Prepare a repository with AI

Repository preparation is an agent-assisted change workflow. It is designed for applications that need more than runtime detection: code may need to read a new environment variable, connect to a managed database, expose a health endpoint, split a worker from a web process, or add an immutable GitHub Actions build.

The result is a reviewable Change Request, not an automatic merge or production deployment.

The agent follows the same repository contract a developer can implement manually. Start with Set up your repository if you want to make or review each change yourself.

What the agent may propose

The coding worker can propose changes anywhere in the temporary repository, including:

• application and API code;
• database clients and migration files;
• package manifests and lockfiles;
• Dockerfiles and .dockerignore;
• autodisc.yml;
• health checks and start commands;
• GitHub Actions workflows;
• service-to-service and managed-resource bindings; and
• documentation required to operate the changed application.

The worker does not receive your GitHub App key, model-provider credential, runtime-provider token, production database password, or project secrets.

Prepare a change

1. Open or create a project.

2. Select a connected GitHub repository and exact branch or commit.

3. Choose Prepare with Autodisc.

4. Describe the outcome in application terms. For example:

   Deploy the API and worker separately. Add PostgreSQL for the existing Prisma models, run safe migrations during release, and expose only the API.

5. Choose whether Autodisc should prepare a draft pull request and GitHub Actions build bridge.

6. Start preparation.

The request is durable. Closing the browser does not cancel it.

Review the Change Request

A completed preparation shows:

• the exact base source revision;
• every changed file and the full stored diff;
• deterministic check results;
• an Application Plan for services, resources, bindings, and networking;
• an immutable build requirement when applicable;
• required secrets that only you can supply;
• a monthly resource quote; and
• an explanation of why each change is needed.

Use Request changes to give another concrete instruction. A new revision is generated from the approved base plus your instruction, and previous approvals are cleared.

Three separate decisions

Autodisc separates authority deliberately:

1. Approve code accepts one exact patch and Application Plan digest.
2. Approve resources accepts one exact quote and any sensitive acknowledgements, such as creating a persistent database.
3. Apply approved plan creates or updates canonical resources and starts the deployment workflow.

Approving code cannot create a database. Merging an unrecognized GitHub commit cannot substitute a different patch for the approved revision.

GitHub pull requests

When the GitHub App has write access, Autodisc can publish the reviewed revision as a branch and draft pull request. The backend performs the Git operations; the disposable coding worker never receives GitHub credentials.

If no GitHub write connection exists, the same revision remains reviewable in Autodisc and can be downloaded or handled through the CLI.

If preparation fails

Open the preparation event timeline and identify the failed phase:

• Source unavailable: refresh the repository connection and verify branch access.
• Unsupported source: use GitHub or an authenticated source upload.
• Agent inference unavailable: retry after the managed model gateway recovers.
• Check failed: inspect the check output, then request a revision with the exact failure.
• Policy blocked: remove privileged container, host-network, secret, or destructive resource behavior.

The failed preparation does not mutate the repository or runtime.

CLI

Use the Change Request commands when the same workflow must run in a terminal:

autodisc change-request preparation inspect PREPARATION_ID --json
autodisc change-request review CHANGE_REQUEST_ID
autodisc change-request approve-code CHANGE_REQUEST_ID \
  --revision REVISION_ID \
  --patch-digest sha256:PATCH \
  --plan-digest sha256:PLAN

The CLI and dashboard use the same canonical objects and approval rules.