TownReporter 0.6.33 beta

Known issue identified after release: Dark Desk can omit relevant captured records from synthesis and the editor brief. The Nelson annexation PDF was successfully captured but excluded by newest-only selection and head-only excerpts. This evidence-selection and adjacent-record attribution issue is corrected in v0.6.34; this dated v0.6.33 release remains unchanged. The attached investigation report also corrects capture counts that mistakenly treated changed records as failed fetches.

TownReporter 0.6.33 beta

This beta fixes native Codex drafting in extracted Windows installations. Codex CLI 0.145 refuses a non-Git working directory unless its documented --skip-git-repo-check option is supplied; TownReporter now supplies that option while preserving the selected sandbox, approval mode, user configuration, authentication, model, prompt transport, web search, and image arguments.

Focused proof includes a native non-Git RED/GREEN probe: the previous invocation refused, while the corrected invocation returned successfully in 17.9 seconds. All five bounded patched-application investigation runs completed from the Git source checkout; none created a handoff or publication. See the five-investigation receipt. These results establish launch and workflow execution only. They do not establish investigative accuracy, independently corroborated findings, completion of the September 3–8 historical replay, Halo deployment, or measured normal-day performance.

The release retains the 0.6.32 evidence-review presentation fixes and prior beta functionality. See docs/HALO-UPDATE-0.6.33.md for the Halo-local staging and promotion procedure after publication.

A separate bounded budget exercise produced two protocol-verified labels. In TownReporter, that label means the documented verification protocol completed; it does not mean two findings were independently fact-checked or corroborated. Human review remains required.

Exact source: 62ae453f595fafb0157c6b681efef302bbab9e53. Application CI passed 14/14 and Windows package/install passed 2/2. Across the aggregate test evidence: 1,953 total, 1,899 passed, zero failed, 54 explicitly skipped (npm script layer 342/333/9; TypeScript test layer 1,611/1,566/45). The Windows ZIP is 7,116,904 bytes with SHA-256 0ad28695ac43114eb962dceca0245f5aabc6be0f6cda41cf645cdb24bcc9f482. Halo-local 0.6.33 prompt.

Published Windows ZIP verification succeeded in 14m01s after downloading and verifying the published assets, installing the private Windows runtime, exercising the browser workflow and restart persistence, checking ownership/credential ACL/lifecycle contracts, and stopping the owned installation.