TaskWraith 1.7.9

TaskWraith 1.7.9

Desktop release with iOS companion build 73. Highlights: iOS performance pass, liquid-glass sheets, transcript parity refinements, Settings 14→9 cleanup, ensemble participant ordering fix, and deterministic roster seeding.

Checksums (SHA-256)

• TaskWraith-1.7.9-universal-mac.dmg: 7a15fb0465ff9a81597f47241bfadaab0465f613b470c255098c596e109e0410
• TaskWraith-1.7.9-universal-mac.zip: a87f46be8ef32f47217d9f7ec1248005de598d88e0cfb38c8809e7ec2de4627a

See CHANGELOG.md for full notes.

TaskWraith 1.7.8

Fixed

• iOS transcript hydration. The Mac bridge now avoids sending oversized rich
  remote-projection frames that could exceed the relay's payload budget after
  the phone had already received the lightweight chat list. Oversized projection
  snapshots fan out through the existing single-envelope channel, and large
  thread snapshots are trimmed to a safe latest-row window instead of leaving
  mobile detail screens stuck on "Loading transcript from your Mac...".
• iOS thread detail refreshes. Mobile detail views now re-request transcript
  snapshots when a metadata-only snapshot says history exists but contains no
  rows, and Mac-side request handling avoids races where an early request could
  be ignored before the bridge connection was ready.
• iOS projection aliases. Remote thread snapshots now carry the aliases the
  Swift client expects, keeping transcript rows, ensemble metadata, and composer
  context attached to the selected chat.
• iOS compact composer. The mobile composer again collapses both the
  above-row controls and the footer telemetry rail when the keyboard drops,
  while preserving the compact diff access affordance.
• iOS Ollama brand parity. Participant health cards now repair generic
  Ollama display stamps with model-specific branding, so Laguna renders as
  Poolside on mobile like it does on desktop.
• Windows CI portability. Release/test scripts avoid POSIX-only assumptions
  that broke the Windows lane.

TaskWraith 1.7.7

Added

• Thread Introspection — memory promotion (read-only MVP). A retrospective
  layer scans recent threads and runs, harvests evidence (run events, approval
  friction, message feedback, correction heuristics), and produces reviewable
  Memory Proposal Packs with scoped, cited lessons (preferences, failure
  modes, repo conventions, provider hints, skill patches, bugs, do-not-repeat
  notes). Settings → Automation → Thread introspection supports a manual
  24h run, proposal review (approve/reject), evidence citations, and a daily
  Enable toggle for read-only scheduled packs. Distinct from per-run
  Evidence Packs and ensemble Blackboard. See THREAD_INTROSPECTION.md.
• Thread Introspection — apply phase 1 (repo conventions). After explicit
  approve in Settings, eligible repo_convention and do_not_repeat proposals
  can be applied to the workspace RepoConventionIndex via
  applyMemoryProposal (stable intro-{proposalId} entries, apply receipt on
  the proposal). Blocked in phase 1: skill/instruction file writes,
  skill_patch, bug, preference, provider_hint, and failure_mode.
  Skill patches remain review-only until a later Skill Patch Manager with diff
  preview and rollback. See THREAD_INTROSPECTION.md.
• Thread Introspection — MCP trigger and review tools. Agents can now use
  tw_introspection_run, tw_introspection_list, tw_introspection_read, and
  tw_introspection_review to create, inspect, and review proposal packs from
  chat. The MCP surface intentionally has no apply tool; applying approved
  proposals remains gated through the Settings/API flow.
• Thread Introspection — decay and supersede lifecycle. Store-level helpers
  can now supersede one memory proposal with another, expire past-due proposed
  items, and preserve bidirectional provenance links. Lifecycle controls are
  internal for now; no Settings, IPC, MCP, or automatic lifecycle policy is
  exposed yet.
• Ensemble roster import/export. Settings → Ensemble Roster can now save,
  import, and export full roster presets as portable JSON so teams can swap
  task-specific ensembles between installs. Imported/exported rosters preserve
  participant shape without carrying live Trusted Session authority.
• iOS Ensemble parity. The companion app can create/switch Ensemble chats via
  the same chat-kind bridge as desktop, repairs thread-list/task-card metadata
  when older Mac projections are sparse, and keeps ensemble status visible in
  the mobile shell.
• Boss controls. Bounded Boss control tools add quota reset summaries,
  timeout expiry, user-response tracking, participant summons, and MCP coverage
  for controlled multi-agent orchestration.
• Agent-pool and tool icon expansion. The agent-pool icon catalog now covers
  more workflow/tool families, including Diff Nice and release/CI-oriented
  glyphs.

Changed

• Trusted Session replaces ambiguous YOLO trust. Full local authority is now
  granted per chat or per ensemble participant lane, not globally to every
  participant in the session. Workspace Write is accented orange and Trusted
  Session/Full Access is accented red in the picker and composer chip.
• Release-class shell posture is narrower. Generic release shell commands no
  longer receive blanket release approval; explicit publish/signing receipts
  still carry their own authority.
• Live activity attribution is clearer. Tool-call and thinking-trace stacks
  show participant-style speaker headers, preserve model/reasoning boundaries,
  and avoid rendering stale historical run cards in the transcript path.
• Composer and workspace controls are denser and steadier. Git/worktree
  rows, utility popovers, goal/plan popovers, diff counters, external-path rows,
  and secondary workspace chrome were tightened for repeated operational use.
• Release packaging checks are more stable. Gate A packaging excludes local
  Cursor workspace state from the app bundle, reduces fixture IO, and keeps
  optional macOS dependency installation scoped to the universal build path.

Fixed

• iOS thread-list recovery. Task-card fallbacks, chatKind repair, and
  metadata snapshot merge keep companion lists aligned with desktop state after
  older or partial broadcasts.
• MCP bridge write hardening. Safe-write paths and tool alias classification
  better distinguish reasoning traces, shell/search aliases, and write-capable
  bridge operations.
• Grok/Cursor composer glow parity. Provider aura styling now matches the
  current composer shell for Grok and Cursor.

TaskWraith v1.7.6

TaskWraith v1.7.6

Release date: 2026-07-05
Tag: v1.7.6 @ 5973182fcfe3e87136a4e4251417aa59af45b5ab

TaskWraith is a local-first macOS desktop workbench for running and reviewing AI coding agents. Execution, history, and workspace state stay on your machine.

---

Highlights

Added

• Prompt caching guarantee tiers. Settings → Providers → Prompt caching shows per-transport posture (Guaranteed / Automatic / Best effort / Unsupported) with optional diagnostics when providers report cache read/creation tokens. Controllable API/BYOK paths honor policy modes (off / auto / explicit); opaque CLI transports remain best-effort only.
• Universal forks. /fork and thread fork actions use a provider capability summary: native fork on Codex (thread/fork), emulated fork (isolated sibling chat with duplicated transcript) on other live providers, with clear native vs emulated labeling.
• Worktree and branch orchestration. Runtime profile workspaceMode: worktree resolves to git worktree lifecycle at run launch. The composer above-row branch control opens a popover to list branches, create a branch, checkout, and create/select/remove worktrees with dirty-tree guards.
• Thinking trace actions. Reasoning traces in the activity stack gain copy-friendly actions consistent with other tool cards.
• Provider thinking summaries. Assistant turns can surface compact thinking summaries where the provider exposes reasoning metadata.

Changed

• Cache usage telemetry. Model Usage and prompt-cache diagnostics surface cache read/creation tokens when providers report them, including on implicit caching transports TaskWraith observes but does not control.
• Run summary diff stats. Diff stat chips align with the current evidence-pack presentation.
• Composer shell polish. Mic affordance and provider-row aura refinements on the composer chrome.
• Sidebar workspace hierarchy. Startup layout for nested workspace trees is tidier and less jumpy on first paint.
• iOS run details token table. Token summary layout simplified on the companion run-details screen.

Fixed

• Branch popover readability and placement. The composer branch/worktree popover reuses the Model/Reasoning picker frosted-glass chrome and opens above the trigger by default so it no longer overlaps the picker row below.
• Grok and Cursor thinking traces. Incremental thinking deltas preserve whitespace and word boundaries instead of trimming each chunk before append.
• Brief mention overlay scroll sync. Ensemble brief @-mention overlays stay aligned while the transcript scrolls.
• Release-class shell commands. Approved release workflows (git push, git_create_pr, and related host commands) bypass redundant shell blocks after explicit user approval.

Documentation

• New guide: SESSION_AND_WORKSPACE.md (caching tiers, forks, worktrees).
• Architecture and Advanced Optional Setup cross-links for BYOK caching caveats.

Full changelog: CHANGELOG.md

---

macOS downloads

Artifact	SHA-256
TaskWraith-1.7.6-universal-mac.dmg	809e3f9d2f3c601ca08288101be419c02001967a8228210b1de27795eae6a0e5
TaskWraith-1.7.6-universal-mac.zip	097555e0dd2948e0c3efe6b5e213392718aca29a220a809e6214f3fcf775b542
latest-mac.yml	74dec6e8a18ff35410af020c497c978649d6cc53c25e15b9d550ce98445ba3c3
sbom.cdx.json	88503231e0bf03c3d8bed1462934f41df2e4c9c377d57cb064cbd4fc590777cb

Checksums file: SHA256SUMS-1.7.6.txt

Notarization note: The universal .app inside the DMG is signed, notarized, and stapled. DMG-level stapling was deferred (stapler validate on the DMG reports no ticket yet); Gatekeeper acceptance was verified on the packaged .app.

---

iOS companion

TestFlight build 65 (marketing version 0.1.0) uploaded to App Store Connect with production push entitlements.

---

Upgrade

Existing installs on the stable channel should pick up 1.7.6 via the in-app updater (latest-mac.yml). New installs: open the .dmg and drag TaskWraith to Applications.

TaskWraith v1.7.5

1.7.5 - 2026-07-05

Added

• Agent Pool icon colour controls. Hue and brightness sliders stay visible for
  seed, asset, and named icons — not only after Shuffle — with editable Hex and
  RGB fields for precise picks. When Tint icon is on, your chosen colour
  overrides pre-baked asset hues; when off, the preview stays monochrome while
  the colour fields still show your last pick.
• Agent Pool leaderboard. Settings → Agent Pool now includes a sortable
  leaderboard table (styled like Model Usage) ranking pooled agents by runs,
  threads, tokens, tool calls, edits, work time, ensemble stage/role, and last
  active. Stats accumulate forward from saved ensemble runs.
• Transcript participant filter rail. Ensemble transcripts gain a compact
  filter rail so you can focus on one participant's messages without losing
  round context.
• Add selection to composer prompt. Highlight transcript text and send it
  straight into the composer as quoted context.
• Composer voice input. Pick a microphone source and dictate locally into
  the composer; on-device transcription fills the prompt without sending audio
  off-machine.
• Ensemble brief presets. Edit reusable brief presets from Settings and
  update them mid-round via a new ensemble_brief MCP tool.
• Stacked fan-out working indicators. Parallel fan-out lanes now show a
  clearer stacked busy state while seats are running.
• Plugin contributions surface. Activated plugin capabilities, tool bundles,
  local services, workflow templates, connector secret setup, health probes, and
  launch runners are wired into Settings and the new-workflow menu, with review
  gates before risky activations.
• Sketch canvas tooling. Agents can open a persisted sketch canvas for
  lightweight markup workflows.
• Poolside Laguna (Ollama). A new local-model preset joins the Ollama catalog.
• Full thinking traces. Extended reasoning traces are surfaced where the
  provider exposes them.
• Compact collapsed model-usage grid. With the sidebar collapsed, model usage
  condenses into a tighter grid layout.

Changed

• Fan-out transcript grouping. Parallel fan-out activity rolls up into grouped
  cards instead of scattering individual lane messages through the thread.
• Ensemble send queueing. Sends issued while the orchestrator is busy route
  through the ensemble queue instead of racing the active turn.
• Solo assistant labels. Single-provider chats label assistant turns with
  provider identities rather than generic placeholders.
• Workspace board creator. Creating a board opens from a main-pane sheet
  instead of a sidebar-only flow.
• Sidebar rhythm and composer tabs. Section spacing is normalized; tucked
  composer tabs cap at 850px so wide monitors do not sprawl.
• Ensemble chars slider polish. Glass styling refinements and removal of
  noisy warning hints on the orchestration row.
• Ollama branding. Mentions, health cards, and iOS working labels carry
  consistent Ollama identity.
• Provider tools catalog. The tools reference is reorganized for easier
  discovery.

Fixed

• Agent Pool tint precedence. Legacy baked accents no longer win over user
  hue/brightness when tinting is enabled.
• Settings roster brief editor width. The ensemble brief editor no longer
  overflows its panel.
• Workflow quick-create from general chats. Quick workflow creation works
  from non-workspace threads again.
• Participant filter rail alignment. Underfilled filter rails bottom-align
  cleanly against the transcript edge.
• Completion-claim warnings retired. Stale completion-claim transcript
  warnings are removed in favour of the current evidence-pack flow.
• Active run provider labels. In-flight runs show the correct provider name.
• Ollama health card brand fallback. Missing brand metadata no longer leaves
  a blank card.

---

SHA-256 checksums (macOS artifacts):

31708ff3a630b7bed131cddcb0368e15d3104d6af89cd55b4c10f67af965c7cc  TaskWraith-1.7.5-universal-mac.dmg
4d94de30fbf7ad6a6554cc4961231b7f552c1f502780db5ce7cc600a559415b6  TaskWraith-1.7.5-universal-mac.zip

Windows and Linux installers are unsigned CI builds attached after publish.

TaskWraith v1.7.4

1.7.4 - 2026-07-04

Added

• Mid-thread ensemble toggle. Top-level chats can now flip in place between
  single-provider and Ensemble mode on the same chat, preserving transcript and
  run history. Solo → Ensemble seeds a single participant from the current
  provider instead of recreating the thread, and Ensemble → Solo collapses back
  to a canonical provider — defaulting to the current Boss participant so the
  thread keeps its lead voice. The toggle is available on started threads when
  idle and stays disabled while a turn is active.
• Ensemble blackboard — a shared scratchpad for your agents. Participants can
  now post durable, bounded shared-memory entries (decisions, verified facts,
  open risks, do-not-repeat notes, concise notes) that every seat sees as a
  compact digest instead of re-deriving the same facts each turn. Agents get
  blackboard_post / blackboard_read / blackboard_delete tools with per-seat
  "seen" tracking, entries carry a round / session / chat lifespan, and the whole
  board is browsable — with its unseen/seen state — from the Pins panel.
• Ensemble fan-out controls, refined. The orchestration-row Fan-Out control is
  clearer to steer, and a parallel writer wave now lands as a single Fan-out
  result card that gathers each seat's contribution in one place instead of
  scattering them down the transcript.
• Live seat-compaction progress. When an Ensemble seat approaches its context
  ceiling, you now see its compaction happening live — an in-progress indicator
  on the participant and a boss-facing cue — rather than only a card after the fact.
• File-change hover diffs. Hovering a changed file in a run now previews a
  compact per-file diff summary (added / removed line counts and the change
  shape), so you can gauge what a step touched without opening the editor.
• Transcript rail reading lens. The user-message marker rail now carries a
  slide-rule-style "lens": a frosted carriage whose height shows how much of the
  thread is on screen and whose position tracks your scroll. Ticks it has swept
  past settle into an "inked" spent state, marking the done/upcoming boundary as
  a second, colour-independent cue.
• Rail jumps now glide. Clicking a go-to-message tick or the ↑ / ↓ arrows
  animates the transcript to the target (distance-scaled, eased) instead of
  teleporting. Any wheel / touch / key input mid-glide cancels it instantly, and
  reduced-motion settings restore the instant jump.
• Collapsed rounds keep their rail markers. In Ensemble chats, user prompts
  hidden inside a collapsed round card still appear on the go-to-message rail,
  anchored at the round's header; clicking one auto-expands the round and glides
  to the prompt.
• Quick controls survive collapsing the sidebar. Hiding the workspace sidebar
  now surfaces a bottom-left vertical glass pill with the sidebar footer's quick
  controls — Settings, Approvals, Shares, and Devices — each opening the same
  popover as its sidebar counterpart, with the same pending/collaborator/device
  glows.
• iOS: richer notification banners. Turn-complete and status banners on the
  companion carry more context at a glance, while genuinely noisy remote
  notifications are dialled back so the ones that arrive are worth reading.

Changed

• Provider / model / reasoning switching no longer stops at first send. In
  normal chats, the composer provider and model / reasoning pickers stay usable
  after a thread has history. Idle changes apply immediately; while a turn is
  active they queue and apply at turn end. The same-provider case keeps the live
  session; only a genuine provider switch resets provider-linked session state.
  The iOS companion mid-thread switch got the same session-hygiene treatment.
• Continuous-mode ensembles keep going on their own. A Continuous round no
  longer stops the moment agents stop explicitly handing off. When the roster
  drains with an active goal, TaskWraith re-dispatches another pass automatically
  — up to the handoff-turn budget — and stops cleanly when the goal is completed,
  blocked, or paused, when the hop budget runs out, or when a round makes no
  progress.
• Ensemble roster is easier to shape. The old minimum-two-participant floor is
  gone, so you can pare a roster down freely, and collapsing or reworking the
  roster no longer discards participants you meant to keep.
• Guest participants removed. The older Guest helper path has been removed from
  desktop, iOS, bridge, and live docs. Historical guest transcript rows remain
  render-safe and inert.
• Corner-pill polish. The bottom-left column pill's dividers span the full pill
  width and its icons are larger. Active pill buttons swapped the old
  left-weighted blue gradient — whose near-white glyph vanished on light themes —
  for the neutral glass blob + hairline ring the hover state already used.
• Quieter sidebar and calmer flourishes. The sidebar's model-usage heatmap was
  retired, the masthead and New button line up with the brand row, transcript
  item separators read as a softer silver, and the sky-diff easter egg is pared
  back to just its drifting +/− line counts.

Fixed

• Priority @-mentions reach the Boss even after they've spoken. Directing a
  round back to the Boss — or the acting Captain when the Boss is away — after
  that authority had already taken its turn printed a "takes routing priority"
  note but then silently dropped the route. The authority is now genuinely
  re-summoned (bounded by the handoff budget so it can't loop), and when a
  re-summon truly can't be delivered the note reports the real reason.
• Dedicated writers fan out in parallel again. Two round-start gating bugs let
  stage reviewers and read-only workers falsely veto a parallel writer wave,
  forcing writers back to a slower serial pass; parallel writer fan-out now runs
  when it's warranted.
• Ensemble steering and recovery hardened. A steer issued during a parked
  window is now honoured, zombie dispatch after a cancel is stopped, and an
  ensemble participant recovered after an app restart is labelled as itself
  rather than mislabelled as a solo run.
• Stale run-queue rows no longer linger. Active-run queue entries left behind
  by a finished or interrupted run are suppressed, so the queue reflects what's
  actually running.
• Chat titles and summaries stay honest. Chat-list summaries are validated
  against each chat's own file so a title can't drift away from its conversation,
  index churn is throttled to keep the list steady, and session checkpoints
  persist more reliably across compaction.
• A pending question now sits at the live tail. In Ensemble chats an
  ask_user_question card could strand itself above the speaker; it now pins to
  the bottom of the live conversation where you'd expect to answer it.
• Rail no longer bleeds under the composer. With the workspace sidebar
  collapsed, the go-to-message rail could overlap the floating composer's left
  edge; it now anchors to the leftmost mounted lane, stays fully to the
  composer's left, and re-measures when layout transitions settle.
• Settings is reachable with the sidebar collapsed. Opening Settings while the
  sidebar was hidden left no tab navigation and no "Back to app"; the sidebar slot
  that hosts the Settings nav now force-mounts for the takeover and slides back
  out on close, without touching the user's collapsed preference.
• Right-dock panels no longer stack from the top pill. The rim buttons (Run
  rail, Media, Notes / Pins, Terminal, File editor, Inspector) route through the
  dock's exclusive lifecycle: opening one surface replaces the previous, and
  closing the active surface collapses the whole dock in a single click.
• Remote approvals that over-shoot the offered tier are honoured. Tapping
  "Accept for workspace" on an iOS approval that only offered a narrower accept no
  longer silently drops it; the accept is down-clamped to the strongest offered
  tier and the command actually runs.

Security

• Roster permission changes respect the workspace auto-edit gate. Setting an
  ensemble participant to an auto-edit-tier posture (workspace-write or full
  access) from a remote device now requires the workspace to permit auto-edit,
  closing an escalation path that bypassed the gate the composer already enforces.
• Cursor native shell and writes unlock only under a real Full access grant. A
  write-capable Cursor run keeps its native shell / write deny-list — routing
  through the TaskWraith broker — unless it's running under a genuine, signed Full
  access grant, matching how Codex full-access is treated.

---

SHA-256 checksums (macOS artifacts):

67c0dd36e0261be49dfd687abb4d1f14cf98d685889b83cd0f2a0af5e0db3dee  TaskWraith-1.7.4-universal-mac.dmg
52fad4aaeeda344c8e3e90bbddba3168925bfe85a8963951035c40a9614e1060  TaskWraith-1.7.4-universal-mac.zip

Windows and Linux installers are unsigned CI builds attached after publish.

TaskWraith v1.7.3

TaskWraith 1.7.3 — Full Access agents can build, sign & ship

Fixed

• Full Access agents can build, sign and ship. A Codex agent running under
  the Full access permission preset now launches without the workspace
  sandbox (danger-full-access) instead of being confined to the repository.
  Under the previous workspace-write confinement it could not reach the login
  keychain (code-signing identities), ~/Library caches (SwiftPM / DerivedData),
  or paths outside the workspace — so an approved release task (iOS
  archive / notarize / TestFlight upload) failed at signing even after the user
  approved it. The sandbox now drops only for a genuine, signed Full access
  grant; every other run stays workspace-confined, and a global "deny shell
  commands" setting still overrides it.

Added

• iOS: grant Full access from your phone. The iOS participant permission
  picker now offers Full access, matching the desktop permission picker, so
  a phone-driven ensemble participant can be given the full-access posture that
  lets it build, sign, and ship.

Companion: iOS build 62 on TestFlight.

---

macOS artifacts (universal, notarized)

SHA-256:

51e7a03b034ac28f6192d38773ee62939006459ce54a7d82e984b02ca368b77d  TaskWraith-1.7.3-universal-mac.dmg
2493048ffe6017286be72010d72163e61b1065f2d1a9d40deeb8e9e0945da410  TaskWraith-1.7.3-universal-mac.zip

Windows and Linux builds are attached separately via CI.

TaskWraith v1.7.2

1.7.2 - 2026-07-03

Added

• iOS ensemble orchestration parity. The iOS companion can now adjust
  ensemble controls that were previously desktop-only: designate the
  Captain (second-in-command who steps in when the Boss is unavailable),
  switch between Turn and Continuous mode, set the max handoff
  turns (hop limit) for continuous rounds, choose the Fan-Out policy
  (Off / Read / Write, where Write resolves to the Boss-gated or
  user-preflight writer lane), and set the shared-transcript character
  budget. Each change round-trips to the Mac and takes effect on the next
  round.

Changed

• Question routing. The agent runtime preamble now steers agents to the
  TaskWraith ask_user_question tool instead of a provider-native question
  prompt, which silently auto-resolves without reaching the user in this
  harness (desktop or the iOS companion).

---

SHA-256 checksums (macOS artifacts):

ed1700dde039b7a4fc790cb9f32f5a922f236c4aa43fe92805fdede7e9d0049e  TaskWraith-1.7.2-universal-mac.dmg
eeac06b1020577b761b1755eac47622849f20a7773b5f8354c27c18bc4c9b920  TaskWraith-1.7.2-universal-mac.zip

iOS companion build 0.1.0 (61) was uploaded to TestFlight with delivery UUID a392efdb-99db-4e47-99a6-c1274daf53dc.

Windows and Linux installers are unsigned CI builds attached after publish.

TaskWraith v1.7.1

1.7.1 - 2026-07-03

Added

• Context compaction, made visible. Solo Claude and Codex chats now show a
  glass compaction card when a context reset completes or fails, and the
  composer context donut grows a pressure state (amber at 80%, red at 95%) with
  a Compact context now button once a linked session is under pressure.
  /compact stops being a cosmetic prompt template and dispatches a real
  provider-native compaction, falling back to the template only for busy or
  session-less chats. Overflow errors get a named remedy hint instead of a raw
  wall of provider text.
• Ensemble seat compaction. Long-running Ensemble seats no longer stall at
  their context ceiling. Each participant carries a rolling summary, the
  orchestrator materializes per-seat compaction cards and boss-facing pressure
  cues, and a post-round auto-trigger keeps Claude, Codex, Cursor, Kimi, and
  Grok seats healthy across a long session with a cooldown so it never thrashes.
• Ensemble stage roles. Participants can now be assigned a scout,
  worker, or reviewer stage. Reviewers wait for the writers instead of
  running against work that does not exist yet, then run together in a closing
  Review wave once only reviewers remain. Stage roles round-trip through
  roster presets, Boss roster edits, and the iOS Roster page picker.
• Message feedback. Assistant messages now take a thumbs-up / thumbs-down
  from the action row or right-click menu, with optional reason chips on a poor
  rating. Votes persist with the message and roll up into per-model casting
  summaries in Settings, laying the groundwork for recasting a weak answer with
  a different model.
• Redesigned right dock. The two-tier glyph strip is replaced with a clean
  surface switcher: a slim header whose active-surface button opens a grouped
  Session / Work / Inspect popover, the Inspector's sub-views regain text
  labels, and a ⌘K command palette searches across every dock surface. The
  active surface is now remembered per chat, and the iOS companion mirrors the
  per-thread inspector memory.
• Agent Pool gets its own page. The reusable Agent Pool moves out of the
  foot of the Ensemble Roster tab into a dedicated AI & Providers -> Agent
  pool settings page, so it has room to grow and is no longer squashed under a
  large roster. The roster tab keeps its per-participant Save to pool action.
• Projects sidebar, finished. The Projects tab added in 1.7.0 gets a full
  polish pass: archived member chats stay visible with an Archived chip,
  expand state persists, chat rows drag straight onto a project, search is truly
  scoped per tab, and the whole tablist is keyboard- and screen-reader-navigable.
  A migrateProjects() step means schema drift no longer drops project data.
• Collaboration contribution rules. Shared chats can now carry a contribution
  preset (read-only, comments, request host action, or auto-draft). Collaborators
  on an enabled share can submit a structured action request that goes to the
  host for review and is never dispatched to a provider, optionally pre-filling
  the host's composer as an external-untrusted draft. A bounded, redacted
  collaboration audit log records share, invite, and contribution events.
• Collaborator reconnect. A dropped collaborator can rejoin their last shared
  chat with a pinned, encrypted identity — no fresh invite token and no repeated
  safety-number compare — while still signing fresh session keys every reconnect.
  Reconnection survives host restarts; revoked participants stay locked out.
• Clickable tool-call file paths. File paths in tool-call rows are now
  clickable and open directly in your editor, with a hover preview. The
  user-message transcript rail also gains a scroll-spy highlight, a progress
  track, and a hover bulge for moving between prompts in a long conversation.

Changed

• Local models follow the standard permission role. Ollama's bespoke
  tool-control tier ladder is retired. Local models now get the full tool
  surface and obey the same Plan / Read-Only / Default / Full Workspace Access
  role as every other provider, governed at the shared approval gate. Read-only
  web search and fetch are now permitted under Read-Only and Plan for all
  providers, so a local model can finally search the web in a General chat.
• Managed-enterprise readiness. The B5 items previously tracked as pending
  have now landed: scoped and redacted audit-bundle export with signed
  verification receipts, managed-policy loading and clamps (settings, update,
  safety, bridge, MCP, and user-MCP allowlisting), encrypted storage for MCP and
  runtime-profile secrets, and dispatch receipts stamped through the queue and
  scheduled-run lifecycle. The claim stays deliberately bounded to local safety
  plus partial managed policy and redacted audit export — no SSO/SCIM, SIEM
  integration, WORM or append-only export, or organization-wide retention is
  implied.
• Claude/Codex-parity scroll follow. Lock-to-bottom now behaves like the
  Claude and Codex apps: a Jump to latest pill appears while a single answer
  streams into view, the jump reliably re-locks even as content grows mid-flight,
  sending a prompt while scrolled up re-arms follow, and any upward gesture always
  wins.
• Sidebar running vs. selection. A running thread now shows a slow-pulsing
  monoline "ghost" mark, and the accent rim is reserved strictly for the selected
  thread — the two were nearly indistinguishable before. Each sidebar section
  previews its first few threads with a show more affordance, and on macOS the
  corner-controls pill clears the window traffic lights when the sidebar is
  closed.
• Ensemble roster controls. Orchestration controls move onto a labeled
  roster-presets second row, the participant chip strip reflows into balanced
  rows of five, and the roster cap is raised from 18 to 20. The composer's
  Ensemble toggle moves to the footer.
• Persistent Grok seats and slimmer resumes. Per-seat Grok ACP sessions and
  slimmed resumed-seat turn prompts are now enabled by default, trimming repeated
  context on long multi-turn Ensemble rounds.
• Composer polish. Diff counters animate as they change, cost estimates line
  up with the remote projection shown on the companion, and working-text glow is
  softened.

Fixed

• Ensemble reliability. A cluster of round-lifecycle fixes: the composer and
  queued-row Steer now act on the first click, queued-row Steer and Delete
  recover after an app restart, seat changes hand over cleanly and preserve their
  stage role, duplicate yield-activity rows are removed, and a clean exit with
  streamed content finalizes as answered instead of hanging.
• Startup opens a fresh chat. Launch no longer restores the most recent
  global chat into a stale, un-hydrated transcript. The app now opens a fresh
  single General chat on your default provider every time; workspaces and
  ensembles remain one click away in the sidebar.
• Right-dock resizing. The inspector divider resizes reliably again, and the
  dock's surfaces are now mutually exclusive so opening one no longer leaves
  another stranded behind it.
• Local tool-call repair. The Ollama tool-call repair loop is hardened: the
  decode grammar allows the full tool catalog rather than only advertised names,
  harness-gate blocks count toward the retry ceiling, argument validation matches
  the executor gate, and each Ensemble seat keeps isolated tool memory.
• Duplicated assistant text. The legacy delta lane now defers to the run-item
  sidecar, so streamed assistant answers no longer render twice.
• Permission honesty. Network-deny is honored for web tools, evidence-pack
  and observability tools are classified as orchestration, and Plan-workflow
  posture is now visibly distinct from Read-Only in labels and over the iOS
  bridge.
• Usage meters. The Claude Fable weekly quota meter reads the current
  session-group payload shape, the Grok /usage probe works against the new
  CLI's terminal queries, and all four Claude quota windows survive on the iOS
  first-launch card.
• iOS companion parity. New task starts preserve their permissions,
  Plan-workflow posture carries correctly over the bridge, and Agent Pool
  transcript identity is de-boxed to match the desktop.

Removed

• Ollama tier machinery. With local models now on the standard permission
  role, the dead tier resolvers, tier tables, mid-run tier-bump path, the
  now-inert tier and run-profile picker UI, and the unused global run-profile
  settings surface are all deleted.
• Retired messaging gateway. The old messaging gateway and its cleanup
  residue are removed.

---

SHA-256 checksums (macOS artifacts):

194af17d4ae59e1127e6db06be26f767ad8aa3328a4122cf3d1828c147dc4966  TaskWraith-1.7.1-universal-mac.dmg
9b19d42f7bbad687bebabc7f9ae49995b4fddda0a37dd8c1b083a7d26f0d96cc  TaskWraith-1.7.1-universal-mac.zip

Windows and Linux installers are unsigned CI builds attached after publish.

TaskWraith v1.7.0

1.7.0 - 2026-07-01

Added

• Plan Mode workflow. Composer mode now separates Plan workflow from Read-only recon. Plan chats can surface a single proposed plan from the designated owner, approve / customize / dismiss it, and carry the approved artifact path into implementation.
• Plan artifact writes. Plan workflow can write markdown plan files under validated workspace paths while the signed workflow posture prevents ordinary read-only recon runs from unlocking that carve-out.
• Codex native-review status cards. Native Codex review progress now flows into review/status cards so transcript activity shows review findings and state without burying the signal in raw provider events.
• Ensemble read fan-out skip. Active read-only scout lanes can be stopped before the writer step when the panel has enough evidence.
• Boss roster-swap choices. Boss participants can inspect live participant ids, provider/model catalogues, context windows, and coarse quota bands before swapping an inactive seat.
• iOS proposed-plan file handoff. The companion now receives plan artifact paths, shows them on proposed-plan cards, and can jump straight into Files mode for that plan artifact when the paired Mac exposes a workspace scope.

Changed

• Claude picker line-up. Mythos 5 is retired from current Claude model pickers while Fable 5 and Sonnet 5 remain visible; historical Mythos chats, aliases, display names, rates, context windows, and CLI normalization stay compatible.
• iOS Plan workflow parity. The companion copy now names Plan workflow directly, and its model context reference mirrors the post-Mythos Claude line-up.
• Ensemble recovery after restart. Multiple interrupted Ensemble lanes in one chat now collapse into one grouped recovery system message with role/participant labels and provider-session hints.
• Ensemble steering reliability. Steered queued prompts now wait for active round cancellation before redispatching, avoiding overlapping fan-out teardown and queue-persist churn.
• Boss and participant boundaries. Agent-driven roster edits and Boss replacement are capped so agents cannot assign Full Workspace Access; only the user can elevate to that level.
• Ollama Ensemble context. Local lanes preserve their assigned participant role and Lead/Boss routing across tool loops, retries, and compaction.

Fixed

• Plan Mode security. Workflow mode is signed into run permission posture, and plan-artifact paths reject symlink / realpath escapes.
• Recovery transcript noise. Restarting after interrupted parallel fan-out no longer appends one nearly identical system row per lane.
• Brokered MCP transcript labels. Brokered MCP tool calls now render with the MCP plug icon and uppercase label.
• Inactive participant pickers. Ensemble participant provider/model pickers stay editable for inactive seats while a round is running, so quota relief and seat swaps do not get blocked by the active speaker.

Known Open

• Grok interrupted-lane resume. Grok session-id persistence remains evidence-first after the steering fix; no source patch is claimed without a fresh post-fix recovery capture.
• Ollama literal write-artifact probe. The Ollama retention/probe item stays accepted-open until a controlled artifact-backed repro says otherwise.

The macOS build is notarized + stapled (universal). Windows (unsigned) and Linux artifacts are attached by CI after the release is published.