vardoger is designed around a non-negotiable constraint: all conversation
history reading and analysis happens on the user's machine. The project does
not open sockets, contact external services, or upload any conversation data.
Analysis is performed by the host AI assistant's own model through a
skill-driven pipeline (see PRD.md, Section 6). The only network traffic in
the normal workflow is pipx install vardoger / uvx vardoger pulling the
package from PyPI.
See PRIVACY.md for the full data-handling policy (what vardoger reads, writes, and never transmits).
If you discover a code path that violates this constraint (e.g., a dependency that phones home, or a bug that exfiltrates local data), please treat it as a security issue and report it privately.
Only the most recent published release on PyPI is supported with security fixes.
| Version | Supported |
|---|---|
| 0.1.x | Yes |
| 0.1.0bN betas | No |
Please do not open a public GitHub issue for security-sensitive reports.
- Email the maintainer at
dstrupl@gmail.comwith:- A description of the issue and its impact.
- Minimal reproduction steps (include vardoger version, Python version, and host platform — Cursor / Claude Code / Codex / OpenClaw).
- Any logs, payloads, or screenshots relevant to the report. Redact conversation content; vardoger does not need it to reproduce most bugs.
- Expect an acknowledgement within 3 business days.
- I will work with you on a fix and a coordinated disclosure timeline; most issues are resolved within 30 days.
In scope:
- Data-exfiltration bugs in vardoger's core or any shipped plugin.
- Path-traversal, TOCTOU, or arbitrary-write issues in the history adapters
(
src/vardoger/history/) and writers (src/vardoger/writers/). - MCP server issues (
src/vardoger/mcp_server.py) such as unsafe tool arguments, sandbox escapes, or prompt-injection paths that influence on-disk state. - Dependency CVEs that affect vardoger's runtime behavior.
Out of scope:
- Social-engineering against PyPI or GitHub maintainers.
- Bugs that require the user to already have arbitrary code execution on their own machine.
- Issues in the upstream AI assistants (report those to Cursor, Anthropic, OpenAI, or OpenClaw directly).
Reporters who follow this policy will be credited in the CHANGELOG unless they request otherwise.