vardoger

vardoger

A cross-platform plugin for AI coding assistants (Cursor, Claude Code, OpenAI Codex, OpenClaw, GitHub Copilot CLI, Devin Local, legacy Windsurf, and Cline) that reads supported local conversation-history formats, extracts behavioral patterns, and generates personalized system prompt additions — making the assistant progressively better suited to how you work. Current format limitations are called out per platform below rather than silently reading undocumented stores.

History discovery, parsing, checkpoints, and generated rules stay on your machine. Analysis is performed by the host assistant you invoke, so selected conversation excerpts follow that assistant and model provider's data policy. Vardoger operates no backend and sends no telemetry.

Prerequisites

Python 3.11+

Platform	Command
macOS	brew install python@3.13 (install Homebrew) or python.org/downloads/macos
Debian / Ubuntu	sudo apt install python3
Fedora	sudo dnf install python3
Windows	winget install Python.Python.3.13 or python.org/downloads/windows

pipx

Recommended for installing vardoger as an isolated CLI tool. Full instructions at pipx.pypa.io/stable/installation.

Platform	Command
macOS	brew install pipx && pipx ensurepath
Debian / Ubuntu	sudo apt install pipx && pipx ensurepath
Fedora	sudo dnf install pipx && pipx ensurepath
Windows	scoop install pipx or pip install --user pipx && pipx ensurepath

Quick Start

pipx install vardoger
vardoger setup cursor        # or claude-code, codex, openclaw, copilot, devin, windsurf, cline

Then tell your assistant: "Personalize my assistant."

Looking for the in-app plugin listings? Track review status for each marketplace (PyPI, Cursor, Claude Code, Codex, Copilot CLI, Devin Desktop, Cline, ClawHub) in MARKETPLACE_STATUS.md. Vardoger is currently live on PyPI, the Claude Code community catalog, self-hosted Codex and Copilot marketplaces, McpMux, and the Official MCP Registry. The former Cursor listing is currently unavailable. ClawHub has regressed to 0.3.1 and is intentionally frozen because its mandatory MIT-0 terms conflict with this Apache-2.0 project. Current OpenClaw history has a live-accepted, opt-in full-read Gateway CLI path through the direct install.

Previous pre-releases. pipx install vardoger now resolves to the stable 0.4.0 release. The beta install paths below stay here for anyone still pinning an earlier release; new installs should not need them.

# opt into future pre-releases (0.2.0bN, etc.):
pipx install --pip-args="--pre" vardoger
# or pin an older pre-release:
pipx install vardoger==0.1.0b3
# or run without installing:
uvx vardoger --help

CLI Commands

Command	Purpose
vardoger setup <platform>	Register vardoger with a platform (cursor, claude-code, codex, openclaw, copilot, devin, windsurf, cline).
vardoger status [--platform X] [--json]	Report whether each personalization is fresh or stale.
vardoger prepare --platform X [--batch N] [--synthesize]	Produce the batched prompts used by the AI-driven skill pipeline.
vardoger write --platform X	Read synthesized personalization from stdin and write it to the platform's rules file (supports YAML-frontmatter confidence metadata).
vardoger feedback accept|reject --platform X [--reason TEXT]	Record whether you kept or rejected the last generation. reject auto-reverts to the prior generation.
vardoger compare --platform X | --all [--window DAYS] [--json]	Compare heuristic conversation-quality metrics before vs. after the latest personalization.
vardoger profile sources [--platform X] [--json]	List saved generations and their stable one-based selection IDs.
vardoger profile preview --source X:N [--source Y:latest]	Compile explicitly selected generations into a normalized cross-host profile and show the portable AGENTS.md diff without writing.
vardoger profile write --source X:N --target ./AGENTS.md [--apply]	Show the same diff; update only Vardoger's fenced block when the user explicitly adds --apply.

How It Works

1. Read — Parses supported conversation files already stored on disk, including opted-in Devin ATIF exports
2. Analyze — The host AI model identifies patterns in your communication style, tech stack, workflow, and preferences
3. Generate — Produces a system prompt addition tailored to you
4. Deliver — Writes the addition to the platform's native config (.cursor/rules/, .claude/rules/, AGENTS.md, etc.)

First run vs. incremental runs. By default vardoger does not apply a time window — the first run reads your full local history (that is when the signal is richest and a windowed default would silently drop older sessions you never get a second chance to feed in). After that, a per-conversation checkpoint store at ~/.vardoger/state.json ensures every subsequent run only reprocesses new or changed conversations, so refreshes stay fast. If you have very large local history and want to cap the first-run cost, pass --since DAYS (e.g. vardoger prepare --platform cursor --since 90); use --full to force a full re-crawl that bypasses the checkpoint.

Cross-host profile compiler

The profile compiler combines only existing Vardoger generations that you select; it does not silently reopen raw transcripts. Generation indexes are one-based, while latest resolves to the newest saved generation for that platform:

# Discover the reviewed generations available for selection.
vardoger profile sources

# Read-only: render the profile, audit summary, and exact AGENTS.md diff.
vardoger profile preview \
  --source claude-code:latest \
  --source codex:2 \
  --target ./AGENTS.md

# Optional controls are applied before rendering and included in the audit.
vardoger profile preview \
  --source codex:latest \
  --min-confidence medium \
  --max-age 180 \
  --redact '@example\.com' \
  --target ./AGENTS.md

# Still preview-only without --apply. This is the explicit write step.
vardoger profile write \
  --source claude-code:latest \
  --source codex:2 \
  --target ./AGENTS.md \
  --apply

The normalized JSON audit is available with profile preview --json. It records source platform, generation, output hash/path, observation time, recency, confidence, conflicts, supersession, and exclusions. Opposing rules such as “Prefer tabs” and “Avoid tabs” are withheld from active instructions until reviewed. Writes preserve everything outside <!-- vardoger-profile:start --> and <!-- vardoger-profile:end -->.

Supported Platforms

Platform	History Source	Prompt Delivery	Integration
Cursor	Agent transcript JSONL	.cursor/rules/vardoger.mdc	Marketplace plugin + MCP
Claude Code	Session JSONL	.claude/rules/vardoger.md	Community/custom plugin
OpenAI Codex	Session rollout JSONL	~/.codex/AGENTS.md	Repository marketplace plugin
OpenClaw	Legacy JSONL or opt-in full read through official Gateway CLI; private SQLite is never queried	~/.openclaw/skills/vardoger-personalization/SKILL.md	Analyzer + Gateway compatibility status
GitHub Copilot CLI	~/.copilot/session-state/<session-id>/events.jsonl plus legacy flat JSONL	~/.copilot/copilot-instructions.md (global) or <project>/.github/copilot-instructions.md (project) — managed inside a <!-- vardoger:start --> fenced section	Custom plugin + published skill
Devin Local	User-enabled Devin CLI --export ATIF JSON under ~/.vardoger/imports/devin/	~/.config/devin/AGENTS.md (global, fenced section) or <project>/.devin/rules/vardoger.md	ATIF + skill + rules integration
Legacy Windsurf / Cascade	~/.codeium/windsurf/**/*.jsonl	~/.codeium/windsurf/memories/global_rules.md (global, fenced section) or <project>/.windsurf/rules/vardoger.md (project, dedicated file)	Legacy Cascade skill + CLI + MCP
Cline	VS Code globalStorage/.../tasks/*/api_conversation_history.json	~/Documents/Cline/Rules/vardoger.md (global) or the existing .clinerules project layouts	CLI + MCP

Development

Requires uv (Python package manager):

git clone https://github.com/dstrupl/vardoger.git
cd vardoger
uv sync
.venv/bin/vardoger --help

Project Layout

src/vardoger/          # shared core — history reading, analysis, prompt generation
.agents/plugins/       # repository-level Codex marketplace catalog
plugins/_shared/       # shared analysis/personalization skill authored once
plugins/cursor/        # Cursor MCP server config, install script
plugins/claude-code/   # Claude Code plugin manifest, skills
plugins/codex/         # Codex plugin manifest, skills
plugins/openclaw/      # OpenClaw skill
plugins/copilot/       # GitHub Copilot CLI plugin manifest, skills
plugins/devin/         # Devin Local ATIF export, skill, and rules integration
plugins/windsurf/      # Legacy Cascade install snippet and rules delivery
plugins/cline/         # Cline integration and marketplace install guidance
tests/                 # all tests, mirroring src/ structure

• Platform-agnostic logic lives under src/vardoger/.
• Platform-specific integration (manifests, skills, install scripts) lives under plugins/<platform>/.
• Tests live in tests/, mirroring the source tree.

See AGENTS.md for full coding standards and quality checks.

Quality gates

CI enforces a combined quality bar on every push and pull request:

• ruff check / ruff format --check — lint (incl. complexity, pylint, return, pathlib, tryceratops rules) and formatting.
• mypy src/ — strict type checking.
• pytest --cov=vardoger --cov-fail-under=80 — tests across Python 3.11–3.13 with a combined 80% coverage floor.
• A parallel security job runs bandit -r src/ and pip-audit --skip-editable to catch common code smells and dependency CVEs.

Run the full bundle locally before pushing:

uv run ruff check . && uv run ruff format --check . && uv run mypy src/ && uv run pytest --cov=vardoger --cov-fail-under=80

Contributing

Contributions are welcome. Short version:

1. Fork dstrupl/vardoger on GitHub and clone your fork.
2. uv sync and create a topic branch.
3. Make your changes with tests and run the quality-gate one-liner above.
4. Push to your fork and open a PR against main.

CI (test on Python 3.11/3.12/3.13 plus a security job) will run automatically on the PR. First-time contributors may need a maintainer to click Approve and run before the first workflow execution.

See CONTRIBUTING.md for the full walkthrough and AGENTS.md for coding standards and commit-message conventions.

Releasing to PyPI

CI runs automatically on every push and PR (lint, type check, tests across Python 3.11–3.13). To publish a new version:

1. Bump the package, generated setup manifests, platform manifests, and registry metadata together; close the matching section in CHANGELOG.md.
2. Run the complete quality, security, plugin-validation, and build checks.
3. Commit and push to main, then wait for all required checks.
4. Create an annotated vX.Y.Z tag and publish the matching GitHub release.

The publish.yml workflow builds the package and uploads it to PyPI via trusted publishers (no API tokens needed). Once complete, pipx install vardoger will pull the new version.

Status

Public beta. Version 0.4.0 is published on PyPI and tagged on GitHub; this source tree tracks that released version. See MARKETPLACE_STATUS.md for live listings and the remaining official-directory submissions. See MANUAL_SUBMISSION_RUNBOOK.md for the owner-only submission, review, and publication steps. See PRD.md for the full product requirements document.

Privacy and security

• PRIVACY.md — what vardoger reads, writes, and (importantly) does not send anywhere.
• SECURITY.md — how to report a vulnerability privately.

License

Licensed under the Apache License, Version 2.0. See also the public Privacy Policy and Terms of Use.