Documentation
Security model
The CLI is inert on installation: package.json defines no preinstall, install, or postinstall script. Explicit commands are required to create files.
Project setup skips existing files by default. --force is required to replace files owned by a command, and --dry-run previews writes. Root AGENTS.md is a protected exception: setup creates it only when absent and never replaces an existing project policy, including under --force. The package does not install Git hooks automatically, execute network scans, read credentials, or contact a telemetry service.
Forced scaffold replacement is staged in the destination directory before the owned file is swapped into place. npm packaging uses an explicit public-document allowlist plus a packed-manifest check; local documentation, generated findings, and audit evidence are rejected from the archive even when they exist in the working tree.
Audit reports begin as Unverified. The tool scaffolds the review; it does not claim that generated text proves a control. Agents and reviewers must attach repository paths, line numbers, reproducible command output, or authorized external evidence.
The start and next workflow commands are read-only. next derives progress from project files and does not create a hidden state file, contact a service, rerun an audit, or authorize active testing. It advances past documentation verification only when the latest recognized report is structurally valid, covers the currently configured required documents, and explicitly passes; malformed, unrelated, ambiguous, or failing evidence is rejected. A more recent failure takes precedence over an older pass.
The suggest reviews command is also read-only. It invokes Git without a shell to obtain selected changed paths, validates an explicit base reference, and prints heuristic review candidates with their matching paths. It creates no worksheet and cannot establish that unmatched changes are safe. Automatic commit or push reminders require a separately configured hook or CI policy.
Security and engineering worksheet commands require .agent-docs/config.json and the configured documentation root before writing. This initialization gate prevents review artifacts from being created outside the project record, but it does not require a passing documentation audit or imply documentation completeness.
Security and engineering worksheets are also validated before the workflow treats them as complete. Changing an Unverified label alone is insufficient: evidence-supported outcomes require a stable control or practice ID plus the fields required by the status, and N/A requires an explicit evidence rationale. Workflow completion means the record is structurally complete, not that every control passed or recommendation was applied.
AI guidance requires repository inspection, explicit unknowns, and human authorization before interacting with a live service, cloud account, network, credential, or production environment.
Publishing uses npm provenance through GitHub trusted publishing when configured. Protect the GitHub release environment, require reviews as appropriate, and never store a long-lived npm token when OIDC is available.