Vanta Omens

Tensorlake Orchestrate application that runs daily, reads open Vanta vulnerabilities, filters them to source-control repository assets, groups them by repository, and dispatches one sandboxed remediation function per repository.

Repository remediation is fanned out with Tensorlake async functions so each repository runs concurrently in its own sandbox.

Prerequisites

• uv
• Tensorlake CLI access through the project environment
• Tensorlake, Vanta, OpenAI, and GitHub credentials

Configuration

Minimum required runtime secrets:

Name	Used by	Purpose	Where to get it
VANTA_CLIENT_ID	Monitor function	Manage Vanta OAuth client ID	Create a Manage Vanta application in the Vanta Developer Console
VANTA_CLIENT_SECRET	Monitor function	Manage Vanta OAuth client secret	Generated with the same Manage Vanta application
GITHUB_TOKEN	Remediation function	Authenticated clones, branch pushes, and pull request creation	Create a GitHub fine-grained personal access token from GitHub Developer settings
OPENAI_API_KEY	Remediation function	API key used by the Codex remediation agent	Create an API key from the OpenAI API keys page

The Vanta application should use the Manage Vanta API client_credentials flow with a read scope such as vanta-api.all:read. Vanta documents this flow as exchanging the application client_id and client_secret at /oauth/token for a bearer token.

For GITHUB_TOKEN, prefer a fine-grained personal access token limited to the repositories this app can remediate. It needs enough repository access to clone, push branches, and create pull requests. For GitHub repositories, that normally means Contents: Read and write and Pull requests: Read and write.

OPENAI_API_KEY is passed to the Codex CLI as both OPENAI_API_KEY and CODEX_API_KEY inside the remediation sandbox.

Common options:

Name	Default	Purpose
VANTA_BASE_URL	https://api.vanta.com	Use https://api.vanta-gov.com for GovCloud
VANTA_SCOPE	vanta-api.all:read	Manage Vanta OAuth scope
VANTA_PAGE_SIZE	100	Vanta page size
VANTA_ASSET_FETCH_CONCURRENCY	10	Maximum concurrent Vanta asset lookup requests
REPOSITORY_URLS_JSON	{}	JSON map from Vanta asset id/name/target id to clone URL
CODE_AGENT_MODEL	gpt-4.1-mini	Cheap, fast coding model for the agent command
CODE_AGENT_COMMAND	codex exec --model {model} --sandbox workspace-write - < {prompt_file}	Remediation command template
DRY_RUN	false	If true, clone and run the agent but do not push or open PRs

REPOSITORY_URLS_JSON is useful because Vanta source-control assets do not always expose a clone URL. Example:

{
  "backend-api": "https://github.com/acme/backend-api.git",
  "a2f7e1b9d0c3f4e5a6c7b8d9": "https://github.com/acme/frontend.git"
}

Deploy

Run deployment from the repository root and deploy the root app.py. Do not run tl deploy app.py from inside vanta_omens/; Tensorlake uses the deploy file's directory as the code root, and deploying from the package directory omits the vanta_omens package name at runtime.

cd /path/to/vanta-omens
uv sync
uv run tl secrets set VANTA_CLIENT_ID=... VANTA_CLIENT_SECRET=... OPENAI_API_KEY=... GITHUB_TOKEN=...
uv run tl deploy app.py

Create the daily Tensorlake cron schedule after deployment:

uv run tl cron create monitor_vanta_vulnerabilities --schedule "0 9 * * *" --input-json null

Local Run

uv run python -m vanta_omens.app

Test

uv run pytest