cland is built by a cland crew: a Coordinator, Implementers, Reviewers and QA on the same hub you would run. After each release we write down what it took, what it cost, what broke and what we changed because of it.
v0.3.1 cuts the cland hub to about a fifth of its memory on our own project and makes sure a message to the Coordinator never silently goes undelivered. 25.5 hours from a voice note to the release, steered from a phone.
v0.3.0 lets you follow and steer an AI coding crew from Telegram or Discord: General both ways, asks and approvals as cards, voice notes as answers. Seven epics, every demo clear on its first run, and a live run on real bots that found the bugs the fakes could not.
v0.2 put the whole workbench in the browser and on the phone: 12 epics, every demo clear on its first verdict. v0.2.1 then ran a free model in every crew role on real tasks. It landed its work with no model-caused rescue. What failed was our ability to prove it.
A merge queue and test tiers cut the time tasks spent waiting for tests from 21% of test time to under 10%, and the crew landed 315 tasks in 91 hours. The night before the release still took five runs and eight hours, and not one of the red runs was a product bug.
We read all 237 test files and 557 run receipts to find out why verification set the crew's pace. The suite is not too big. It runs at the wrong level, and the fix is test tiers enforced by the hub plus a merge queue.
The crew merged 2.9 tasks an hour instead of 1.2, and more of them cleared review on the first pass. The test suite did not keep up: the queue shrank, but each run got longer, droplets cost $18.11, and the release took nine runs.
In 77 hours a cland crew landed 94 tasks across 10 epics for $1.17 of cloud runners. The crew was not the bottleneck: the verification queue was, and not one red integrated run found a product bug.