Cotabby — McHamster’s Build 0.6.2.2 (37.50%)

Cotabby — McHamster's Build 0.6.2.2

An independent experimental build of Cotabby, published from McHamster's master branch. This is not an official upstream release.

Changes since McHamster 0.6.2.1

• Finish words without fighting spell-check. Unfinished words can reach the completion engine; typo suppression and correction wait for a committed word boundary. Final unusable model output can fall back to an exact-prefix ending from local dictionaries, document vocabulary, or the existing glossary, without another model request.
• More stable ghost text. Presentation checks the assembled word at the caret, caches spelling decisions per word, and rejects late partials after a generation finishes. A short bounded delay follows typing cadence while a word is unfinished.
• Respect dismissals. Escape temporarily remembers a dismissed continuation for the same field and nearby text, reducing repeated suggestions. This memory is bounded, expires, and is not persisted.
• Compact context and adjusted sampling. Surface context uses compact format/app/domain/title/field labels, and the default repetition penalty changes from 1.05 to 1.025. Existing context scope and backend selection remain unchanged.
• Dictionary and metrics improvements. Dictionary language selection remains available when typo correction is disabled, and recovered word completions are counted consistently in quality metrics.
• Expanded evaluation tools and evidence. Includes contextual phrase scenarios, three-worker evaluation, baseline reports, experiment and completion-prefix analysis scripts, and the completed September 17 experiment reports with reproduction instructions. Adds regression tests for caret-word boundaries, local prefix matching, streaming, dismissal memory, timing, and context formatting.
• Release packaging. Each DMG now includes its own version's release notes instead of always copying the first McHamster release notes.

Evaluation context

The checked-in experiment report compares a controlled baseline with the selected compact-context/sampling configuration using Qwen3.5-0.8B-Base Q6 and a synthetic English corpus. Its integration comparison reports exact next-word accuracy of 34.772% → 37.824% with screen context and 28.506% → 31.222% without screen context. These are configuration comparisons on that benchmark, not a measurement of overall release-to-release writing quality or a guarantee for other models, languages, or apps. See the full results and limitations.

Installation and identity

• Apple Silicon, macOS 14 or later. Apple Intelligence requires compatible hardware and macOS 26 or later. No Intel build or model weights are included.
• Download Cotabby-McHamster-0.6.2-mchamster.2-arm64.dmg and drag Cotabby McHamster.app into Applications, replacing the previous McHamster build. Quit the app first.
• The app keeps org.mchamster.cotabby, your McHamster settings/model storage, and Jorge Miguel Casler's Developer ID team 8RN882MNR5.
• Upstream Cotabby remains a separate installation. Quit the other build before enabling autocomplete. Upstream Sparkle updates remain disabled; Check for Updates opens this fork's releases.
• Screenshot/OCR processing and the new dictionary fallback remain local. The existing explicitly selected endpoint backend can receive bounded text context; this release adds no new hosted transmission.
• Original Cotabby authorship, licensing, and third-party acknowledgements are retained.

Tag: mchamster-v0.6.2.2. In-app version: 0.6.2-mchamster.2. Build: 2026091702.

Release verification

Developer ID signed, accepted by Apple notarization, and stapled. Gatekeeper reports Notarized Developer ID. Validation passed: 156 targeted Swift tests, 52 Python tooling tests, and native token-healing tests. The mounted DMG passed signature, identity, version, release-note, and secret-scan checks. SHA-256 checksums are attached.

Cotabby — McHamster’s Build 0.6.2.1 (34.50%)

Cotabby — McHamster's Build 0.6.2.1

An independent, experimental McHamster build of Cotabby, based on upstream commit ac2699e4ab40b68cf7429b7d13d5a7206c99dbbc (after v0.6.2-beta). This is not an official upstream Cotabby release.

What changed in this fork

• Llama prompt-cache reuse: restores a validated prompt prefix before another generation, including after cancellation; adds native cache checkpoints and diagnostics instead of permanently abandoning reuse after one failed trim.
• Word continuation: adds native token-prefix constraints and a Swift token-healing buffer so partially typed words can continue across token boundaries without replaying their already typed prefix.
• Typing and acceptance: refines active-tail reconciliation, cancellation ownership, and streamed suggestion presentation timing as the user keeps typing or accepts text.
• Local visual context: refreshes screenshot/OCR context while the same field remains focused, bounds and selects relevant excerpts, avoids secure fields, checks permission and focus changes, and clears stale context after failed captures. Screenshot/OCR processing remains local; the existing explicitly selected endpoint backend may receive bounded text context.
• Prompt construction: adjusts section budgets and the stable conditioning prefix to support cache reuse and cleaner continuations; improves output normalization.
• Evaluation coverage: adds typing-session scoring, a 1,337-case phrase-prediction fixture and runner, token-healing tests, visual-context tests, and native cache integration coverage. These are development tools, not a claim of measured quality or speed gains for every model.

The app changes originated in fork commits fb47cb2 and b345ced. This release also includes the accompanying native-runtime changes as patches/cotabbyinference-mchamster.patch, applied to CotabbyInference commit 7574a21. The release script builds this exact patched dependency rather than relying on a developer's adjacent checkout.

Separate from upstream

• App: Cotabby McHamster.app; bundle identifier: org.mchamster.cotabby.
• Signed using Jorge Miguel Casler's Developer ID, Apple team 8RN882MNR5.
• Separate preferences, Keychain service, macOS permission grants, downloaded models, and file logs. Existing upstream settings and models are not migrated automatically.
• Upstream Sparkle updates are disabled. Check for Updates opens this fork's GitHub releases; updates are installed manually.
• Feedback links point to this fork. Upstream's release workflow, update domain, and Homebrew publishing are guarded against use in the fork.
• Release tag: mchamster-v0.6.2.1; in-app version: 0.6.2-mchamster.1.

Installation

Apple Silicon Mac, macOS 14 or later. Apple Intelligence requires compatible hardware and macOS 26 or later; other engines retain their existing requirements. This asset does not include an Intel build or model weights.

Download the Cotabby-McHamster-0.6.2-mchamster.1-arm64.dmg asset, open it, and drag Cotabby McHamster into Applications. Grant Accessibility and Input Monitoring to this separately named app; grant Screen Recording only if using visual context. Quit the other Cotabby build before enabling this one: separate identities allow both to be installed, but both monitor typing and should not run autocomplete simultaneously.

Original authorship and license remain with Cotabby and its contributors; see the repository's LICENSE and acknowledgements. The upstream Homebrew cask installs upstream Cotabby, not this build.

Release verification

The Apple Silicon DMG is Developer ID signed, accepted by Apple notarization, and stapled. Gatekeeper verification reports Notarized Developer ID. SHA-256 checksums are attached. Native token-healing tests and all six phrase-evaluation CLI tests passed against the tagged release sources.