Public alpha / Apache-2.0 / local CLI

Catch false-green AI-agent work before you accept it.

Synrail is a local proof gate for coding agents. It does not write code. It asks whether the agent is allowed to claim the task is done.

Open
Apache-2.0
Local
No hosted lock-in
Alpha
Narrow and honest
CI
Tests + audit path
synrail check

Agent:

done, tests passed

python3 alpha.py check

Synrail: Status: Proof Invalid

Reason: verification command not executed

Next: repair final_result.json

After real verification:

Synrail: Status: Accepted

Only Status: Accepted means the task may be reported as complete.

The skeptical question

Is this just post-review?

Not exactly. A normal post-review asks: is this code good? Synrail asks a narrower question first: is the agent allowed to claim this task is done?

If you personally read every diff, run every check, and track every agent step, Synrail may be unnecessary overhead. In that case, you are already acting as Synrail manually.

Synrail becomes useful when you stop being the runtime supervisor: repeated agent runs, long context, failed repairs, second-operator handoff, or proof-sensitive changes.

Where it fits

Use Synrail when verification debt compounds.

Batch agent work

You ask an agent to update many files or run repeated small tasks, and manual proof tracking starts getting expensive.

Failed repair loops

The agent says it fixed the issue, another agent says it is still broken, and you need a bounded next step instead of chat archaeology.

Handoff-safe proof

A second operator or agent needs to understand what was accepted, what was blocked, why, and what is allowed next.

You may not need Synrail if...

The task is tiny, you inspect every changed line, you run verification yourself, and a false-green costs less than running the gate. Synrail is intentionally not a universal correctness machine.

What Synrail checks

claimed_done != accepted_done

Synrail keeps acceptance separate from execution. The agent can produce code, but the proof gate decides whether the work may be accepted.

Scenario
Agent rules
CI alone
Synrail
Tests claimed but not run
Usually missed
Maybe later
Blocks before accept
Proof does not match changed files
Weak
Partial signal
Requires matched proof
Repair handoff is unclear
Chat context
No answer
Names one bounded step

Run locally

Try the false-green demo in about two minutes.

The public alpha is a local CLI. It is designed to wrap existing coding-agent workflows, not replace your editor, CI, or code review.

git clone https://github.com/USBVadik/synrail cd synrail make install-dev make demo

Fallback without make:

python3 -m pip install -e ".[dev]" -c constraints-dev.txt

./examples/false-green-demo/run_demo.sh

Operational transparency

Built in public by NexusShellAI.

Synrail is the current public product of NexusShellAI: a focused developer tool for AI-agent reliability, proof, bounded repair, and handoff.

Alpha feedback

Use Synrail on one real agent task and tell us where it hurts.

The next milestone is not more abstraction. It is 3-5 real external testers using Synrail around Claude Code, Cursor, Codex, Gemini CLI, Aider, or custom agent loops.