Preconfiguration Ships Its First Release With Preconfig Doctor for Failed Setups

A coding agent in the cloud usually begins each task on a machine built from scratch, and a wrong setup tends to surface only when the agent is already halfway into its work. Preconfiguration deals with that before the agent arrives. You write the machine down once, in preconfig.yaml, and the preconfig command turns it into a setup file for each platform: a dev container definition, the setup steps GitHub Copilot runs, Cursor’s environment file, a cloud-init file and a plain shell script. Version 0.1.0, with Preconfig Doctor built in, has been on the releases page for Linux, macOS and Windows since 8 October.

Four Commands Before the Agent Starts

detect reads files a repository already keeps, such as .python-version, requirements.txt and a compose file, and writes a first draft of the spec. On the Alpha’s sample service, the draft came out identical to the hand-written spec, comments aside. build writes every platform’s file from the spec. check reads setup files that people wrote themselves. On the same service it found 15 errors and 2 warnings: 11 of the errors and both warnings in three hand-written files, and 4 errors for files that were missing. verify runs the setup and then the project’s own tests on a clean machine. In three runs it reached READY in 49.6 to 73.0 seconds, and a run with no Redis in the spec got through the setup but failed its ready check, with the tests showing why.

Doctor Reads the Failures

preconfig doctor takes the log of a failed setup, finds the step that broke and names the cause from 31 fixed rules, with no AI model involved. If the spec is at fault, --fix edits preconfig.yaml itself and every platform’s file is rebuilt. The rules were first fitted to 28 recorded cases. Then they were frozen, and 11 more cases were run against them: 8 came back right, 2 came back “unknown”, and 1 was a miss, a network failure where Doctor didn’t name the host. Over all 74 answers it hasn’t once made a wrong change to a spec, and each fix it wrote cleared its failure when the setup ran again on a clean machine.

On the Platforms Themselves

The Beta, planned at about 12 weeks, takes the generated files onto the platforms themselves, Copilot, Cursor and Codespaces among them, with real repositories and real failed logs for Doctor to read. Preconfiguration and VPN Works come at AI agents from two sides: one gets the agent’s machine ready, the other decides where the agent’s traffic may go.