, so a reader could catch it empty, and a note left by Helme's own tests was believed as real. Rules files are now written whole, a note is checked before an agent is restarted because of it, and leftovers older than three days are cleaned up.
- Pushing a branch that already contains the main branch works again. The check that holds back changes to workflow files now looks only at commits that no remote has yet.
Good to know
- The plan only ever makes the night do less: it never starts or approves anything by itself.
- The skills held back stay hidden until you add their names to
settings.md.
v0.5.154: The phone waiting list uses far fewer Claude runs
The waiting list on your phone now updates in minutes without a run of Claude for every small edit, so it spends far less of your quota.
New
- Changes to the waiting list are gathered before the phone page is written. Helme used to start a background Claude run for nearly every edit of the list: 1,406 runs between 14 and 22 October on the light account. Now it waits by the kind of change, counted from the first time the change is seen and never pushed back: an item added or closed, or with other buttons, waits at most 5 minutes; a change of only the words or the how-to of an item waits 30 minutes (a structural change that comes sooner takes it along); a change of only the account that holds the Chief waits 60 minutes and never causes a run by itself. A burst of changes is one run.
- Replayed on your real log, the same changes would have been about 510 runs instead of 1,406 (64% fewer; 394 runs on 21 October would have been 102). The longest wait of a new or closed item was 5 minutes, never longer.
helme mobile-waitingsays what is gathering. It lists the kinds of change waiting and when the next write starts.
Good to know
- A new item still reaches your phone within a few minutes. The words of an item you already know can take up to half an hour to reach it; you read those in the Helme window.
- Nothing is sent when the list is the same as what the page holds, as before. Asking for a write by hand (
helme mobile-waiting sync --force) writes at once and is not held back. - The gathering windows start again whenever Helme starts.
- The figure after the change comes from replaying the log, not from measuring a new day. The first days after installing will show the real number.
Not on the main branch yet.
Viết bởi: helme project agent · tài khoản không rõ · model claude-sonnet-5-5 · phiên 3f5ec1ab-0d3c-4ffb-a24a-ce4176ef94b5