Changelog
What shipped, when, and what it changed for you. Written in plain words, not commit messages.
A finished task leaves a mark nothing can erase
The work record stopped depending on anyone remembering to write it down, and the site started answering machines as well as people.
- Creating or completing a task now writes a row to an append-only ledger, from inside the database itself.
- The work log reads that ledger, so a day that has passed stops changing after the fact.
- Time is reported per project, and time nobody can account for is shown as unknown instead of counted as zero.
- Search engines and AI answer engines get real surfaces: sitemap.xml, robots.txt, llms.txt and structured data on the homepage.
- At the end of a session the agent is asked what it learned, and it cannot file a proposal without your explicit yes.
Who did what, and how much
The question a person leading a project asks every week finally has an answer that does not require opening a single task.
- A work log per person: time actually worked, derived from recorded activity rather than from what anyone claims.
- It says where the timezone came from, so the hours are not a mystery number.
- The handoff now tells a project when its local MHQ files are older than the server.
- The VS Code extension checks the setup that is actually running, not only the files sitting on disk.
One folder per agent, so two agents stop fighting
Running several agents on one project used to mean they edited the same files at the same time. Now each one gets its own copy.
- A task can get its own working folder and its own branch, both created from a clean starting point.
- Two buttons in the VS Code extension: start an agent worktree, finish an agent. No git commands to type.
- A branch is allowed to land only after the tasks it declares a dependency on have landed.
- A task shows which folder an agent is currently working in, so you can see who has what.
A task another agent already took now says so
Two sessions picking up the same task was silent until it produced conflicting work. It is visible from the first list.
- The task list flags work a live agent has already claimed.
- The setup check runs per user instead of comparing project against project, which was producing false alarms.
Take back a skill, and find things when you search
Two things that were quietly wrong: a way of working could never be retired, and multi-word searches returned nothing.
- A skill that turned out to be wrong can be withdrawn instead of governing the project forever.
- A way of working that failed can be diagnosed and closed out rather than silently repeated.
- Searching for several words now ranks the results instead of demanding that every word appear.
The system notices when it keeps failing
Guidance that produces bad work is worse than no guidance. This release made the failures visible to the system itself.
- When a way of working keeps failing, it gets diagnosed instead of being handed out again unchanged.
- A rewrite counts as a defect only when it really is one, not when you simply changed your mind.
- Reminders arrive on a fixed cadence per kind of work, instead of firing on every task.
Older work exists before these dates. This page starts where the release notes start.