Releases: tronschell/shinbo
Release list
v0.8.1
Download Shinbo for macOS (DMG)
Open the disk image and drag Shinbo onto Applications. The ZIP asset supports automatic updates.
Features
- polish desktop navigation and computer control for v0.8.1 (#70) by @tronschell
- Add desktop content minimaps, dedicated Changes and Git review panes,
and collapsible unchanged diff context. - Show nested task hierarchies and configurable tab colors with quieter
activity indicators. - Let Stop and Escape revoke computer control while the agent continues,
and remove the 20-call limit while retaining the ten-minute timeout. - Avoid unavailable advisor calls and unnecessary delegation, and omit
generated build directories from folder context. - Refresh documentation and remove obsolete audit reports and tracked
Python caches. - These new navigation and appearance surfaces are desktop-only; the
phone bridge is unchanged.
- Add desktop content minimaps, dedicated Changes and Git review panes,
Full changelog: v0.8.0…dd0c0e7
v0.8.0
Download Shinbo for macOS (DMG)
Open the disk image and drag Shinbo onto Applications. The ZIP asset supports automatic updates.
Breaking changes
-
rename Emma to Shinbo and ship the 0.8.0 performance and bug-audit work (#68) by @tronschell
- The app, CLI, data directory, environment variables, and all bundled
skills and tools are now named Shinbo instead of Emma
(SHINBO_DATA_DIR,SHINBO_FAST_BUILD,shinbo-cli,Shinbo.app).
Existing Emma data directories, profiles,EMMA_DATA_DIR, and renderer
settings are picked up automatically. - New local model brands: llama.cpp, LM Studio, Ollama, and oMLX.
- Runtime and renderer performance work across threads, timelines,
markdown, attachments, workspace output, the phone bridge, and the host
library, with regression tests for each. - Bug-audit fixes for cancellation, voice teardown, secret prompts,
scheduling, storage recovery, and council/workflow runtimes. - Reasoning details are surfaced from the OpenAI-compatible gateway.
- Imported skill names mentioned in messages are highlighted and open
their SKILL.md in the preview, and the preview dialog header puts the
title first with its controls grouped on the right.
Breaking change: the app, data directory, and environment variables are
renamed from Emma to Shinbo. - The app, CLI, data directory, environment variables, and all bundled
Full changelog: v0.7.1…a4b1fa4
v0.7.1
Open the disk image and drag Emma onto Applications. The ZIP asset supports automatic updates.
Fixes
- improve thread and workflow behavior for v0.7.1 (#66) by @tronschell
- Preserve per-thread model and thinking selections across restarts and
workflow runs. - Show clearer model and tool activity and expandable context-compaction
summaries. - Simplify workflow editing with collapsible schedules and a graph
inspector that preserves unsaved edits. - Add context-bar widget colors and restore new-thread creation in
Priority view. - Improve inline visual layout, actions, loading states, and error
feedback. - Clear inherited vision routing when no separate vision model is
configured.
- Preserve per-thread model and thinking selections across restarts and
Full changelog: v0.7.0…5a544de
v0.7.0
Open the disk image and drag Emma onto Applications. The ZIP asset supports automatic updates.
Features
- ship context handoffs and performance improvements in v0.7.0 (#64) by @tronschell
- Add optional Fresh context handoffs and checkpoint prompts before
context rollover, with a bounded recovery record when no handoff is
supplied. - Improve thread serialization, terminal scrollback retention, machine
probes, lazy search setup, and renderer polling and loading. - Harden goal validation, blocked-goal accounting, local IPC URLs, and
Windows file URI and LSP handling. - Limit second-model fallback requests to three models and preserve
Windows transport retry handling. - Trim Windows package locales and retain bundled search license
notices. - Select CI checks from the changed files, exercise packaging changes
before promotion, and provide a direct macOS DMG link in release notes.
- Add optional Fresh context handoffs and checkpoint prompts before
Full changelog: v0.6.0…24d4f1f
v0.6.0
Features
- launch apps and see the ones Emma started (#58) by @tronschell
computergains alaunch_appaction. Give it an app's name —
Notepad,Calculator,Google Chrome, or the namelist_appsprints
— and Emma asks the user once, starts the installed app, returns its
identity and PID, and grants control of that instance for the turn. It
never takes a file path and resolves only against installed
applications.- Apps that Emma itself started are visible again. Both native helpers
now hide only Emma's own binaries — the helper, the process behind
--blocked-pid, theemma,emma-host,emma-cli,emma-pty,
emma-computer,emma-option-tapandemma-transcribeexecutables,
and anything under the helper's or Emma's own directory. Process
ancestry is no longer consulted, so an app opened from Emma's shell or
bylaunch_appappears inlist_appsand can be read and controlled. - The
terminaltool now says not to start graphical apps from the
shell: on Windows its job object kills the command's whole tree when the
call returns, so a GUI app opened there dies immediately. Use
launch_app. The job also carriesJOB_OBJECT_LIMIT_BREAKAWAY_OK, so a
child that deliberately asks to break away can. docs/computer-use.mdanddocs/tools.mdcover the new action, the
resolution order on each platform, and the exclusion rule.
Fixes
-
keep macOS-only window calls off Windows and let approvals lapse gracefully (#61) by @tronschell
- Windows computer use no longer crashes on its first app read. The
activity cursor overlay was calling a macOS-only window method, which
killed everyget_app_stateon Windows withsetHiddenInMissionControl is not a function. - Models can find
launch_appagain. The tool description the harness
advertises had never been updated for it and still told the model to ask
you to open an app it could not see, so models started apps from the
terminal instead — where the shell kills them the moment the command
returns. - An app approval you did not get to in time is no longer treated as a
refusal. Emma tells the model there is no answer yet and lets it ask
once more; only Don't, Escape or a stop bans an app for the rest of
the turn.
🤖 Generated with Claude Code
- Windows computer use no longer crashes on its first app read. The
Other changes
-
release 0.4.0 (#30) by @tronschell
-
install and launch both packages on the runners (#54) by @tronschell
- CI now installs and launches both the macOS disk image and the Windows
Squirrel installer on the runners, drives the installed app over the
DevTools protocol to prove the workspace renders, uploads a screenshot
of it, and uninstalls the Windows build again, so a broken package fails
the build instead of reaching a release. ciaccepts a manual dispatch with apackageinput that runs the
macOS and Windows packaging jobs, with the install smoke test, against
any branch.
🤖 Generated with Claude Code
- CI now installs and launches both the macOS disk image and the Windows
-
bump to 0.6.0 (#59) by @tronschell
Emma 0.6.0 ships Windows x64 alongside macOS: a Squirrel installer with
auto-update, PowerShell as the shell the model targets, computer use
with app launching, durable terminal sessions, a taskbar and installer
identity of its own, and zvec-grep downloaded on demand instead of
bundled, which shrinks the Windows installer from 570 MB to 143 MB.
macOS is unchanged in behavior and gains the same on-demand zvec-grep
and the credential-store hardening.🤖 Generated with Claude Code
Full changelog: zvec-grep-v0.2.1…8904954
zvec-grep 0.2.1
Prebuilt @zvec/zvec-grep 0.2.1 trees Emma downloads on demand when semantic search is turned on.
v0.5.0
Features
- ship harness handoffs and desktop workflow refresh (#47) by @tronschell
- Add cross-harness result handoffs, per-run model and thinking
controls, and workspace navigation for active CLI work. - Add configurable web-search providers with free fallbacks, provider
branding, bounded caching, and private-network guards. - Refresh setup, archive, settings, sidebar, context, model-plan, and
pull-request status interfaces. - Replace the macOS installer flow with a verified drag-to-Applications
DMG and repair vendored zvec-grep symlinks for signing. - Remove the retired research and autoresearch paths while simplifying
durable host and thread storage.
- Add cross-harness result handoffs, per-run model and thinking
Full changelog: v0.4.4…14b152e
v0.4.4
Features
- finish Electron desktop and self-improvement migration (#43) by @tronschell
- Complete the Electron desktop surface and retire the unused GPUI app and marketing artifacts.
- Add the self-improvement bench, model board, trace ledger, and verification coverage.
- Harden harness routing, phone bridge boundaries, updater state, and macOS packaging workflows.
- Bound loopback harness cleanup paths and stabilize native Windows helpers.
Full changelog: v0.4.3…56a6ad6
v0.4.3
What's Changed
- fix: prevent stale app launches by @tronschell in #40
- chore: release 0.4.3 by @tronschell in #41
- release: promote dev to main for 0.4.3 by @tronschell in #42
Full Changelog: v0.4.2...v0.4.3
v0.4.2
What's Changed
- chore: remove dead code by @tronschell in #37
- chore(dev): release 0.4.2 by @tronschell in #38
- release: promote dev to main for 0.4.2 by @tronschell in #39
Full Changelog: v0.4.1...v0.4.2