Rangabot updates

Meaningful changes, with their evidence attached.

A chronological record of what changed, why it matters, what stays private and whether the work is a candidate or an actual release.

Implementation, verification and release are different facts.

Every entry names its exact status and build. Candidate work is not presented as shipped, and missing or not-run evidence stays visible.

Updates are newest first. Future entries will retain the same plain-language structure: user problem, change, privacy boundary, evidence, limitations and release status.

Recovery · local context · clearer identity

Failed work can return as a reviewable draft.

Rangabot 1.2.0 adds user-triggered failed-turn recovery, stronger project and dataset binding, revisioned local dataset context, Profiles v1 candidate workspaces, and the restored transparent assistant chat mark.

Goal preserved

Recovery restores the exact stored request, mode and reply receipt only after conversation, history, project, dataset and code bindings revalidate. Failed partial output is excluded.

Nothing automatic

The recovered request returns to the composer for review. Rangabot does not rerun the model, reopen private context or press Send on the user's behalf.

Local context stays bound

Dataset descriptions, aliases and confirmed joins use revisioned local storage. Concurrent windows and changed resources fail closed instead of silently changing the request's meaning.

Evidence and release boundary

Independent source-level recovery acceptance passed 24/24 cases. A separate binding harness moved from 0/5 protected cases on the merged base to 5/5 on the release source. The deterministic repository release gate also passed.

These are source checks, not model-quality, signed-package, TestFlight, Store, clean-install or public-binary evidence. The 23 August Mac 0.1.0 submission and earlier Light artifacts are historical candidates, not 1.2.0 binaries.

Open the exact source release →

Mac distribution · experimental analytics

Signed Mac beta submitted; safer SQL planning added to source.

Rangabot reached two meaningful milestones without hiding their remaining gates: a signed Mac App Store package entered TestFlight Beta App Review, and the source gained stronger schema-grounded SQL planning.

Mac beta

The Mac App Store build was signed, uploaded and submitted to TestFlight Beta App Review. At submission it was Waiting for Review, not publicly released.

SQL source

Relationship-aware schema linking, typed table and column meaning, confirmed joins, operation decomposition and intent checks now sit inside the local read-only analytics path.

Privacy boundary

No model weights or user data are in the app package. Dataset access remains explicit, local and read-only; semantic context cannot add permissions or a new requested operation.

Evidence and release boundary

Disclosed internal SQL matrices improved from 571/720 to 720/720 and from 146/640 to 624/640 after engineering. Those are regression results, not independent universal proof.

On external BIRD Mini-Dev, typed schema context improved exact results from 1/500 (0.2%) to 18/500 (3.6%). That is measurable progress and also clear evidence that Rangabot is not yet a general SQL expert.

The signed Mac beta still requires Apple approval before testers can install it. Public App Store release, Windows public release and a general-SQL readiness claim remain open gates.

Inspect the source and evidence →

Smaller downloads · reuse what is installed

Light desktop editions omit the Ollama runtime.

Rangabot can now be packaged without a second copy of Ollama when a compatible local Ollama service is already running.

What changed

Light packages connect only to 127.0.0.1:11434, accept stable Ollama 0.32.x, and do not search for or launch an executable from the system path.

Why it is smaller

No model was ever bundled. Light additionally removes the runtime itself: roughly 396 MB from the Mac app and 1.9 GB from the unpacked Windows payload.

What users provide

Ollama installation, updates, logs and model storage remain separate. Rangabot neither stops nor reconfigures that external service.

Evidence and release boundary

The Mac Light ZIP is exactly 165,584,612 bytes with SHA-256 134e1298971f0836228f2434c8e317ba958b6d4219ba73e572e8f3a052d8533d; its extracted application passed the governed artifact verifier. It remains ad-hoc signed and unnotarized. The Windows Light binary has not yet been produced on the pinned Windows SDK, so its exact size is unknown.

Neither binary is a public download. Release requires platform signing and clean-machine acceptance of the exact final bytes.

Convenience · local ownership

Rangabot for Mac

Rangabot is becoming an application people can open and close like any other Mac app—without starting it from Terminal.

What is changing

The local Rangabot experience is being packaged inside a native macOS application with a Finder-first launch and a private per-user data boundary.

Why it matters

Starting Rangabot should be simple enough for everyday use while preserving the local capabilities and ownership model behind the project.

What stays local

The desktop architecture keeps the server on an isolated loopback address and separates bundled application resources from private user data.

Evidence and release boundary

The current Apple-silicon verification candidate passes source checks, signed-bundle integrity checks and packaged readiness, private bootstrap and authenticated API smoke tests.

This is a coming-soon development announcement, not a download or public release. Developer ID signing, notarization and broader release testing remain before general availability. Windows and Ubuntu feasibility follow the macOS release path.

User control · local evidence

Local Response Feedback Pulse

A quick, reversible way to mark a completed Rangabot answer Helpful or Needs improvement—without surrendering the conversation that produced it.

What changed

Completed responses receive two clear controls. A selection can be changed or cleared, and Rangabot confirms exactly what was saved locally.

Why it matters

Users can signal whether an answer helped in one step while keeping control of the feedback and the conversation.

What stays private

No prompt or response text, title, reason, memory, attachment or personal content enters the feedback record.

Evidence and release boundary

The candidate passed independent technical, privacy and integrity review plus five synthetic persona protocols, including a bounded 20-turn conversation. Manual assistive-technology details, induced persistence-failure recovery, cross-platform coverage and a live aggregate transfer remain not run.

This entry records verified candidate behavior. It does not claim production adoption, satisfaction, public availability or release. The status will change only after the corresponding build is released.

Inspect merged pull request #102 →