LESS GUESSING. MORE EVIDENCE.
A native Windows diagnostic workstation for technicians who want evidence, not guesswork.
Built by ROOKWELD · Open source · Designed for real investigations
Evidence Trials (live demo) · See the workstation · How it works · Get started · Safety · Field Kit
An actual WPF + WebView2 Windows application. Not a browser tab.
Explore the interactive Evidence Trials →
Three reproducible synthetic investigations expose FieldRook's source-linked evidence, hypotheses, action boundaries and independently measured outcomes: a suspicious five-minute task (cause unverified), a simulated worker recovery (VERIFIED_RESOLVED), and a reported fix that fresh evidence rejects (NOT_RESOLVED). No customer data, real host changes, or live-model performance claims. The public methodology and export scripts allow independent inspection.
FieldRook turns an unfamiliar Windows problem into a structured investigation. It helps a technician capture observations, run bounded diagnostics, examine traceable evidence, test competing explanations, and record whether a resolution was actually verified.
The AI is an assistant to the technician—not an unrestricted shell, a source of invented facts, or an authority that can silently declare a machine fixed.
| Investigate | Understand | Verify |
|---|---|---|
| Describe the issue and select the affected computer. | Inspect observations, provenance, hypotheses, and the next useful question. | Record what changed, distinguish claims from measurements, and export the case. |
FieldRook 0.5.0 introduces SIGNALFORGE: a retro-futuristic instrument-inspired interface, with deep plum surfaces, coral highlights, selective acid-lime signals, and restrained circuit-trace motion.
Its new native WPF shell embeds the existing React workstation through WebView2 and manages the packaged local FastAPI server. The usual experience is one Windows window, not a manually launched backend plus a browser tab.
The SIGNALFORGE Workbench — actual FieldRook 0.5.0 UI capture, not concept art.
More native-app captures — evidence, Collector, resolution, and settings
▶ Watch the 33-second SIGNALFORGE release trailer (MP4) · Download v0.5.0
Real-product footage. The trailer is edited from an OS-level capture of the shipped v0.5.0 native Windows application, running isolated simulated demo investigations. Slow camera pushes, typography, transitions, and an original electronic sound bed were added in post-production; the application's UI and case data were not fabricated or altered. Decorative interface motion is not telemetry.
The background motion is deliberately subtle. The interface offers Alive, Quiet, and Off preferences and supports reduced-motion behavior. Decorative activity is not presented as a real hardware measurement.
Technician describes an issue
│
▼
Choose target: this PC / another PC / not yet known
│
▼
Run approved local diagnostics or import a Collector package
│
▼
Timestamped evidence + provenance + visible uncertainty
│
▼
Review hypotheses and the next useful check
│
▼
Technician observes, confirms, or supplies new evidence
│
▼
Document outcomes and export a case report
| Capability | What it does |
|---|---|
| Native Windows workstation | Opens in its own WPF/WebView2 window with the SIGNALFORGE interface. |
| Guided investigations | Keeps the problem, current questions, evidence, and investigation history together. |
| Windows diagnostics | Uses registered, bounded diagnostics with explicit target and tool boundaries. |
| Portable Collector | Gathers read-only evidence on a Windows computer and exports a versioned, hash-verified ZIP for import. |
| Evidence and provenance | Shows where observations came from and distinguishes collected measurements from user reports and model interpretations. |
| Knowledge support | Retrieves relevant technical references as quoted context with citations; irrelevant matches are filtered conservatively. |
| AI provider choice | Supports optional local-model inference and configured cloud inference; simulated operation remains labeled. |
| Technician-controlled actions | Keeps proposals, approvals, execution, and verification separate. The only implemented corrective execution targets FieldRook's own demo worker. |
| Case persistence and export | Reopen investigations and produce technician-readable case reports. |
| Motion preferences | Tune the ambient UI between Alive, Quiet, and Off. |
Important: An imported Collector package does not give FieldRook remote command execution. Host-local measurements cannot be silently substituted for a different machine, and a remote reading cannot be rechecked against the operator's own PC.
The Collector is a separate Windows executable. It is designed to run on an affected computer without requiring a Python installation, gather selected read-only diagnostics with consent, and create a package that FieldRook can validate and import.
Affected Windows PC Technician's FieldRook
┌─────────────────────┐ ┌─────────────────────┐
│ FieldRook Collector │─── evidence ZIP ────▶ │ Import + validate │
│ read-only capture │ │ investigate │
└─────────────────────┘ └─────────────────────┘
Collection is not silently uploaded. Some elevated collection categories have not yet been validated in the release candidate; permissions and unavailable measurements must be reported honestly.
The newest Windows desktop package is v0.6.0-rc.2. It includes the Field Kit dock, Evidence Lens and Network Lab, plus M030 local-model reliability and security fixes: cache recovery, fast cancellation during Ollama probes, honest API-style reporting and loopback-only probe validation. This is a pre-release, not a stable or independently certified build.
- FieldRook-0.6.0-rc.2.zip — portable Windows desktop app; extract the whole folder and run
FieldRook.exe. - SHA256SUMS.txt — release checksum manifest.
- BOM.txt — exact toolchain, source commit and hashes.
- RELEASE_VALIDATION.md — verified test results and remaining limitations.
Keep the extracted application files together. On first launch, configure the desired inference mode. Windows 10/11 x64 and WebView2 Runtime are required. The binaries are unsigned; verify the published checksum before running. No model is bundled by default, and offline evidence workflows do not depend on paid cloud inference.
This release does not include a new, separately packaged Collector download. The older v0.5.0 Collector remains available below, but cross-version compatibility has not been certified.
For details, see the v0.6.0-rc.2 release notes and provider configuration.
The earlier v0.6.0-rc.1 remains available for reference, including its Windows ZIP and release notes.
The earlier v0.5.0 release is retained for reference and includes two separately published packages:
- FieldRook-0.5.0.zip — previous portable desktop app.
- FieldRook-Collector-0.5.0.zip — portable read-only Collector.
- SHA256SUMS-v0.5.0.txt — release checksums.
v0.5.0 does not contain the Field Kit instruments introduced in v0.6.0-rc.1. Historical native-delivery notes are in M026.
This repository contains the Python/FastAPI backend, React/TypeScript frontend, native Windows shell, and packaging scripts. Development requires the corresponding build tools; end users of the packaged Desktop do not need Python or Node.js installed separately.
From a Windows development checkout with Python, Node.js/pnpm, and .NET 10 SDK, follow the desktop build documentation and the M026 delivery manifest. The native shell and backend are packaged separately, then assembled into the complete distribution. Do not confuse the older browser-based development launcher with the packaged native Windows application.
FieldRook deliberately makes certain operations harder to do by accident.
- Read-only by default. Windows diagnostics use registered capabilities and bounded inputs, not arbitrary model-authored PowerShell.
- Target-aware evidence. Collector imports are tied to the affected computer; mismatched identities are not silently mixed.
- Human control. An AI suggestion is not permission to execute it. Corrective actions require explicit approval and a registered execution path.
- Measured verification. An intervention being attempted or reported does not mean the problem was resolved.
- Local-first processing. The packaged service binds to loopback and uses application-owned local storage. Cloud-provider use requires deliberate configuration and consent.
- Provenance and privacy. Collector archives are integrity-checked, and diagnostic data is not automatically transmitted elsewhere.
- Separation of fact and inference. Model explanations and confidence language never replace measured evidence.
Security documentation: FORGE privacy · Tool boundaries · Technical highlights
| Layer | Implementation |
|---|---|
| Desktop host | .NET WPF + Microsoft Edge WebView2 |
| Frontend | React + TypeScript + Vite, SIGNALFORGE design system |
| Local API and investigation engine | Python + FastAPI |
| Case persistence | SQLite in per-user application data |
| Windows evidence | Allowlisted diagnostic toolpacks and portable Collector |
| Inference | Optional configured local or cloud providers; no unrestricted shell authority |
| Validation | Pytest, Playwright, .NET tests, diagnostic gauntlet, frozen-app smoke tests |
FieldRook was developed as part of the Meta Global AI Developer Hackathon 2026. This project is independent; participation is not an endorsement by Meta.
For the historical v0.5.0 / M026.1 release-candidate verification (plus green GitHub Actions on the merged source), the records show:
| Gate | Reported result |
|---|---|
| Backend | 687 passed, 4 environment-dependent skips |
| Playwright end-to-end | 47 / 47 passed |
| .NET tests | 66 / 66 passed |
| Diagnostic gauntlet | 38 / 38 passed |
| Native packaged journey | Investigation, Collector import, export, and persistence exercised |
| Security probes | 11 / 11 passed |
These are local release-candidate validation results, not a claim that every user environment, Windows machine, AI provider, or troubleshooting problem has been certified.
Still to validate: unsigned-binary reputation and antivirus behavior across vendors; Windows machines without WebView2; multi-machine field trials; upgrade-over-existing installation; selected elevated Collector categories; lower-end CPU-only local-model responsiveness; some performance combinations on other GPUs.
The right-hand dock is now a modular, context-aware technician toolbox. Shipped in this source tree as evidence-backed instruments with bounded diagnostics and clear local-versus-remote behavior:
- Evidence Lens -- deterministic before/after record compare over persisted evidence, with contradiction citations and export linkage.
- Network Lab -- read-only networking instrument (adapters, addresses, DNS config, resolution test) with consented re-checks and a remote-target guard.
- Field Kit dock -- instrument framework with pins, keyboard access, and narrow-overlay support.
These instruments are included in the published v0.6.0-rc.1 desktop pre-release, and are covered by the demo scenario in M028_THREE_MINUTE_DEMO.md. They are not included in the earlier v0.5.0 binaries. The remaining modules (Process Watch, Event Decoder, Command Deck, Field Notes) are still design-prototype concepts, not shipped features.
- M026 native desktop delivery
- Hackathon build log
- Provider documentation
- Diagnostic tool reference
- Security and privacy
- License







