DobeWorks provides a bounded Worker and a read-only Observer. The shared terminal interface runs both alongside DEGS, the engineering evidence gate.
Install Git and Python 3.9 or later, then run these commands from a directory where you keep projects:
git clone https://github.com/tmccoy678/DobeWorks.git
git clone https://github.com/tmccoy678/DEGS.git
cd DobeWorks
python3 -B interface/command_center.py menuNo Python packages, model account, or API key are required. On macOS, zsh DobeWorks.command opens the same menu with the screenshot's Terminal settings. Keep the two folders beside each other, or use --degs-root /path/to/DEGS before menu.
Choose 01 for the paired demonstration, 07 to check both cores, or 00 to close. The examples use synthetic records: a fresh Observer record passes, a stale record blocks, and Worker output stays staged until separately accepted. Local run evidence is saved under interface/Runs/.
Interface guide and command reference · Worker and Observer details
This release contains the implemented synthetic Worker and Observer cores, the shared interface, examples, tests, and supporting documentation. Its Git tag is initial-release-final. Internal policy, phase, correction-generation, and benchmark numbers keep their existing meanings.
The Sashiko/Live-SWE executor is proposed follow-up work. The current examples do not operate a device, grant execution authority, or qualify a production system. Source archives are available with the release; the shared interface requires Git clones with an origin remote.
- Core verification and correction record and paired warning-contract verification
- Benchmark method, results, and reproduction
- Illustrated system guide
- Engineering assurance standard, citations, and historical package tests
I will briefly introduce Dobeworks and DEGS. First, DobeWorks and DEGS were originally intended for personal-use after I became quite obsessed with LLM hallucinations and developing a way to perform auditable context transfer between windows for Codex and OpenAI models in August, 2026.
I am aware that it is impossible to eliminate errors. It follows then that DobeWorks and DEGS do not try to fix errors; and one strength of the systems is their ability to prevent errors from moving forward into the next ticket or part of the work.
For a time, I used DobeWorks and DEGS in front of every message entering my regular Codex terminal. As I made the governance more exacting, ordinary work became buried under gate preparation. I created a separate external clean Workbench lane to regain balance. That history explains why a Tier 2 record with required human approval marked REJECTED was once allowed to return PASS, but it was not a truthful gate result and has been corrected in DEGS. A proper Doberman is resolute, quiet, and poised in the house; it does not need to splash in a river to prove that it is moving. DobeWorks should remain an umbrella over AI while applying governance in proportion to the work.
| September 8: recognizing the governance burden | October 22: documenting the installed external Workbench menu |
|---|---|
 |
 |
The September image is a surviving contemporaneous screenshot from the discussion that led to the external Clean Path. The October image documents the currently installed Workbench launcher after it was opened solely for this record; its capture date is not a usage date. See Governance balance and external Clean Path history for provenance and interpretation.
DobeWorks and DEGS are my personal engineering system for methodological systems, software, research infrastructure, and special projects.
If you would like to discuss DobeWorks and/or DEGS privately, please email me at: [email protected].
Thank you,
Taylor
Except for material identified in third-party notices and under its own license, DobeWorks is available under Apache License 2.0. Use issues for non-sensitive questions and SECURITY.md for vulnerability reporting.