Security and data custody #

Docwriter handles source repositories, generated datasets, model-provider credentials, and agent trajectories. Treat generated artifacts as private until they pass an explicit publication review.

Reporting a vulnerability #

Do not place credentials, private source, or exploit details in a public issue. Use the contact route listed at https://cameron.stream and include the affected command, revision, and a minimal reproduction that contains no live secret.

Repository guarantees #

• API credentials are read from the process environment only for explicitly executed provider operations. They must never enter committed files, task workspaces, fixtures, logs, or command-line arguments.
• artifacts/, source caches, datasets, task workspaces, sessions, and local environments are ignored by Git.
• Paid generation and training require an execution flag and explicit cost ceiling. Dry-run planning does not call a provider.
• Repository-agent tasks run without Bash and require a filesystem namespace that hides Git history, host home, source caches, the hidden target, the oracle, and unrelated credentials.
• A generated page or trajectory is not trusted because a model completed it. Deterministic source, mutation, tool, termination, and license checks run before model judgment.

Publication review #

Before pushing a release or generated artifact:

1. enumerate the exact Git publication surface;
2. scan it for secrets, absolute local paths, private repositories, sessions, artifacts, and unsupported source text;
3. verify source and curriculum license manifests;
4. run tests, static checks, and dependency audits without ambient credentials;
5. inspect the resulting commit rather than relying on the working tree;
6. verify the remote commit and anonymously readable files after push.