Check the work before your AI agent calls it done.
+Reflex Control checks protected-file writes and selected risky commands, then runs your configured tests when a coding agent attempts to finish.
+ +Start with the documented Claude Code integration.
01 / agent requests a write
target: .env
→ write denied · protected path
02 / agent attempts to finish
run: configured project tests
→ a test fails · return failure to agent
agent revises the change
run: configured project tests
→ tests pass · turn may finish
Three checkpoints.
One agent turn.
Hooks add checks around an agent's actions. Your configured rules determine what is protected and which tests count as verification.
Set boundaries for file changes.
Check supported file writes against configured protected paths. The default policy includes environment files, private keys, and secrets directories.
Hook behavior varies by agent. These checks are not a sandbox. Read the adapter coverage and limitations ↗
+A released prototype.
A trace you can inspect.
A documented Claude Code session includes blocked writes to protected paths and a failing test followed by a correction and a passing test. The record describes the setup, observations, and untested paths.
This is early research evidence, not production validation. Other adapters have different levels of verification.
-
+
- Read the live Claude Code test ↗October 2, 2026 · Setup and observed results +
- Inspect the v0.5.0 release ↗macOS, Linux, and Windows packages +
- View the CI history ↗Public build and check results +
- Check agent adapter coverage ↗Supported hooks, constraints, and validation status +
Give test failures
more useful context.
The next proposed feature is optional Claude API analysis of user-selected code diffs and test-failure evidence: explain the issue and suggest concrete next steps.
The existing Claude Code hooks and this proposed API feature are separate. Content selection and sensitive-data handling need implementation and evaluation before user trials.
- 01Build the optional Claude API integration
- 02Evaluate explanations against known failures
- 03Track request cost and cap retries
- 04Try the workflow with initial volunteers
Built by Murat.
I build Rust-based tools for AI-assisted software development, with a focus on verification, bounded execution, and inspectable results. Reflex Control sits alongside my agent runtime and repository intelligence experiments.