TownReporter 0.6.31 beta

TownReporter 0.6.31 beta

TownReporter 0.6.31 is a beta prerelease from source 58f5760704d45dce972814976ab07c6f7f76d6d8. PR36 passed all 16 checks and merged as 6f1ca348847609281c2d4772623a695535104cb4. This GitHub release does not deploy the paper on Halo.

What is included

• Community focus: The Queue keeps the full lead set and offers an optional focus on 3–5 existing eligible leads. The editor keeps manual control and reviews evidence before drafting.
• Claim-level review: PR33 and PR34 add private editor judgments against captured evidence and bind them to the saved draft. This is review support, not exhaustive claim coverage or independent fact-checking.
• Approved notice roundups: Six owner-approved families cover library, recreation, community and arts events, registration deadlines, waste and recycling, and public-meeting logistics. Each source and format requires explicit owner approval; ambiguous or unsupported material is refused or held for review.
• Routine editions: Approved sources can feed Today in town, This weekend, and Deadlines approaching through fixed templates and public source links. The scheduler runs daily, not continuously. An approved source change can correct an automation-owned edition on a same-day rerun.

Install

Download the beta Windows ZIP from the 0.6.31 release asset. Extract it, open Install TownReporter.cmd, create the owner account, and configure AI with Configure AI.cmd or a supported signed-in CLI/local endpoint. The stable Windows release remains 0.6.30.

The installer provisions private Node/PostgreSQL runtimes, persistent storage and Chromium, and checks local HTTP readiness. It does not configure Halo, a public tunnel, or a paid provider fallback.

Limits and acceptance

Daily scans allow up to 12 accepted sources with one scheduled run per newsroom day. Draft batches contain 1–5 leads with existing rate and per-lead job protections. Routine editions use at most five useful logistics items and publish no empty edition. The 60-minute editor-work and 3–5-useful-item figures are acceptance targets, not guarantees.

Existing subscription tools and local runtimes are supported. Ordinary reporting remains human-reviewed and human-published; automatic publication is limited to the approved notice families after eligibility, evidence, conflict, history and idempotency checks. Broader community coverage and field acceptance remain open. Five real Dark Desk investigations remain a Halo-local exercise. No new metered paid fallback is enabled.

Receipts

• PR36: all 16 checks green, head 58f5760.
• Application CI: run 34209082147: scripts: 342 total, 333 passed, 0 failed, 9 skipped; TypeScript: 1,609 total, 1,564 passed, 0 failed, 45 skipped; combined: 1,951 total, 1,897 passed, 0 failed, 54 skipped.
• Windows CI: run 34209082139 green.
• Merge: 6f1ca348847609281c2d4772623a695535104cb4.
• Beta ZIP: TownReporter-0.6.31-windows-x64.zip, 7,120,307 bytes, SHA-256 3171479a4cd7b5d9325518ad6a27d2550e559b87c17e48fcaebd764a03899c81.
• Local packaged acceptance: PASS with Ollama gemma4:12b: supplied-only draft, human body edit, explicit keep-evidence decision, confirmed publish, Stop/Start, and anonymous article persistence. Two additional quality samples were not run. This does not prove 3–5 useful items per day, 60-minute editor time, or a model-duration benchmark; no such performance claim is made.

Documentation/version checks passed before the final source-copy integration: versions-agree 1/1, scoped ESLint passed, and git diff --check passed.

The 54 aggregate skips comprise 44 opt-in real-PostgreSQL tests (covered by the separate green PostgreSQL job), six Windows/PowerShell-only checks, two build-dependent checks, one authenticated GitHub security-setting check and one opt-in live-model evaluation. Separate Windows/build jobs passed; the authenticated repository-setting check and live-model evaluation were not run in CI. The local-model editorial exercise above is separate evidence.

Published Windows verification also passed: run 34213268197 downloaded this release's assets and completed installation, browser, restart persistence, stale-build refusal, ownership and credential ACL checks in an 11m38s job. This fresh Windows exercise uses a manual article with AI disabled; the real local-model editorial exercise above is separate. All three downloaded release assets matched the tested package byte for byte.