SENTRYINDEPENDENT AI CODE REVIEW & QAOBSERVE.
INVESTIGATE.
ACCOUNT FOR IT.

Generation isn’t
verification.

THE VERIFICATION CHAMBER

Review code through independent AI investigations, preserve the evidence behind every finding, and understand what was verified—and what wasn’t.

Three independent AI harnesses.
One accountable review process.

OpenCode · DeepSeek · GLM via Claude Code
Independent investigations. Separately attributable evidence.

SENTRY DESKTOP / AUTHENTIC INTERFACEINDEPENDENT APPLICATION
Authentic Sentry desktop dashboard showing managed projects and QA activity
Product interface capture. The cinematic process below uses authored examples, not a live Sentry run.
01 / THE VERIFICATION CHAMBER

One source.
Independent perspectives.

Follow a frozen revision through three separate investigations. Every movement represents an architecture boundary, an evidence relationship or a change of representation.

Choose the sample harnesses

One, two or three. Demonstration controls only; no provider is called.

Skip the experience ↓
THE VERIFICATION CHAMBER / ILLUSTRATIVE
3 harnesses ▾
Illustrative harness selectionNo provider invocation. At least one harness remains selected.
Skip ↓
LINUX RUNTIME / CANONICAL COORDINATIONMANAGED PROJECT REVISION / ILLUSTRATIVEOffline Task ManagerOne revision. A traceable starting point.FULL QA / FROZEN MANIFESTEXAMPLE-S01 · MANIFEST + SOURCE BYTESOC / E-OC-01DS / E-DS-01GL / STATUS ONLY
01 / INDEPENDENT HARNESSOpenCodeNot started
PRIVATE DISPOSABLE WORKSPACE
DEMO-OC ← EXAMPLE-S01
ILLUSTRATIVE SOURCE
1export function createTask(title: string) {2  const task = { title: title.trim() };3  save(task);4  return task;5}
SOURCE REFERENCE / E-OC-01

An empty title can become a task.

src/tasks.ts:2–4
DEMO-OC
02 / INDEPENDENT HARNESSDeepSeekNot started
PRIVATE DISPOSABLE WORKSPACE
DEMO-DS ← EXAMPLE-S01
ILLUSTRATIVE SOURCE
1export function save(task: Task) {2  localStorage.setItem(task.id, JSON.stringify(task));3}
SOURCE REFERENCE / E-DS-01

A storage failure has no explicit recovery path.

src/storage.ts:1–3
DEMO-DS
03 / INDEPENDENT HARNESSGLM via Claude CodeNot started
PRIVATE DISPOSABLE WORKSPACE
DEMO-GL ← EXAMPLE-S01
ILLUSTRATIVE SOURCE
1export function createTask(title: string) {2  const task = { title: title.trim() };3  save(task);4  return task;5}
SOURCE REFERENCE / No evidence reported

No finding reported.

No source reference
DEMO-GL
RUN OVERVIEW / INDEPENDENT RECORDSOpenCode / DEMO-OCCompletedDeepSeek / DEMO-DSPartial outputGLM / DEMO-GLFailedONE HISTORY. NO SYNTHETIC CONSENSUS.
FROZEN REVISION → INDEPENDENT AGENTRUNS → PERSISTED HISTORY
READ THE ARCHITECTURE / EQUIVALENT STAGE EXPLANATIONS
01Every review begins with a revision.

A review starts in Sentry Desktop or its supported CLI broker. The Linux runtime owns the work; the interface presents its recorded state.

Input
A managed project revision and review request.
Operation
Establish project identity and runtime ownership.
Output
A specific revision ready for preparation.

The supported CLI is a frontend through Desktop. Neither renderer nor CLI owns the canonical database or constructs its own harnesses.

02One frozen source of truth.

The runtime freezes a manifest and source bytes from the managed revision. Every selected harness receives the same baseline in its own workspace.

Input
Managed revision, selected harnesses and scope.
Operation
Freeze the snapshot; validate targeted paths against its manifest.
Output
An immutable baseline with explicit review scope.

Full QA and Targeted QA use the same independent-agent topology. Targeted QA freezes a TargetSpec and validates it against manifest entries, including the intended focus and dependencies.

03One revision. Independent investigations.

One source becomes separate investigations. Only selected harnesses receive work; a visible inactive option has not executed.

Input
Frozen snapshot and non-empty harness selection.
Operation
Fan out to OpenCode, DeepSeek and GLM via Claude Code as selected.
Output
Independent AgentRuns with private workspace identities.

Selected adapters are preflighted and launched concurrently. These are three implementations, not quality, security and correctness departments.

04Separate workspaces. Separate investigations.

Each selected harness independently plans and investigates in a disposable private workspace. It does not write into its peers’ workspace or the retained snapshot.

Input
One immutable source baseline per selected AgentRun.
Operation
Create private working copies and independent investigation contexts.
Output
Separately attributable activities and observations.

The workspace factory associates baseline hashes and snapshot identity with each AgentRun. Durable artifacts and records remain after ordinary workspace cleanup.

05Findings need evidence.

Follow an authored code reference from an observation to a finding. A source-based concern is not a reproduced failure, and agreement cannot manufacture stronger proof.

Input
A specific source region and an independent observation.
Operation
Keep the evidence reference attached to its finding and origin.
Output
An inspectable finding with its evidence and limitations.

All code, identities and findings in this film are illustrative. No repository is imported, no provider is called, and no code or test is executed.

06Evidence survives imperfect runs.

The example investigators reach different outcomes. Completed peer results and usable partial output remain visible even when another harness fails.

Input
Each selected AgentRun’s terminal status and usable output.
Operation
Preserve peer results; derive the parent lifecycle status.
Output
A completed, partial or failed example review—not a consensus score.

In this sample DeepSeek’s AgentRun is failed with usable evidence, labeled partial output. All completed yields completed; otherwise usable output yields partial; no usable output yields failed. Runtime cancellation has separate precedence.

07Independent findings. One inspectable history.

Results arrive in the shared run overview while retaining their originating harness, evidence and status. Spatial convergence is collection, not an AI merger or verdict.

Input
Separate reports, findings, evidence and limitations.
Operation
Persist independently sourced results in the same run history.
Output
One overview containing distinct AgentRun results.

No mandatory merger, verifier, clustering, deduplication or agreement-based confidence upgrade is introduced. Historical and retest compatibility records are a different path.

08See the evidence. Understand the result.

Select a harness result and inspect its finding, source reference, snapshot identity and limitation. A failed reviewer is never silently presented as complete.

Input
Persisted run identity and individual results.
Operation
Present records and terminal-run JSON/log exports.
Output
A review history you can inspect and export.

This DOM presentation follows inspected fields and UI structure. It is authored sample data, not a live application or genuine persisted run.

09Independent verification. Actionable evidence.

Sentry remains an independent application. Its recorded results can inform the next engineering decision without granting execution or completion authority.

Input
Review history, evidence and unresolved concerns.
Operation
Keep results accessible beyond temporary workspaces.
Output
Traceable observations for the next decision.

An optional return toward Nexus is intended ecosystem architecture. The inspected native Nexus Sentry page does not expose a working Sentry API.

02 / THE THREE HARNESSES

Independent by
implementation.

Each selected harness independently plans, investigates and reports. They are not three fixed departments. A similar observation from two reviewers still needs its own evidence.

01

OpenCode

OpenCode harness

Its own session, disposable workspace, events and report. Readiness and configured provider access are required before execution.

02

DeepSeek

DeepSeek harness

Its own session, disposable workspace, events and report. Readiness and configured provider access are required before execution.

03

GLM via Claude Code

Claude Code harness · GLM

Its own session, disposable workspace, events and report. Readiness and configured provider access are required before execution.

03 / FULL QA & TARGETED QA

Change the focus.
Keep the independence.

Full QA investigates the project broadly. Targeted QA freezes an explicit focus against the snapshot manifest, with investigation of relevant dependencies. The selected harnesses remain independent in both modes.

EXAMPLE-S01 / FROZEN MANIFESTsrc/tasks.tsProject scopesrc/storage.tsProject scopesrc/view.tsProject scope
FULL QA / BROAD REVIEW

The project is the review subject.

This comparison changes scope, not execution topology.

OpenCodeSelected · own workspaceDeepSeekSelected · own workspaceGLM via Claude CodeSelected · own workspace
04 / EVIDENCE & PROVENANCE

A finding keeps
its origin.

Know who reported the concern, where its evidence came from and which frozen revision it describes. Source inspection remains different from a reproduced failure.

  1. FindingEmpty task title acceptedDEMO-OC / source example
  2. Evidencesrc/tasks.ts:2–4E-OC-01 / code reference
  3. AgentRunOpenCodeOrigin remains attributable
  4. SnapshotEXAMPLE-S01Frozen baseline + manifest

Illustrative identity chain. No executed test, confidence upgrade or live finding is implied.

05 / BUILT FOR REAL PROJECTS

Beyond a
single response.

Managed projects, review history and independently inspectable agent results keep the work understandable after an investigation ends.

Authentic Sentry UI fixture showing independent agent results and a finding inspector with source location and evidence
Authentic product screenshot from a documented UI fixture. The fixture’s findings, times and statuses are sample data.

Managed revisions

Project operations preserve managed revisions and snapshot identity.

Cancellation & recovery

The runtime owns cancellation and settlement. A disconnected interface does not prove a run has stopped.

Durable results

Findings and evidence survive ordinary disposable-workspace cleanup.

JSON / log exports

Terminal-run exports retain individual results and provenance.

06 / DESKTOP & SUPPORTED CLI

One runtime.
Two ways in.

Desktop and the packaged CLI use the same projects and canonical review history. Open Desktop first; the CLI talks through its supported broker.

VERIFIED COMMAND REFERENCE / NOT LIVE OUTPUTsentry open
sentry cli capabilities --json
sentry cli project list --json
sentry cli run status run_ID --json

Replace run_ID with a real run identity. No command is executed here. Credentials remain in Desktop settings; client identity is informational, not authentication.

Inspect runtime and interface responsibilities

The Linux runtime owns SQLite persistence, coordination, adapters and project operations. Desktop main owns runtime connection, host path translation, credentials and native export dialogs. The renderer reads snapshots. The CLI never creates a separate domain database or its own agents.

07 / AN INDEPENDENT PERSPECTIVE

Evidence for
the next decision.

Sentry is an independent application in the ArthurLabs ecosystem. The intended Nexus relationship carries scope and evidence into review and returns findings to the relevant work.

Native Nexus integration is unavailable in the inspected Desktop. A supported Sentry CLI contract does not establish an implemented Nexus UI connection.

08 / DOWNLOAD SENTRY

Independent review.
On your machine.

Install Sentry as its own application. Review the prerequisites, verify the archive, and configure the harnesses you intend to use.

VERIFIED PUBLISHED ARTIFACT / 1.0.2

Windows 11 + WSL2

x64 · Unsigned ZIP · online runtime bootstrap
122.9 MB · packaged production CLI included

Download Sentry 1.0.2 ↓
BEFORE YOU INSTALL
  • Windows 11 x64 + WSL2; Ubuntu preferred.
  • Online setup; Linux Node ≥22, pnpm 9.12.3, systemd user services, bubblewrap and socat.
  • Provider tooling, configured accounts and prerequisites required. Provider usage charges are separate.

This is an unsigned archive with online WSL bootstrap. It is not a signed installer or an offline package.

Fresh installation & upgrade instructions ↗
Artifact identity and checksumsentry-1.0.2-2640efcd7874-windows-x64.zip
SHA256: 5265713af248297c543bf0d41093258bd0e7d8c177e0b1c544e733750347286f
Source: 2640efcd78744cb02c517a405e648722e662bd60

Published release metadata checked 22 October 2026. Native Linux Desktop is deferred; macOS support is not established. The Linux execution runtime through WSL2 is a separate requirement.

09 / DOCUMENTATION & ECOSYSTEM

Keep the evidence.
Understand the system.

Installation guidance and implementation references belong alongside the download. Explore Sentry independently, or see its intended place in the wider ArthurLabs workflow.

02 / ENGINEERING

Inspect the boundaries.

The production pipeline contract distinguishes separate investigations from historical finalization concepts.

Read the architecture ↗
03 / SOURCE

Follow the implementation.

Inspect the published source revision behind this verified distribution. No live run is performed by this website.

Explore the source ↗
INDEPENDENT BY DESIGN

A perspective within
a greater workflow.

Arthur Studio holds ideas. Nexus coordinates work. Sentry investigates independently. Their connected workflow remains a conceptual demonstration where integration is unavailable.

Explore the ArthurLabs Suite ↗