Releases: boggspa/TaskWraith
Releases · boggspa/TaskWraith
TaskWraith 1.6.8
1.6.8 - 2026-06-28
Added
- Scheduled messages in the visible queue. The composer Schedule clock now
creates a timer-locked queued message instead of a separate hidden scheduled
task. Scheduled rows keep Edit/Delete, show a live countdown where Steer
usually appears, and dispatch automatically when due. - Larger composer uploads. Drag/drop, paste, and attachment picker flows now
keep up to 15 composer attachments. - Detected Tailscale relay setup. The iOS bridge settings now show this
Mac's detectedwss://Tailscale relay door with Use this, Copy, and Test
actions so cellular pairing no longer depends on users hand-entering the
MagicDNS URL and port.
Changed
- Cleaner trust-mode UX. YOLO / Trust This Session now appears as a compact
composer chip with an explanatory tooltip and click-to-disable behavior,
replacing the warning-style banner. - Clearer approval affordances. Permission, elevation, unattended workflow,
Ollama parity, and provider sign-in actions now have tooltips that explain the
consequence and lifetime of each choice. - Softer visual polish. Sky visual effects have softer orb edges, and image
attachments show better placeholder behavior while previews load. - Message actions moved into timestamp footers. Transcript copy/delete
actions now sit in the message footer area instead of floating over the
message body.
Fixed
- Share invite reliability. Shared-chat invite copy and relay setup are more
robust, including stronger validation around collaborator join payloads. - iOS relay pairing fallback. Tailscale and relay pairing paths recover more
reliably when the initially advertised route is not usable, and release builds
honor the configured Tailscale Serve front-door port during startup and
self-heal. - Desktop polish fixes. Helper subprocesses no longer appear as extra app
icons, spellcheck context menus behave correctly in editable fields, thread
rename persistence is more reliable, and the sidebar rename editor is harder
to trip up.
TaskWraith 1.6.7
1.6.7 - 2026-06-28
Added
- Universal composer scheduling. Workspace-backed single and Ensemble chats now have a Schedule clock in the composer controls. It opens a glass date/time picker with quick presets and uses the existing scheduled-task pipeline so delayed prompts stay restart-safe and run through the same unattended-authority guardrails.
Changed
- Scheduled task visibility. Scheduled-task pills now show a live countdown and switch to "due / waiting" when the timer has elapsed but the chat is still busy.
- More flexible iOS relay pairing. Device pairing can include a manually configured relay door alongside LAN/Tailscale discovery, making release builds less likely to get stuck on the wrong network path.
Fixed
- Sidebar rename reliability. Inline chat rename fields focus and select consistently, rename can be started deliberately from the row or menu, and renamed titles avoid unnecessary truncation in edit fields.
- Scheduled-task validation. The main process now rejects invalid or already elapsed schedule times before saving delayed work.
Release Verification
- macOS universal app notarized and stapled.
- iOS TestFlight build 56 uploaded successfully.
TaskWraith 1.6.6
TaskWraith 1.6.6
Added
- Decomposition groundwork. App and main-process orchestration have started moving out of the large root files into focused helper, hook, and IPC modules, making future changes easier to review without changing user-facing behavior.
- Sidebar workspace path actions. Workspace path controls are easier to reach from the sidebar during local project work.
Changed
- More robust iOS remote access. Dev and release builds recover Tailscale / relay routing more reliably, retry transient status probes, prefer recently successful relay doors on reconnect, and restart the embedded bridge on demand when pairing needs it.
- Smoother desktop transcript streaming. Electron transcript reveal and scroll ownership now stay active during provider output so streamed text feels less chunky and long-running replies do not fight the user's scroll position.
- Navigable app notices. Welcome / first-launch notification cards can now be dismissed permanently and moved through with explicit carousel controls.
Fixed
- Ensemble orchestration reliability. Role and Bossman mention routing now takes priority over ambiguous provider tags, stale round lifecycle state recovers cleanly, and shared-chat invite joins are harder to trip up.
- Cursor MCP bridge compatibility. Cursor and compatible runtimes can use the brokered TaskWraith MCP tool names that include hyphens while reserved TaskWraith tool namespaces remain protected from repo-provided collisions.
- Security hardening. Release 1.6.6 tightens agent trust boundaries across approval actions, Cursor MCP allowlisting, favicon fetching, external path grants, raw provider event persistence, PTY session ownership, git IPC scope, transcript local-link opening, relay connection caps, and encrypted relay resume/ACK handling. Retired Gemini OAuth profile material is purged instead of being recreated.
Verification
npm run validate:releasepassed all required gates.npm run build:mac:notarizedcompleted successfully.- macOS app and DMG are Developer ID signed, notarized, stapled, and Gatekeeper accepted.
npm run validate:mac-update-feedpassed after final DMG stapling.- Packaged Electron smoke and launch smoke passed.
TaskWraith 1.6.5
1.6.5 - 2026-06-27
Added
- User MCP server manager. Settings can now manage user-defined MCP servers,
import/export provider snippets, show per-server readiness, and copy
provider-specific config for Codex, Claude, Cursor, and compatible JSON/TOML
targets. Remote server URLs, bearer headers, naming collisions, and invalid
exports are validated before they reach provider runtimes. - iOS diff review sheet. The companion's files-changed / diff summary rows
now open a dedicated glass diff sheet for quick review from the composer area. - Shared chats on iOS. Shared-human chat projections and the shared-chat type
chooser now appear on the companion, with sidebar actions restored on desktop. - Transcript user gutter. User messages gain a compact gutter/jump affordance
so long transcripts can move back to the originating request more reliably.
Fixed
- iOS thread rename parity. Chat renames from the iOS companion sync back to
the Mac, full titles are preserved across the bridge, and headers/sidebar rows
avoid unnecessary truncation in rename fields and wide-enough title areas. - iOS Bossman removal. Roster updates from the iOS companion now distinguish
old clients that omit the Bossman marker from current clients explicitly sending
isBossman: false, so turning Bossman off clears the Mac-side assignment and
removes the crown instead of preserving the previous participant. - Transcript jump stability. Message jump targets now converge more reliably
through virtualized transcript rendering and stay anchored when navigating to a
specific user message. - Transcript table readability. Wide Markdown tables wrap and align more
consistently across Electron and iOS transcripts. - Claude picker availability. Fable remains visible but disabled while
unavailable, and reasoning options such as Extra / Ultracode stay visible but
unselectable on models that do not support them.
Changed
- MCP setup copy is clearer. Provider-tool setup now labels active MCP
servers, routes settings searches to the right config panes, and explains the
Cursor/Grok bridge fallback without implying retired Gemini setup work. - Media previews are richer. Electron previews and inline audio waveform clips
gained focused preview affordances and cleaner transcript presentation.
TaskWraith v1.6.4
Added
- Composer-aligned slash picker. Slash commands now open in a wider,
composer-width picker with grouped sections, richer command descriptions, and
custom monoline icons for review, planning, side-chat, ensemble, Gemini,
workflow, settings, and utility commands. - More slash command coverage. Multiview side panes, ensemble controls,
workflow helpers, Gemini passthrough commands, prompt templates, model/context
tools, and settings shortcuts can be run from typed slash commands or picker
selection.
Changed
- Consistent theme opacity. Every named system theme now obeys Settings ->
Appearance -> Main pane opacity, instead of only Light, Dark, Alabaster, and
Obsidian honoring the slider. - More readable popovers. The Settings menu, Approvals, Shares, and Devices
mini pickers now use a 75% background material so their contents stay legible
across glass-heavy themes. - iOS display scaling. The companion gained a display-size scaling control for
tuning the remote UI density on-device.
Fixed
- Slash routing hardening. Typed slash commands now route through the shared
registry, preserve drafts when a pane handler redirects, resolve participants
per focused pane, refresh Gemini command discovery after/commands reload, and
handle discovered commands with argument placeholders. - Share reconnect polish. Shared chat discovery and reconnect signals are more
resilient when a linked device or collaborator returns after a brief disconnect.
TaskWraith v1.6.3
1.6.3 - 2026-06-26
Added
- Shared chats. You can now invite other people into a chat. Sharing runs over
the same end-to-end-encrypted relay as the iOS companion: invitees join by
confirming a short verification code, see a live projection of the conversation,
and you stay in control from a People header with per-participant revoke and a
one-click "Stop sharing." A new Shared section in the sidebar, a "New Shared
Chat" entry in the + New menu, and a dedicated Shares settings tab manage the
whole lifecycle; collaborator comments are clearly marked as external/untrusted. - A bigger media toolkit for agents. Agents can now mix multitrack audio and
encode, overlay and concatenate video natively (VideoToolbox, no ffmpeg),
transcribe audio on-device, and inspect audio segments and video frames with an
interactive scrubber, an NLE-style filmstrip and a DAW-style waveform player —
all over the un-forgeable trusted-media channel. Producer outputs gain automatic
posters, badges, an expand-to-view viewer and Finder/copy/save actions, and any
player can be torn off into its own resizable pane. - Audio & video playback on iOS. The companion streams and plays audio and
video inline in the transcript over the encrypted link — chunked so it seeks and
plays without downloading the whole asset first — with posters and metadata. - Model context lengths. A new Context lengths view (in the sidebar Model Usage
overlay and a Settings table, on both desktop and iOS) shows each model's
official context window, and the composer context meter is now a clickable
popover with per-participant detail. - Sidebar control cluster. A traffic-light footer — Approvals, Shares and
Devices — with at-a-glance popovers for pending approvals, active shares, and
paired devices with a live-connected indicator. - Composer plan popover. A plan control in the composer for setting a run's
plan inline. - Trust guide & setup docs. A top-level trust-and-safety guide (safe first-run
path, capability matrix, local storage/reset notes, provider data boundaries,
release-verification commands and known limits), an advanced optional-setup
guide, a composer-variant gallery and refreshed README screenshots.
Changed
- Codex surfaces media. Codex runs now show their generated images and trusted
audio/video in the transcript, at parity with the other providers. - Looping workflows show live progress. A looping workflow reports per-iteration
progress as it runs, and its verifier can run on a different provider than the
maker. - Media tools are gated like shell and file access. Media editing is now its
own audited, approvable capability with read-only presets, rather than riding
along with general tool access. - Cursor/Grok bridge clarity. Write-capable Cursor and Grok runs use
TaskWraith's scoped MCP broker without manual provider-side config, and the
shared bridge copy no longer names retired Gemini as an active setup target.
Fixed
- No more repeated notification banner (iOS). The companion no longer raises a
banner every time it registers for push notifications. - iOS recovers instead of bricking. A keychain identity error (-34018) now
recovers the device identity instead of leaving remote unusable, plus sidebar
polish — single-line chat rows and the workspace utility icons unified into one
glass pill. - Collaboration hardening. Shared projections scrub secrets and collapse
sensitive paths, appends are rate-limited, display-name impersonation is
tightened, and shares are revoked when their chat is deleted. - Composer plan control. Stays visible and renders its participant plan lanes
correctly, and the side-chat composer gained the same parity controls.
TaskWraith v1.6.2
Added
- Refractive "liquid glass." An optional refractive-glass material that brings a
subtle liquid-glass refraction to the composer, sidebars, pickers, dashboard and
first-launch sheet in place of the flat frosted look. Toggle it under
Settings → Appearance — it's independent of the other visual effects. - Workflows can loop. A workflow can now run a maker → verifier → decide loop:
each iteration is independently judged (accept / revise / reject), with a durable
run history. Desktop shows per-iteration history in the sidebar; iOS shows a
"N× · accepted" loop-progress badge. - Audio, video & image tools for agents. Agents can play audio and video inline
in the transcript, run waveform-backed audio analysis and native single-frame
video decode, and generate images with a bring-your-own image-API-key UI — all
over an un-forgeable trusted media channel. - Bossman — ensemble manager. An opt-in manager that can reorder or replace
ensemble participants and drive a shared goal, gated by an auto-approval policy
with a full audit ledger.
Changed
- General chats are friendlier. "Global Chat" is now "General" across desktop and
iOS, with a less technical welcome and softer chrome for non-coding conversations. - App-icon picker (macOS). The Monoline and WWDC26 variants now render the correct
canonical artwork in the Dock, and the picker thumbnails are colour-managed
(8-bit sRGB). - Branded macOS installer. The DMG now uses a branded background from the brand kit.
Fixed
- The transcript stays where you put it. Scrolling up is no longer overridden by an
auto-snap to the bottom on new activity, and the streaming tail no longer flashes;
the side-chat panel gained the same scroll-up release (desktop). - No more dropped characters. Several per-keystroke text fields — ensemble role and
goal/brief, the roster preset name, audit budgets and the guest model id — no longer
drop or jumble characters as you type. - More reliable iOS ↔ Mac pairing. Dev and release builds use distinct serve ports,
dead relay doors are dropped from the QR code, and off-LAN pairing tries the right
door first with clearer, actionable error copy. - iOS polish. Smoother streaming with a reliable "Show more", a fixed transcript
scroll crash on ensemble send, and no more false "Mac busy or asleep" banner when a
request times out while still connected. - Empty "New Chat" tabs are tidied up. Never-used New Chat tombstones are reaped
automatically (drafts, ensembles and workflow-compose are left alone). - Grok remembers across turns. Grok's default transport re-injects cross-turn
context instead of starting fresh each turn. - Composer glass polish. The native-glass composer renders its exact shell variant
(no stale replica) with a smooth top edge.
TaskWraith 1.6.1
TaskWraith 1.6.1 — local-first macOS workbench for running and reviewing AI coding agents.
Added
- Composer drafts persist across switching threads, backgrounding, and an app restart (desktop + iOS).
- Per-participant context meter — the composer context donut is clickable, opening a context-meter popover; in ensembles it follows the focused participant and live-ticks the running one.
- iOS — rich completion notifications — a banner with the final message line and a coloured "N files · +A −B" diff summary, composed on-device over the encrypted link.
- iOS — inline @mention tinting in the composer (single-provider and guest chats too).
- Provider API rate limits in Model Usage.
Changed
- Smoother streaming — a matched type-out reveal plus render-coalescing and scroll/markdown perf retunes (desktop + iOS).
- Kimi defaults to K2.7 Code.
- Grok native goals via slash commands; Codex long-context estimates.
Fixed
- Ensemble completion alerts fire once per round, not per participant.
- Multiview diff stats are per-pane in the Create-PR row.
- Live usage meters — Codex reads live; no false 0% on stale windows.
- iOS transcript stability — fixed a crash scrolling during an ensemble send, reduced streaming flicker, kept the @mention hue while typing.
macOS universal build is notarized + stapled. Windows and Linux artifacts are added by CI.
TaskWraith v1.6.0
Added
- Multiview — split the workbench into up to four live panes. Open several
chats side by side over one shared environment, each with its own composer,
agent aura, and ambient FX; drag the dividers to resize. Welcome and split
states render correctly per pane. - Workflows are now a first-class chat type. A dedicated Workflows welcome
screen with its own compose controls and a "Run as ensemble" toggle where the
ensemble feature gate is enabled, plus a Workflows sidebar section. Scheduled
workflows recover after a restart, and workflows project to paired iOS devices
(read/view). - Reusable notification area. The welcome and first-launch screens now host
a dismissable notice card for significant changes — provider deprecations, new
models or providers, shipped features — rotating through more than one with the
same swipe effect as the activity heatmaps. - Hover labels on the composer footer icons. Screen Watch, Goal, Copy
transcript, and Multiview reveal a frosted glass label on hover/focus, so the
icon-only buttons are discoverable without taking up resting space. - Grok Composer 2.5 Fast. Grok now runs on Composer 2.5 Fast — the new
default — selectable from the Grok model picker. - iOS — offline Demo Mode. Explore the companion with no paired Mac:
interactive sample replies, a populated Inspector (Changes / Agents / Side
chats / Notes / Usage), an offline File Editor + Diff Studio, and demo chat
creation. Also new on iOS: a Workflows section, an iPad sidebar that opens
collapsed to headers, a live "still working" anchor during tool calls, and
attached images that render inline in the transcript. - Ollama local models — per-chat tool tiers, run profiles, and Cloud
sign-in. A two-pane composer picker sets how much tool access a local model
gets per chat (a four-step tier ladder, Tier 4 = full provider parity) with an
independent run profile; the run path and mid-run tool gates honor the per-chat
tier, and the chip warns when a Tier 4 selection can't take effect. You can
also sign in to Ollama's hosted Cloud models from the provider login terminal,
Settings → Providers, and onboarding. - TaskWraith Canvas — a live web surface agents can see and drive. A
read-only preview surface, then ref-based click / fill and Set-of-Mark
annotation, plus signed-elevated evaluation behind a dedicated, non-grantable
approval lockbox. Open it as a floating window or an embedded multiview pane
from the composer; read-only Canvas previews project to paired iPhones with
close / reload controls, and an iOS-simulator device driver can drive a
screenshot-only canvas. - Multi-host iOS pairing. One iPhone can pair with several Macs or PCs —
per-host session management and per-OS host glyphs — plus optional Tailscale
tailnet discovery and on-demand pairing. - Settings → Roster, plus a dedicated iOS Roster page. A full
ensemble-roster preset manager: per-participant provider, model, reasoning,
permissions, role, brief, and turn order, with orchestration controls and
cross-device preset load / save / delete on iOS. - App icon switcher. Choose between Regular, WWDC26, Monoline, and Glass app
icons from Settings → Appearance (the desktop dock) and the iOS Settings sheet. - Proposed plans. An agent can propose a structured plan that appears as an
inline card you approve before it implements — and the approve / respond /
dismiss round-trip works from paired iPhones too. - Cross-thread recall. An agent can resolve a vague reference to a past run
on another thread and read how far it got, through gated tools: zero-prompt for
your own workspace, gated across workspaces, allowlist-scoped from a phone, and
with verifiable citations. - Inline transcript media. Provider image outputs, tool image results, and
workspace images embedded in markdown become first-class transcript media with
full-size preview, and project to iOS. - Run and launch targets. Discover and run VS Code shell tasks and Xcode run
targets, and open previews straight from launch-output URLs, with launch
attempts surfaced in the run rail. - Inline questions + an iOS first-launch guide.
ask_user_questionrenders
as an inline transcript card on desktop and iOS, and a new iOS first-launch
guide mirrors the desktop onboarding with live provider readiness from your Mac.
Changed
- Gemini has been retired. Google ended the Gemini CLI sign-in, so Gemini is
no longer available for new runs and is removed from new-run and picker
surfaces, the sidebar usage meter, Settings (Providers + MCP), onboarding, and
new iOS run surfaces. Existing Gemini chats, labels/glyphs, transcripts, and
usage history are preserved; new chats default to Claude with sticky last-used. - Grok and Cursor are full first-class providers. The experimental
eligibility gate has been removed — both are accepted everywhere like the other
providers, write-capable and gated by approval mode. - Cleaner startup. The window stays masked until the renderer finishes
hydrating, so there is no flash of half-loaded UI. - Readability + performance polish. Slash / mention / context / palette
popovers are frosted; welcome heading sizing is consistent across composer
shells; the iOS Pair-with-Mac screen was rebuilt in TaskWraith chrome;
re-visiting a chat reuses its cached hydrated state to skip a redundant render;
and iOS streaming auto-follow is smoother. - Automatic provider failover (opt-in). When a provider hits a quota wall,
TaskWraith can pause and reroute the run to a healthy provider without
escalating its permissions. Off by default, behind a setting. - Settings, reorganised. A more discoverable settings sidebar and a safety
overview on desktop and iOS, and the Tailscale remote-access helper now sits
above Bridge networking on the pairing screen.
Fixed
- iOS image attachments work end to end. Attached photos now reach the agent
and render inline in the transcript, and the encoder no longer silently drops a
dense image. - iOS Grok composer keeps focus. The composer shell no longer flips its
layout — losing the tucked tabs and focus — when you tap in. - Multiview robustness. Per-pane composer parity and routing, a
reasoning-options crash, focused-state leaking into resting panes, per-pane
Screen Watch gating, and welcome cards/dashboard hiding when the view is split. - iOS polish. The roster editor no longer dismisses when you focus the role
field, the jump-to-latest pill stops sticking, long markdown list/quote items
wrap instead of truncating, and demo mode can't contaminate a real paired
session. - iOS no longer gets stuck "running" after backgrounding. A run that finishes
while the app is backgrounded now reconciles its streaming state on reconnect,
so the thread leaves the running state and shows the final message. - Read-only Grok seats can't hard-cancel a run. The Composer
Shelltool is
denied on read-only seats instead of cancelling the whole run. - Clearer activity diffs. Inline tool-diff add / delete lines read as red and
green through theme-aware tokens.
TaskWraith v1.5.9
TaskWraith v1.5.9
A same-day follow-up to v1.5.8 with two field-feedback fixes.
Changed
- iOS composer collapses to one line when idle. On the phone, when the composer isn't focused (keyboard down) it now collapses to a single-line input + model pill + send — the diff/changes rows, participant roster, queued prompts, and telemetry rail tuck away and reappear when you tap in.
Fixed
- Ensemble "Task Complete" summary covers the whole round. The end-of-round completion card now sums tokens and cost across every participant, unions their file changes, and spans the full round duration — previously it showed only the last participant's numbers, which badly understated long continuous ensembles.
macOS: notarized + stapled universal (.dmg / .zip + update feed). Windows/Linux installers are attached by CI shortly after publish.