Code Hangar 0.1.2 alpha
Pre-releaseCode Hangar 0.1.2 judge candidate
This local candidate packages the eligible OpenAI Build Week work while keeping
event material isolated from reusable development.
Candidate highlights
- reconstructs AI-assisted changes from local session, Git, current-file, and
saved-review evidence, with coverage and unknowns shown first; - improves large-project navigation through bounded progressive Hangar Map
loading and separates dependency-cache observations from direct project
signals; - keeps graph/API/MCP list requests within a hard 1,000-item ceiling;
- adds the explicit OpenAI GPT-5.6 Connector preset and direct-OpenAI
max_completion_tokenscontract while preserving compatible providers; - exposes the existing scoped/audited Code Hangar MCP sidecar as the primary
GPT-5.6 judge journey through ChatGPT-authenticated Codex; - retains the hard Local-versus-Connector build boundary, secret checks,
Protected Zones, exact-request review, and reversible correction workflow.
Editions
Code-Hangar-AI-Connector_0.1.2_x64-setup.exe— primary judge build with
opt-in AI Assist and local MCP sidecar.Code-Hangar_0.1.2_x64-setup.exe— Local isolation-proof build with no
AI-provider or MCP frontend surface.
Both packages are Windows x64 preview builds. Final hashes, signing inspection,
and any clean-install limitation are recorded in the evidence manifest rather
than copied from an older release.
Proof and owner-gated external actions
A synthetic subscription-backed GPT-5.6 Sol + Code Hangar MCP proof passed
against the final compiled sidecar, including two scoped audited reads,
temporary credential revocation, and zero persisted Codex session files. The
public assets were re-downloaded and their SHA-256 values matched the published
manifest. Exact host lifecycle was not repeated for these release hashes. An
empty network-disabled Windows Sandbox refused the final downloaded Local setup
under guest Application Control before product execution; its fail-fast result
left Connector unexecuted. Clean disposable-machine proof and the final public
installed-product video remain outstanding. A separately billed in-app OpenAI
API call is optional and has not been run.
The privacy-sanitized source repository is public. This candidate does not claim
/feedback, YouTube publication, or Devpost submission until the owner
explicitly performs or authorizes those account-bound actions.