Audit the system, not the files.
Most coding agents inspect files in isolation. System Coherence reconstructs how behavior should work, then checks whether the layers still agree.
Find the bugs between the layers.
Install the complete suite · See the methodology · Explore the docs
A file can be correct and a test can be green while the UI, state, domain, persistence, workers, or external systems disagree.
The ten Agent Skills turn an unfamiliar codebase into an evidence-backed behavioral case, one handoff at a time.
system-coherence routes the next stage from the current artifacts.
Specialists record evidence, uncertainty, freshness, and the next safe
transition.
Install the complete suite into the repository you want to audit:
npx skills add "https://github.com/MPSMeridiaN/Auditor" --skill '*' --copy --yesRun it from the target repository. Or give this repository URL to your coding agent and ask it to install the complete System Coherence suite.
Invoke system-coherence in the target repository:
Run a System Coherence audit of this repository. Start from
.coherenceif it exists; otherwise initialize it. Preserve evidence for every conclusion and tell me the next safe step when a handoff is incomplete.
The optional Python coherence verifier supports contributor,
protocol, and release checks. It is not required by the installed skills.
Advanced installation → Getting Started · AI Agent Installation Contract
Agents can change. The behavioral memory stays.
Each stage writes an inspectable artifact into .coherence/. If a
session stops, the next agent reads the current handoff and resumes from the
last valid stage.
Imagine a workspace that is registered, then deleted outside the application. The UI and registry still say it exists; the filesystem says it does not.
Missing state: registered-but-missing · Broken transition: external deletion → reconciliation
The included web-cache-staleness fixture
is a runnable analogue: the database changes, but a derived cache keeps
reporting the deleted object.
The repository keeps a paper trail for structural checks, fixture behavior, and release boundaries.
The checks are intentionally scoped: a passing artifact graph or fixture run does not claim universal behavioral coverage. See Evaluation for the evidence and limits.
- Getting Started — install, start, resume, and inspect a target workspace.
- Architecture — the skill collection and optional verifier boundaries.
- Methodology — reconstruct behavior from the outside in.
- Artifact Protocol — durable handoff envelopes and payloads.
- Evaluation — fixtures, probes, and interpretation.
- Compatibility — hosts, installers, Python, and operating systems.
- Security · Contributing · Changelog
MIT. See LICENSE.