Skip to content

privacy-threat-model.md

Latest commit

 

History

History
66 lines (50 loc) · 2.85 KB

privacy-threat-model.md

File metadata and controls

66 lines (50 loc) · 2.85 KB
Raw

Privacy And Threat Model

Code Mower's job is to route code review work through multiple agents while making each trust boundary explicit. The safe default is local setup and inspectable artifacts; hosted review and cloud reporting are opt-in.

Assets

Protect these by default:

  • repository source code and diffs;
  • pull request titles, bodies, comments, labels, and commit metadata;
  • local checkout paths and machine/user names;
  • GitHub tokens, provider API keys, deploy keys, and CLI session state;
  • reviewer raw stdout/stderr, prompts, and context-pack material;
  • benchmark results before the user chooses to share them.

Trust Boundaries

Boundary What can cross it Default posture
Local CLI provider Prompt, diff, selected files, context packs Explicit lane config and doctor visibility
SaaS reviewer GitHub App Pull request diff and repository context Manual or informational until calibrated
Local model endpoint Prompt and selected code context User-controlled endpoint; informational by default
GitHub Actions Generated workflows, labels, comments, artifacts Least privilege, guarded triggers, no surprise cron
Cloud benchmark export Sanitized report bundle Opt-in only; inspect bundle before upload

Data Minimization Rules

  • Send diffs plus bounded context packs, not entire repositories by default.
  • Keep raw reviewer outputs local unless the user intentionally commits or uploads them.
  • Redact auth probe output content. Store shape diagnostics such as return code and line count instead of account text.
  • Prefer secret file references for local keys and avoid printing secret values.
  • Keep provider spend and hosted reviewer triggers explicit.
  • Treat generated calibration manifests as shareable only after a privacy scan.

Public Repository Hygiene

The public OSS repo should not contain private reference-repo names, personal paths, raw provider outputs, private account identifiers, or live calibration artifacts from proprietary products. Use anonymized summaries in docs and generic examples such as owner/repo.

Cloud Benchmarking

The hosted benchmarking service is a commercial surface. The OSS core should produce local reports and an inspectable export bundle. Upload should require an explicit command, clear destination, and a chance to review the bundle contents.

Future cloud uploads should support:

  • anonymous or organization-scoped submission modes;
  • source-free metric summaries;
  • optional redacted finding text;
  • explicit retention policy;
  • user-controlled deletion/export; and
  • clear separation between public aggregate benchmarks and private product reports.

Release Gate

Before release, run the privacy scan and inspect any changed calibration artifacts. A release should fail if it contains personal paths, private repo slugs, raw auth output, or likely secrets.