AI coding agents can claim that work is done before the proof is trustworthy. Synrail adds a proof-first control layer that separates agent claims from accepted results.
AI coding agents are powerful, but they often produce plausible completion narratives before the work is actually verified. They may claim tests passed, show a plausible diff, or leave handoff state unclear while the operator still lacks trustworthy proof.
Claimed tests that were not actually run or were bypassed by the agent's internal narrative.
Proof artifacts that do not match the changed files or reflect an older state of the codebase.
Persuasive explanations of completion instead of concrete, reproducible runtime evidence.
Failed repair cycles that leave the environment in an inconsistent state without a clear next step.
Synrail keeps the boundary between execution and acceptance explicit. An agent saying “done” is not enough. Synrail checks whether proof is real, matched to the task, and verified. If proof is weak, mismatched, stale, or unverified, Synrail blocks acceptance and gives one bounded repair step.
Loading proof from ./.synrail/proof.json...
Comparing diff against claimed changes...
⚠ Proof Mismatch: claimed "Fix bug in parser" but diff shows "Update README"
Status: Rejected (False-Green Detected)
Recommended repair: "Ensure the actual code changes match the reported intent before claiming success."
Initialize a task scope that the agent must work within.
The agent generates machine-readable proof: test results, trace logs, and verified diffs.
Validate the artifacts against the actual state of the workspace.
If non-green, repair only the named gap and check again until success.
Quick Install
git clone https://github.com/USBVadik/synrail
cd synrail
make install-dev
make demoStandard Workflow
synrail start "Describe the bounded local change"
# agent works...
synrail checkDevelopers using Claude Code, Cursor, Aider, Gemini CLI, or custom agentic loops who want to automate the verification of "done" claims.
Teams running repeated small AI-assisted code changes where manual verification of every diff is becoming a bottleneck.
Operators who still manually verify agent work but want a tool to catch the "easy" hallucinations before they look at the code.
Developers inheriting a failed repair from an AI agent who need an explicit handoff state to continue safely.
Blocks claimed-done closure until proof reaches accepted status.
Keeps proof on explicit artifacts instead of narrative trust.
Names the exact proof repair when acceptance is blocked.
Preserves trusted local recovery points when they exist.
Supports bounded continuation and second-operator handoff.
Synrail is currently a narrow local alpha product. It is focused on one local trusted worktree, one bounded local agent task, explicit proof artifacts, and synrail check as the acceptance gate. It is not yet a broad hosted orchestration platform, a generic CI/CD replacement, or a universal correctness system.
NexusshellAI is building developer infrastructure for safer AI-assisted software work. As coding agents become more capable, teams need better control over when generated work is allowed to be trusted. Synrail is our first product in that direction: a focused reliability layer for proof, acceptance, bounded repair, and handoff.
Synrail is model-agnostic and can be used around different coding-agent workflows. As the product evolves, NexusshellAI plans to evaluate larger model-inference and agent-reliability workloads, including coding-model evaluation, proof-quality checks, false-green detection benchmarks, and multi-agent repair/handoff scenarios.