Managed AI
How Autodisc uses governed models for repository changes, deployment diagnosis, and repair proposals.
Autodisc uses managed AI for bounded platform tasks. The main workflow is repository preparation: a disposable coding worker inspects an immutable source revision, asks the managed model to propose repository changes and an Application Plan, runs checks, and returns a Change Request for review.
AI output is never treated as deployment authority.
Managed AI is optional. A developer can implement the same portable repository contract by following Set up your repository.
Supported workflows
Repository preparation
The agent may propose application code, API changes, database migrations, runtime configuration, managed-resource bindings, Dockerfiles, and CI workflows. The trusted backend reapplies and validates the result against the immutable source before it becomes reviewable.
Revision requests
A reviewer can ask for a new revision with a concrete instruction. The new source patch and Application Plan receive new digests, and earlier approvals no longer apply.
Deployment diagnosis and repair
Where enabled, Autodisc can analyze bounded, redacted build or runtime evidence and explain a failure. A repair is another Change Request; it is not edited directly into the running production container.
Credential boundary
The disposable worker receives one expiring, job-scoped capability. It does not receive:
- the model-provider token;
- the GitHub App private key or installation token;
- Dokploy or runtime-provider credentials;
- production database credentials;
- project environment secrets; or
- a general Autodisc user token.
The worker can call only the internal endpoints allowed for its exact job. The backend performs model calls, GitHub publication, resource provisioning, and deployment.
Repository data
Autodisc builds a bounded repository context from the immutable source snapshot. Secret files and common credential patterns are omitted before model inference. The model result is schema-validated, stored as a content-addressed artifact, and tied to the job's source digest.
Do not include secret values in preparation instructions. State the variable name and purpose instead.
Approvals
Generated work uses separate controls:
- Code approval names an exact source patch and plan digest.
- Resource approval names an exact plan, quote, and sensitive-change acknowledgements.
- Apply is a separate, idempotent operation over approved immutable artifacts.
The model cannot merge a pull request, approve a database charge, or deploy to production on the user's behalf.
Usage and availability
Managed AI uses the active organization's inference allowance. Autodisc reserves a bounded amount before calling the provider, records actual usage, and releases the reservation when the request fails.
Provider configuration, internal model routing, and credentials remain private. The dashboard reports capability availability and customer-visible usage rather than exposing a provider setup form.