Repository navigation
Releases: ShiftHackZ/ellmsha
Release list
eLLMsha 2026.10.22-patch.4
eLLMsha — October 22, 2026
New
- Desktop browser tabs now keep separate sign-ins, cookies and browsing data for each project. Chats in the same project share a browser profile; chats outside projects use a shared default profile.
- Auto Router can choose models for the main agent and its subagents. Its visual editor lets you assign different models to tasks such as coding, research and review, with separate saved settings for each mode.
- Usage statistics are available for individual chats and across the app. See input and output tokens, cached input, context size and tool usage, with global filters by date, provider and model.
- Added bundled community skills for research, writing, interface design and diagrams.
Improvements and fixes
- Browser tabs preserve filled forms and popup windows when switching projects. Browsing history and data-clearing controls are separate for each project.
- Improved website compatibility with an updated desktop browser engine.
- Fixed browser automation hanging when an agent reconnects to an already open tab.
- Agents read project rules before starting work, including tasks delegated to subagents.
- Reduced unnecessary context sent to AI models when many tools are enabled.
- Fixed the main window taking focus when releasing a dragged desktop pet.
- Refreshed the application icon and corrected its Dock size.
eLLMsha 2026.10.14-patch.11
v2026.10.14 Patch 11 — chat controls and subagent status
Release candidate prepared on 2026-10-14 (Europe/Kyiv). Publication, installer validation and installation of this candidate are pending. The release pipeline builds signed and notarized macOS Apple Silicon installers, an unsigned Windows x86_64 installer and four Linux/amd64 VPS images in parallel; publication requires every branch to pass.
Changes
- Show and drag the desktop pet through nonactivating macOS panels, preserving the separate Dock action that opens the main window.
- Open command completion consistently at a valid slash invocation in either composer. Preserve surrounding text and caret position when selecting a command, handle empty results and input methods, and highlight known skills in the draft.
- Render BMAD document links and inline artifact buttons with readable Light/Dark colors, keyboard focus and disabled states inside their Shadow DOM.
- Keep finished or waiting subagents from becoming falsely active after a parent interaction. A genuine new child turn restores the running state; duplicate terminal events are suppressed.
- Fit subagent avatars within the available header width. Working children have animated circular borders; an ellipsis exposes the hidden children and their states. Reduced-motion preferences are respected.
- Automatically open a new Office document only when a live assistant reply in the exact current chat explicitly references it. Shared-workspace file events alone do not open unrelated spreadsheets. Existing files, historical messages and hidden content do not authorize opening; navigation cancels pending actions.
Source verification
- The required AionUI
just pushgate passed strict lint, formatting, TypeScript, translations and 3,511 tests, with five skipped, across 420 passing test files and one skipped file on Node 24.11.0 and Bun 1.3.9. Published source:6a5214044e2900021f7a2dd4cde82b76bdbbf8e3. - The required AionCore
just pushgate passed migration checks, lint, formatting and 8,443 tests, with 26 skipped. Published source:388e07f99e7f7c5c6fe1eb70b505386079a9b181. - Three independent reviews examined the diff, boundary cases and acceptance criteria. Findings about root-directory paths, files arriving during the initial folder scan and duplicate child terminal events were repaired. Subsequent focused checks passed 35 Office/Markdown tests and eight Core child-session tests.
- Earlier focused checks passed 57 composer tests, seven pet-window tests, 18 avatar/sidebar tests and 14 Core lifecycle tests. TypeScript, lint with zero errors, translations and the Electron build passed.
- Browser checks used the actual production components and styles with service/context mocks. Both composers passed slash/caret/input-method/highlight scenarios. BMAD link actions, keyboard focus and Light/Dark contrast passed. Avatar rendering at three widths stayed within its bounds, exposed the hidden list and respected reduced motion.
All 44 root release/helper contracts passed, and an independent review of the version metadata and release notes found no actionable issue. Hosted build and distribution results will be recorded after completion.
Runtime verification boundaries
An isolated Electron application received a real filesystem event for an unrelated XLSX and correctly kept its preview closed. The positive real-agent scenario stopped before document creation because its test backend reported codex runtime unavailable: bundled codex 0.162.1 is unavailable; automatic opening of a real generated file is not claimed as accepted. Current-chat positive cases and initialization/navigation races pass focused regressions. Native DOCX/PPTX generation was not run.
An isolated pet fixture created visible, nonfocused windows. Physical dragging while another application is foreground and the Dock interaction remain unverified because native automation was unavailable. Component and unit checks do not establish installed acceptance.
The working application, profile, chats and tasks were preserved during source verification. Installation of this release will take place only when agent turns are idle and after downloaded macOS bytes, signature, notarization and Gatekeeper checks pass. Native Windows runtime and upgrade acceptance remain unverified.
Distribution policy
macOS must retain the expected Developer ID Application identity and team, valid app/helper signatures and designated requirements, a stapled notarization ticket and Gatekeeper acceptance. Windows remains unsigned. publish controls publication only; it does not weaken verification. The release date must match the actual publication day in Europe/Kyiv. Build metadata and image manifests record the exact root revision and maintained-fork revisions; no earlier release tag is moved or overwritten.
eLLMsha 2026.10.14-patch.10
v2026.10.14 Patch 10 — release automation and installer downloads
Published on 2026-10-14 at 08:41:25 UTC (11:41 Europe/Kyiv), from exact tag revision c672615577a72013d45acc24d64622b5d1195157. The twelfth hosted run passed signed/notarized macOS and unsigned Windows packaging, all four Linux/amd64 image builds and isolated smoke checks, full-sized image archive handoff, and GitHub/GHCR publication. Independent checks of all eight downloaded assets, macOS signature/notarization and anonymous pulls of all four GHCR images passed. Release artifact and distribution acceptance is complete. Installed native updater and native Windows runtime/upgrade acceptance remain unverified. Earlier CI results below retain their original candidate dates and provenance.
Released changes
- Build native macOS Apple Silicon and Windows x86_64 installers from the root repository's exact maintained-fork revisions.
- Build the Linux VPS stack as versioned GHCR images with a release deployment compose file.
- Check the latest published eLLMsha GitHub release from the desktop and show a download action beside the sidebar toggle.
- Download and verify the complete platform installer, then open its DMG or EXE through the operating system. Opening an installer does not automatically terminate running agent tasks or replace the application.
- Keep signing keys in CI secrets and a temporary runner keychain. Every CI macOS build requires a Developer ID Application identity and notarization; Windows installers remain unsigned. Manual verification does not publish a release.
Acceptance status
Final public artifact and distribution checks: PASS. All eight uploaded assets were downloaded independently and matched their GitHub API sizes and SHA-256 digests. Individual installer hashes and build metadata also passed verifyReleaseArtifacts against the exact release root and its nine gitlinks. Both app bundles extracted from the public DMG and ZIP passed the expected Developer ID Application identity/team and app/helper designated requirements, strict deep signature verification, stapled-ticket validation and Gatekeeper assessment (source=Notarized Developer ID). DMG integrity verification passed and the image was detached. Anonymous pulls of all four GHCR images passed, including version/SHA/latest digest consistency, Linux/amd64 platform and OCI labels.
These are the verified individual installer checksums, distinct from the historical Actions archive digests below:
| Public installer | Bytes | SHA-256 |
|---|---|---|
eLLMsha-2026.10.14-patch.10-mac-arm64.dmg |
804,062,017 | 62c1fce2876ec91632f2625129e1f45233a17494701d6312a223ca3ecbbcecf6 |
eLLMsha-2026.10.14-patch.10-mac-arm64.zip |
840,263,790 | ad3fd9fd5f86dcdda9b0ec984a75adadf5015660ac333e774ea8f505d130ae69 |
eLLMsha-2026.10.14-patch.10-win-x64.exe |
474,215,284 | d87a1498643f1afabaabe812350360d20a8d05e89a98051dc570d5e42a824f27 |
Deferred runtime checks: the preserved Patch 9 installation lacks Patch 10's ReleaseInstaller and still uses legacy quitAndInstall. Real acceptance of the new downloader requires an installed Patch 10 and a genuinely newer published release because same-version downloads are rejected; a mock feed does not establish that result. Native Windows installation/runtime/upgrade QA remains unverified. These distribution checks did not launch or replace the production app, restart tasks or change its profile.
The following build and test records retain their original historical scope.
The required AionUI publication gate passed 3,448 tests (5 skipped), strict lint, formatting, TypeScript and translations. Focused updater checks passed 64/64; release pipeline helper checks passed 17/17 and workflow validation passed. Light/Dark renderer fixture checks used mocked IPC and did not open an actual installer. All four Linux/amd64 images, including the full web/Core/RS image, passed hosted builds and isolated smoke checks in the eleventh run. The seventh through tenth runs passed Developer ID signed/notarized macOS packaging and unsigned Windows NSIS installer creation, verification and artifact upload; the tenth run was subsequently canceled. Installed native updater behavior and native Windows installation/runtime/upgrade QA remain unverified. Independent reviews repaired a cancellation race and a stale latest-image promotion path.
A dedicated Developer ID Application certificate was issued and its key match verified. All six signing/notarization secret names are present in the GitHub release environment, restricted to main and release tags. Actual signed/notarized macOS CI acceptance passed in the seventh run below. Windows installer creation does not establish Windows runtime or upgrade acceptance.
The first hosted verification run found two packaging-environment defects: Windows incorrectly required an npm .cmd shim despite using Bun, and macOS reached Node's approximately 2 GiB heap limit during Chrome MCP bundling after Rust release compilation succeeded. The Windows prerequisite repair was reviewed and pushed. Desktop CI now gives Node a bounded 4096 MiB heap; an isolated Chrome MCP bundle using the same Node 24.11.0 completed successfully, focused contracts passed, and independent review found no blocker. Failed packaging also retains non-secret dependency/Rust caches for retries. The second run was deliberately stopped before repeating the known macOS failure. Failed and canceled evidence is retained; neither run is package acceptance.
The version date matches the actual publication date in Europe/Kyiv, following the project's release-date rule.
The third hosted verification run built and verified the unsigned macOS arm64 installers and uploaded the artifact at source 45036f3. That unsigned run did not establish Developer ID/notarization or installed-app acceptance. Windows passed Chrome MCP bundling, then found a separate Unix-style electron-vite shell invocation rejected by Windows cmd. Cache retention after this failure succeeded. The launcher repair uses checkout-local JS entries with Node and separate arguments for both Vite and builder, including retry paths. It passed 12 root contracts, 16 UI build-script tests and independent review. Windows installer acceptance was still pending after that run; it subsequently passed in the fifth run below.
The final launcher publication gate passed 3,450 UI tests (5 skipped), lint, formatting, types and translations. The complete root release/helper contracts passed 25/25. The retry packaging behavior contract and build-script suite passed 19/19, followed by independent review. AionUI source 8bd8228 is published; these source checks do not replace hosted installer acceptance.
The fourth verification run was canceled when the maintainer clarified the required single-pipeline policy: macOS always signed/notarized, Windows unsigned. Signing is now independent of publish; a missing Apple secret or failed signing/notarization stops verification too. The changed pipeline/helper contracts passed 28/28 and Actionlint passed. These checks establish configuration behavior, not completed Apple notarization.
The fifth verification run completed at source 021315b:
- The Windows job passed NSIS installer creation, installer verification and artifact upload. This establishes hosted packaging acceptance for the unsigned Windows x86_64 installer; native Windows installation, runtime and upgrade QA remain unverified.
- The macOS job successfully prepared the signing keychain and notarization credentials. Electron-builder 26.15.2 then rejected the configured signing identity's full
Developer ID Application:prefix. The generic DMG retry packaged an unsigned app, and stapled-ticket validation correctly failed with exit 65 because no notarization ticket existed. No accepted macOS artifact was uploaded by this run; Developer ID signing and notarization acceptance were still pending at that point. - The image and publication jobs were skipped. This run published neither a GitHub release nor GHCR images.
The signing repair preserves the full certificate name for native validation while stripping the builder-forbidden prefix only from its selector. Prepackaged recovery now validates the app/helper certificate, team and designated requirements, plus the notarization ticket and Gatekeeper for public builds, before repacking. Early configuration validation rejects partial/hash selectors with a clear error. The complete root contracts passed 30/30, targeted UI packaging tests passed 24/24, the actual electron-builder 26.15.2 selector contract passed using synthetic identity metadata, and independent review passed. The required AionUI just push gate passed 3,455 tests / 5 skipped, lint, formatting, types and translations; source a13cf1d is pushed. These source checks did not establish signed/notarized hosted acceptance; the sixth run below exposed the next blocker.
The sixth hosted run, at source 3f14ed5, selected the Developer ID identity and began macOS signing. It then failed with EMFILE (too many open files) while scanning the bundled Tencent dependency tree. The strengthened prepackaged-recovery gate blocked retrying that incomplete result; that run did not est...
eLLMsha v2026.10.13 Patch 7 — Reliable tools & browser control
More reliable tools, more predictable browser actions
Patch 7 includes the previously unpublished Patch 6 improvements.
- Browser control: exact tab targeting, reliable input focus, independent same-URL tabs and explicit errors for closed targets. Background application state no longer causes a false focus failure.
- Dynamic MCP guidance: relevant server documentation accompanies discovered tools; paginated catalogs and changed schemas refresh safely, retaining the last complete catalog after a failed refresh.
- Task continuity: a conversation-local working brief tracks user constraints, corrections and evidence. Guidance distinguishes current architecture checks through CodeGraph/source from historical memory and chat history.
- Model recovery: one bounded retry after an empty response keeps tools available; repeated empty responses report an incomplete task. Orphan tool events now terminate correctly.
- Integration fixes: safer automatic titles, supported Linear image-markup normalization, native empty-field clearing, and isolated evaluation sessions.
Download & install
For macOS Apple Silicon: download the DMG below and replace eLLMsha in Applications. Existing chats, memory and account connections remain in the persistent application profile. ZIP and SHA-256 checksums are also provided.
Ad-hoc signed; not Developer ID signed or notarized. Windows/Linux binaries are not included.
Verified scope
Installed, bounded GPT/Qwen evaluations used independent synthetic fixtures, with actual tool results and native readback. Browser input/lifecycle, MCP guidance/catalog refresh, the synthetic Linear adapter and native field clearing passed. An unchanged Computer Use helper retained access after a cold post-update launch. Automated regression gates and independent reviews also passed.
These assisted checks do not guarantee autonomous model accuracy or eliminate every regression. Tool/config isolation is not an OS sandbox. Future helper identity changes may require renewed macOS permission.
Component patches: AionUI 7 · AionCore 7 · AionRS 4 · Chrome DevTools MCP 3 · Open Computer Use 1.
Full changelog · Verification scope
The installed evaluation was performed on the October 12 candidate. The final October 13 build updates release-date metadata and consumes the committed fork pins. Packaging/signature checks passed; backend and Computer Use helper binaries match the evaluated installation. The final dated build has not replaced that installation.
DMG SHA-256
0885994dba37063c41984052050e3e58ee6f9d155cea1528ecd339c1cfdfa297
eLLMsha v2026.10.11 Patch 5 — Better recall, code graph on demand
Better recall, clearer evidence
- Scoped memory retrieval: project/global boundaries, project aliases, exact IDs, dated latest-task evidence, and explicit pagination/completeness.
- Code graph on demand: agents can inspect, allocate and release graph runtimes; persistent indexes survive release and reload.
- Honest tool results: preserve partial-failure warnings and source caveats, propagate MCP errors, and reject malformed tool arguments.
- Current architecture guidance: inspect the repository/code graph for current implementation facts; use memory for historical decisions.
Fixes
Context searches now use trusted project scope, invalid aliases are filtered, and retrieval health metadata reaches explicit search/context callers. Chat-history requests route to scoped history tools.
Download & install
Download eLLMsha-2026.10.11-patch.5-mac-arm64.dmg below and replace eLLMsha in Applications. Existing chats, memory and account connections use the persistent eLLMsha Desktop profile.
For macOS Apple Silicon. Ad-hoc signed; not Developer ID signed or notarized. Windows/Linux binaries are not included.
Verification boundaries
Bounded checks covered scoped retrieval, complete pagination, chat-history isolation, graph lifecycle/persistence and fresh GPT/Qwen recall. Both models recovered the main latest-task fact; Qwen still omitted caveats and made a factual-detail error. Model quality and latency remain different. Current-architecture tool use is guidance, not a mandatory tool-call gate. Full memory reindexing is not an agent tool in this release. Claude inference remains unverified.
Component patches: AionUI 5 · AionCore 5 · AionRS 2 · TencentDB Agent Memory 3. Patch 4 was an intermediate local build, included here.
Full changelog · Verification ledger
The version date follows the release date in Europe/Kiev. Original root material is AGPLv3 with a separate commercial licensing option; upstream licenses remain in force.
DMG SHA-256
504c16416309b0a512366fa8f4be933058de487d11adac36a06e6fc58df37ccf
eLLMsha v2026.10.10 Patch 3 — GPT-6 & managed runtimes
Highlights
- Codex CLI 0.162.1 and live model discovery, including GPT-6. Actual GPT-6 Astra inference and project-memory access verified in the installed macOS app.
- Managed / System / Custom CLI selection with resolved path/version, independent managed updates, active-session pinning and rollback when a previous managed version exists.
Fixes
- Fixed stale model catalogs rejecting available models before live validation; added bounded catalog pagination.
- Moved harness-managed Codex skills outside user repositories while preserving discovery and isolation.
- Fixed permission-profile setup and resumed Tencent memory-tool compatibility.
- Verified network access in Auto and Read Only modes while retaining filesystem restrictions.
- Improved runtime button/loading consistency and Light/Dark contrast.
Download & install
Download eLLMsha-2026.10.10-patch.3-mac-arm64.dmg below and replace eLLMsha in Applications. Existing chats and memory use the persistent eLLMsha Desktop profile.
This build targets macOS Apple Silicon. It is ad-hoc signed, not Developer ID signed or notarized. Windows/Linux binaries are not included.
Verification boundaries
Live GPT-6 Astra, memory reads, resumed network access, denied writes in Read Only mode, runtime source selection and native Light/Dark UI were verified. Cross-version managed rollback has focused test/review coverage but was not exercised live. Claude inference was not tested.
Version policy
Release versions now use the actual release date in Europe/Kiev: 2026.10.10, Patch 3. Historical tags remain unchanged.
DMG SHA-256
4490e139d4c829d39a755d8a8fb6079564ce7add15d7bbf53157c6d2814c9197
eLLMsha v2026.10.9 Patch 1 — Your agents. Your workspace.
Meet eLLMsha 🚀
Your AI agents. Your models. Your workspace.
eLLMsha — extensible Large Language Model smart harness architecture — brings cloud and local coding agents into one desktop workspace. This is our first public release.
What’s inside
- 🧠 Persistent project memory and code indexes across multiple repositories.
- 🤝 Visible agent and subagent workflows with read-only child histories.
- 🔌 Plugins, MCP tools, and independent Linear / Clockify accounts.
- 🌐 An in-app browser with agent-controlled tabs.
- 📱 Mobile automation for iOS and Android, plus macOS Computer Use.
- 🏠 Local desktop data, with documented Docker Compose deployment for self-hosting.
Download & install
Download eLLMsha-2026.10.9-patch.1-mac-arm64.dmg below, open it, and move eLLMsha to Applications. This binary is for macOS on Apple Silicon. Windows and Linux binaries are not included in this release.
Signing: this build is ad-hoc signed, not Developer ID signed or notarized. macOS may require manual approval to open it.
Existing eLLMsha Desktop chats and memory use the persistent Application Support profile.
Verification & scope
The macOS app, real profile, browser workflows, Mobile MCP device enumeration, and a harmless Computer Use interaction were exercised. Clockify was checked with one account and an empty report; populated reports and live multi-account Clockify remain unverified. Live Claude inference was not tested.
The source tag also includes clean-clone build preparation fixes and documentation completed after this DMG was built; those changes do not alter the desktop application implementation.
DMG SHA-256
5e7d943448b35b4d67ae5203560019ea98d81aa626dbb586d9fef5b82e3497cb
Build, explore, contribute
git clone --recurse-submodules https://github.com/ShiftHackZ/ellmsha.gitSee the README for architecture, screenshots and workflows, and release status for validation details.
Original eLLMsha code is AGPL-3.0-only, with a separately negotiated commercial licensing option. Upstream components retain their own licenses. See licensing.
Built by Moroz Inc. If eLLMsha helps your workflow, star the repository and share what you build.