This page summarises the user-visible features added in AITM 2.6 and the updates that followed it.
What's New
Pipeline heals itself instead of stopping
The biggest change is philosophy: a task should only end as failed when the configuration is wrong or the AI is missing data. Every other problem is now retried or repaired automatically.
- Merge conflicts self-heal. A task that hits a conflict while merging into
devis retried with rebase and merge strategies, then handed to a fix agent. - Post-merge tests fix forward. If tests fail after a merge, AITM isolates the cause and fixes it on the task branch rather than reverting the merge immediately.
- Type-check auto-fix. A failing type-check goes to a fix agent for up to three rounds before the task fails.
- Infrastructure errors are classified separately from real test failures: a crashed worktree is rebuilt, and dependents are no longer cascade-skipped.
- Watchdog covers every task type, including agent-mode tasks.
- Stale merge locks and stranded merges are reclaimed automatically on the next scheduler tick and after a restart.
- Post-merge test fixes stay within the task's scope. When tests are fixed after a merge, the fix cannot touch files outside the task's scope — the same scope-gate used elsewhere in the pipeline.
Bug tasks must prove the bug
- Reproduction gate. When a task is a bug fix, the architect step must write a Reproduction section that demonstrates the symptom before the fix and its absence after.
- Diagnostic pack. Bug tasks that AITM files automatically carry the crash lines, a blame summary, affected files and the linked known-error id.
- Business-scope gate. Work outside the task's scope is blocked with an AI judge verdict.
Full test suite is opt-in
Running the whole test suite after every merge was the main cause of memory pressure. It is now off by default (Pipeline → Full test enabled) and, when enabled, runs only when the merge queue is empty.
Deliberate restart after a self-merge
When AITM merges a change to its own code, it no longer restarts under you. The default (Git & Storage → Server restart after self-merge = manual) waits for the queue to drain and shows a Restart button in the header menu.
Settings you could not reach before
- Day / night CPU profiles (System tab): reserve more threads for the dashboard during the day and let tests use the machine at night.
- Disk Janitor and Disk Watchdog (Watchdog & Resources): retention rules and absolute GB budgets per location.
- Swap management (System): recommended swap size computed from RAM and free disk.
- Dedicated Project Check tab, and a Test connection button for the translation model.
- A new project gets a default set of skills right at creation (documentation, UI patterns), instead of waiting for its first task.
Dashboard and panels
- Dockable panels (Chat, Brainstorm, Contract Explorer, task consoles) snap into real layout slots.
- Monitor bar gained a Queued tab, a CPU sparkline, and an Estimation tab with queue ETA.
- English is the default UI language.
- Task cards: dependency numbers are clickable, and the recovery-agent button asks for confirmation.
Licensing
A false-positive licence block can be unlocked from the UI. Project slots used / max are shown in Settings → License. When creating or importing a project in the Wizard, if the licence has a free slot, the wizard offers to assign it right away.
Resources
/tmp sweeper recognises leaked Chrome and Playwright directories and cleans them after two hours by default.
Tips
- The pipeline now retries and self-heals most failures instead of stopping — give the watchdog a chance before intervening manually.
- Filing a bug-fix task? Expect the architect step to write a Reproduction section first.
- The full test suite is now opt-in — enable it in Pipeline settings if you want it to run after every merge.