Frankenreview v12.1.0

Lean, stateless, zero-cost code reviewer and research agent powered by Gemini via browser automation.

Architecture: Pull Model | Cost: $0.00 | Runtime: Playwright + Chrome

[!CAUTION] Data Privacy Notice: Google AI Studio may use submitted code for model training. Do NOT submit proprietary, confidential, or sensitive code.

Overview

Frankenreview packs a repository into structured XML context and submits it to Gemini via browser automation (no API key required).

Current browser flow:

• Starts or attaches to a Chrome debug session.
• Opens AI Studio on the requested model and verifies the active UI model.
• Enables temporary chat for one-shot reviews when available, avoiding delete cleanup.
• Sets AI Studio Thinking Level to High by default when the model exposes that control.
• Uploads repository context as a file-backed prompt, then triggers Run via native DOM click with mouse fallback.
• Extracts responses with AI Studio's native Copy as markdown action first, then falls back to DOM scraping.

Installation

pip install frankenreview && playwright install chromium

Run Frankenreview from the project directory or pass --path explicitly. Filesystem roots such as /, C:\, and UNC share roots are rejected before any workspace files are created or scanned.

Usage

frankenreview --start-chrome              # Start Chrome debugger
frankenreview -r                          # Review current directory
frankenreview -r --path "C:\path\to\repo" # Windows: review an explicit project
frankenreview -r -m gemini-3-pro         # Use specific model
frankenreview -r -m lite                 # Alias: gemini-flash-lite-latest
frankenreview -r -m gemma                # Alias: gemma-4-31b-it
frankenreview -r --thinking-level high    # Set AI Studio thinking effort if available
frankenreview -r --json                   # JSON output for CI
frankenreview --research --prompt q.md    # Deep research
frankenreview -t                          # Token eaters + token tree

# Model Discovery
frankenreview --available-models           # Cached model list (instant)
frankenreview --fetch-models               # Live fetch from Google AI docs

# Security Gates
frankenreview --secrets                    # Count exposed secrets in repo
frankenreview --secrets-gate               # Toggle git-push secret scan
frankenreview --key-gate                   # Toggle shell key-exfiltration guard

# Session Management
frankenreview --open-chat                 # Open chat for manual use
frankenreview -r --continue               # Continue previous session
frankenreview -r --continue --prompt "Summarize next steps"
frankenreview --delete-chat               # Delete saved session

CLI Reference

Run fr --help for authoritative argument details and aliases.

Security Gates

Frankenreview masks secrets before XML dump writes when mask_secrets: true.

Manual checks:

frankenreview --secrets

Optional shell gates:

frankenreview --secrets-gate   # Wraps git push with a pre-upload secret scan
frankenreview --key-gate       # Guards common shell output paths for key-like data

Both gate toggles edit the current user's shell config and are reversible by running the same command again.

Configuration

Edit config.yaml:

chat_url: "https://aistudio.google.com/prompts/new_chat?model=gemini-flash-latest"
delete_after_review: true
use_temporary_chat: true
prune_dirs: [.git, node_modules, __pycache__]
model_failover: [gemini-3-flash-preview, gemini-2.5-pro, gemini-2.0-flash]
thinking_mode: true
thinking_level: high
mask_secrets: true

thinking_level is best-effort. If a model does not expose the AI Studio control, Frankenreview logs the miss and continues.

Model aliases:

Run selector discovery when AI Studio changes:

frankenreview -ds --port 9222                  # Selector + deterministic generation check

Selector discovery exercises the temporary-chat toggle at the start and ends in a saved chat so the production delete flow can be checked. Temporary chats do not expose Delete Prompt. Discovery submits exactly one probe and does not rerun failed turns or switch models; normal reviews may still use use_temporary_chat: true for one-shot cleanup.

Documentation

• CONTEXT_v11.md - Archived v11 architecture reference
• FR_FEATURE_CATALOG_v11.md - Archived v11 feature reference
• MODEL_COMPATIBILITY_2026.md - Models
• TESTING_PROCEDURES.md - Release verification matrix
• SELECTOR_DISCOVERY_GUIDE.md - AI Studio selector maintenance
• COMMAND_OUTPUT_REFERENCE.md - Generated non-review CLI output

License

MIT - See LICENSE