StarNet v0.12.4

StarNet v0.12.4

Full update notes: https://github.com/androoAGI/starnet/releases/tag/v0.12.4

Projects now bring their orchestrator conversation, crew, and delegated work together in COMMS. This release also adds selected Google file access and optional Windows accessibility computer control, alongside extensive conversation, station-saving, and recovery improvements.

Projects and crew coordination

• Open a project into one persistent orchestrator conversation with the project's working context.
• Choose and save a preferred crew for each project.
• Follow delegated work inside COMMS, with inline status updates and expandable activity details showing the agent, task, result, tools used, outputs, and directions.
• Send a direction to a working agent or request that its work stop from the activity view. Queued directions remain visible and are included in the orchestrator's review.
• Keep the station visible while moving between project conversation, crew, and activity views.
• Improve project control spacing, activity readability, and consistency with COMMS' glass appearance. Status announcements occur when an update actually changes.
• Fix saved projects getting stuck during loading because of an undefined session status.
• Preserve crew controls when reopening projects, keep activity order stable, and prevent duplicate worker controls during refreshes.
• Keep transcript navigation out of project detail views. Failed or revoked project opens no longer relabel the previously open conversation, and late responses cannot pull focus back after navigating away.
• Show unavailable project context and activity-refresh failures explicitly, retaining the last confirmed activity while reconnecting.

Overseer and delegated conversations

• Coordinate background workers from existing COMMS orchestrator conversations, with durable worker conversations and results returned to the originating conversation for review.
• Restore worker conversation history and project context when continuing or resuming delegated work.
• Reconcile delegated conversations and completed results without switching the conversation you are reading.
• Deliver review summaries without exposing internal task labels in the conversation.
• Refresh queued conversation context before execution and retain compatibility with legacy session dispatch and the first-save bridge.
• Keep automatic coordination scoped to COMMS orchestrator runs. Stop, halt, emergency recovery, and update handling respect coordinator boundaries.
• Refuse new overseer work if its durable halt state cannot be saved, and preserve worker/review state through restart and failure paths.

Conversations, attachments, and COMMS

• Restore conversation history before sending a new message, including when history arrives slowly or a conversation is reopened while a read is pending.
• Keep unsent messages when history cannot be restored and report the failure instead of sending without the expected context.
• Discard stale history responses after switching, clearing, or removing a conversation; preserve drafts and conversation identity through reconnects.
• Require confirmed completion before marking a streamed reply complete. Interrupted replies retain their partial content instead of appearing successfully finished.
• Drop files anywhere across the chat panel, with a visible attachment target, multiple-file handling, previews, and the existing attachment limits.
• Handle cancelled drags, unavailable files, disabled chat, and folder drops. Accidental drops outside an attachment target no longer navigate away from the station; normal text and link editing is preserved.
• Adjust COMMS layout and expand/restore behavior for narrow and short windows while retaining drafts and panel sizing.
• Retain bounded conversation-continuity and page-error diagnostics to help investigate interrupted sessions without recording message contents in the continuity trace.

Station building and movement

• Autosave station design changes while editing, before leaving build mode, including bursts of changes without losing conversation state.
• Reuse the existing environment render during object placement to avoid unnecessary floor rebakes in populated stations.
• Let agents yield, step aside, or retreat through narrow hallways so opposing traffic can reach its destination; improve following distance and handling around seated and working crew.
• Keep the website demo aligned with the station, placement, COMMS, onboarding, project, and reliability improvements.

Google files and connector storage

• Add Selected Google files in Abilities: use Google's file picker to grant access to specific Docs and Sheets through the limited drive.file permission.
• Provide 11 tools for accessible-file metadata/export, reading/creating/editing Docs, and reading/creating/editing Sheets. This is not whole-account Drive search; broad Drive, Docs, Sheets, Gmail, and Calendar connections remain deferred.
• Use CHOOSE GOOGLE FILES on a connected service to select more files. Cancelled, empty, or broader-than-supported grants preserve the previous connection.
• Require a durable refresh grant and recheck the limited permission after restart and refresh; refreshing cannot silently widen access.
• Encrypt desktop connector credentials using an encryption key backed by the operating system credential store, with migration from existing plaintext connector state.
• Preserve original credentials and recovery copies when migration, writing, key access, or JSON validation fails. A locked or malformed credential store reports an error rather than replacing the original key or state.
• Export encrypted connector settings through the running app with its credential store unlocked, while removing credentials and machine-specific permissions from recovery exports.
• Clarify privacy disclosures: encrypted credentials do not encrypt retained file contents. Disconnecting removes local credentials, not Google's grant or previously retained work.

Windows computer control

• Add optional CUA · Accessibility under Abilities → Computer Control for Windows x64, alongside Windows · Classic and Off. Installing CUA does not enable it; the selected driver persists across restarts and applies to new runs.
• Give authorized agents app/window discovery, screenshots with accessibility elements, semantic and pixel input, menus, window positioning, clipboard actions, and bounded state verification.
• Verify the downloaded driver's hash, version, and helper before activation. Check Driver and Reinstall CUA support diagnosis and repair; a failed replacement download preserves the current installation.
• Isolate native control sessions per run, serialize their actions, and close them on completion, cancellation, transport failure, or driver changes. Applications opened for you remain running after the agent's session closes.
• Recover expired sessions on a subsequent call, reject stale element references, and report refused or unverified actions without automatically repeating an ambiguous mutation.
• Keep driver telemetry disabled and provider credentials out of its child environment; report cleanup failures without logging desktop content.
• Native CUA control is currently Windows x64 only. It does not add always-listening audio or change voice permissions, and restricted runs cannot start the native driver.

Saving, permissions, routines, and recovery

• Preserve the last known saved station when reads fail or save acknowledgements are missing; distinguish local restore failures from a disconnected backend.
• Strengthen durable-store recovery so read/write failures cannot masquerade as empty state or confirmed changes.
• Serialize permission changes and reject stale responses, preserving the latest confirmed grant, revoke, or bypass decision.
• Validate routine, messaging, and other asynchronous control acknowledgements before showing success.
• Keep marketplace skills, project, connector, routine, and run-history reads from treating service failures as verified empty results. Disabled or unauthenticated connectors no longer count as ready capabilities.
• Preserve spending authority through failed writes, incomplete or truncated ledger reads, and hard restarts. Uncertain accounting no longer appears as a trustworthy zero balance or silently restores permission to spend.
• Allocate limited routine execution slots to the oldest due work to prevent starvation.
• Elect a single writer during concurrent stale-workspace recovery using immutable ownership generations; protect whole-root recovery and workspace activation from takeover races and crash windows.
• Refuse recovery when process ownership cannot be established, including ambiguous routine-lock ownership, rather than assuming another process is gone.
• Improve queue and browser-probe cleanup so failures do not strand pending work or timers. Isolate seeded lifecycle and recovery-test workspaces to avoid cross-session interference.

Desktop compatibility, packaging, and maintenance

• Bundle the shared specialty catalog with the desktop frontend and validate staged boot assets.
• Replace regular-expression lookbehind and newer timeout-API assumptions that prevented older macOS WebKit from parsing or running affected chat, widget, and history paths; preserve escaped-pipe table rendering.
• Improve onboarding sign-in-code contrast on dark glass in the desktop app and website demo.
• Stage production dependencies in desktop bundles and update Sharp and archive dependencies to patched versions.
• Expand regression coverage for persistence faults, ownership races, spending uncertainty, conversation continuity, project recovery, file drops, responsive layouts, native control, and credential migration.
• Improve QA isolation, screenshot readiness, shared-host gate deadlines, installed Mac version receipts, and upgrade comparisons that account only for the new...

StarNet v0.12.3

Download

Platform	Installer
Windows (x64)	StarNet_0.12.3_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.12.3_aarch64.dmg
macOS (Intel)	StarNet_0.12.3_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.12.3

This update remasters the station itself: industrial materials, a redesigned command bridge, a furnished default station with every essential ability already installed, a browsable Build library, remastered agent skins with grounded walk cycles, and an onboarding that goes from what you want to do straight to a first useful task. It includes fixes and hardening for customer-reported update, boot, rating, delegation and conversation-history problems reported since 0.11.2.

These notes cover the changes shipped since 0.11.2. The 0.12.0, 0.12.1 and 0.12.2 candidates were held before publication; this final build includes the complete overhaul and the final sprite-animation merge.

• Close-to-tray: finishing startup or reloading the document no longer reopens a window you closed to the tray. Initial display waits for the completed document; explicit tray Open and a second launch still reveal the app.

• Saved model choices: a temporary catalog error or offline fallback list no longer clears your selected model, including during startup and refresh.

• Rebuilt walk animations: 33 redesigned skins now use eight-frame alternating walk cycles in all eight directions, with corrected manifest pointers. Capybara retains its earlier walk frames; the retained original skins are unchanged.

A remastered station

• Industrial station materials: walls, floors, ceiling coping, corridor doorways and shells are rebuilt in a darker industrial style with orbital shells, clean station floors and consistent structural frames. Wall heights, doorway reveals and corner joins are aligned across every room shape.
• Redesigned command bridge and props: the workstation, tactical table, capability equipment (terminal, files cabinet, comms dish, memory core, media studio), lounge furniture, storage, lab and workshop machinery are re-authored at station scale with readable forms, fitted screen effects and truthful activity (screens, mechanisms and indicators follow real harness state).
• Pre-equipped default station: a new station starts as one furnished room with a workstation, all five essential ability props and plants, so a fresh Commander can hand over a first task without placing equipment. Seven station presets (DEFAULT, QUIET RETREAT, COZY WORKSHOP with a connected conveyor line, CREATIVE STUDIO with a prepared draft-and-review line, RESEARCH, ENGINEERING and OPERATIONS stations) can be applied from Build and the previous station restored.
• Build library: Build Mode is a spacious, searchable prop library sectioned into FURNITURE, EQUIPMENT and ABILITIES with category menus, previews, clear placement/cancellation feedback, inline move cancellation and keyboard Escape. Build selections now match the visible artwork.
• Remastered agent skins: 37 finalized skins with grounded strides, corrected cardinal facings, natural turns, talking and sitting motion, workstation seating with proper foot occlusion, and layered couch seating. Legacy skin selections keep working.
• Movement and lighting: agents route clear of raised doorway walls, keep a natural travel pace, reconcile their walk after collisions and plant correctly when stopped. Station lighting pools, fixture illumination, back-wall light and exterior navigation lights are rebalanced; the shipped CRT look is preserved.
• Renderer recovery and performance: silently lost station canvases are recovered and remaster materials restored before a re-bake; the canvas watchdog no longer performs synchronous readbacks; lighting caches are reused. On the 96-prop stress layout the integration measurements recorded frame intervals of 8 to 15 ms at p95 with zero render faults.

From an idea to a result

• Goal-first onboarding: setup follows what you want to do, with a cinematic agent arrival and a first-task handoff. The optional quick tour explains the equipment already in the station and never asks you to place props; a typed first task survives the tour and returns as an editable draft.
• Personal journeys: WORK → QUESTS connects ambitions, plans, next actions and outcome reviews (Now / Goals / Progress / History). Planning suggestions can be reviewed, adopted, edited and undone; drafts survive refreshes and delayed saves; finished steps link to the exact run output.
• Creative Studio guidance: a guided setup and verified sample workflow for the Studio, with a prepared draft and review conveyor in the furnished preset.

Everyday improvements

• Compact speech bubbles and fewer redundant delivery acknowledgments.
• Conversation history stays in order: refreshing or retrying a conversation no longer reorders or rewrites the local transcript (the "my turns vanish / an old message moved to the front" report).
• Composing stays in your chosen session: background workshop events and delayed connector checks respect the focused composer, unsent drafts and attachments. Explicit session navigation continues to work.
• Corrections replace old memories: explicit memory corrections preserve an audit history while keeping one current belief in recall. Approved requirements can stay pinned within their project or conversation, and stale reviews cannot overwrite newer edits.
• Incomplete work is explained: repeated promises without tool actions end with an explicit incomplete-work error. Unsupported claims that a preference was saved are checked against actual notebook-write receipts.
• Saved files stay clickable: the "▤ saved " and media rows an agent produced are replayed when a conversation is reopened, switched to, reloaded or retried, instead of surviving only in the model's prose.
• XP and ratings: earned XP refreshes in open views, skipped rating credit can be recovered, and interactive replies in a conversation that also contains scheduled routine activity can be rated on their own run record. Ineligible replies now say exactly why.
• Secondary panels render correctly at enlarged UI scale and narrow widths; provider text and dropdown affordances are readable; CRT curvature is accounted for in hover and click targets.
• Static CRT settings (0–200%) persist across reload independently of theme and glass.

Reliability and support repairs

• Boot guard self-heals a slow engine: if the desktop window loads before the station engine answers, the shared catalog is retried with backoff and the page reloads once on success instead of painting "STATION FAILED TO BOOT" (the shared/specialties.js reports on Mac and Windows). A genuinely broken file still shows the banner, now with the retry ledger for support.
• Account unlink recovery: unlink and status-read failures stay visible with a retry and a working LINK STATION path instead of silently hiding the controls.
• ChatGPT sign-in at setup: a live ChatGPT sign-in takes precedence over an old saved OpenAI API key when the key field is blank. A key you explicitly type still selects the API route, and connection errors identify the route used.
• Browser element discovery: the internal browser now finds JavaScript-driven cards, ARIA targets and open-shadow controls (the campaign-listing report).
• Telegram and channel leads receive the crew briefing matching the delegation tools they were actually granted; workers cannot delegate recursively.
• Local models: small-talk turns omit task-only prompt material and per-run ids ride last in the prompt so caches survive; Ollama output has a configurable ceiling and capped casual replies stop without automatic continuation.
• Response safety: auxiliary budgets and responsive fallback cancellation are preserved; durable unload saves are recognized when their acknowledgement was lost; the tutorial activity guard stays scoped to orientation.
• Delegated specialists keep access to connected MCP tools; cross-provider fallback keeps the selected route; task cancellation, review retries and routine/session binding repairs from the 0.11.2 follow-through are included.

Upgrade notes

• Existing stations keep their layout, props, crew, conversations, routines, goals and settings. The furnished presets never replace a saved station; they are applied only from Build and can be restored.
• Update from SYSTEM → SETTINGS → UPDATE CENTER, or download the installer for Windows or macOS from the release page. The desktop shell clears the stale WebView cache automatically on a version change.
• The installer is substantially larger than 0.11.2 because it includes the new station and character artwork. Review-only art is excluded; the textures needed to enable the remaster are included.
• The customary 48-hour attended soak was waived for this release. Candidate-specific installed smoke, scripted reliability and responsiveness receipts are recorded with the release preparation evidence.

StarNet v0.11.2

StarNet v0.11.2

Released September 11, 2026. This update expands agent personality and setup, refreshes the station's worlds and walking animations, improves report readability, and repairs save, provider, delegation, graphics and desktop-update behavior.

These notes cover the changes shipped since 0.11.1, including the final CRT static controls and Windows pointer repair.

Station visuals and performance

• Six rebuilt backdrops: THE NURSERY gains layered gas clouds, filaments, dust and twinkling cores; NIGHT CITY gains a lit street grid, building windows, rivers, bridges and beacons; OCEAN gains stepped waves, crest lines, reflections and pixel clouds; THE BELT gains detailed small asteroids and dust; THE MOON gains richer regolith, crater rims, ejecta and rocks; FOREST gains a dark aerial canopy, conifers, undergrowth, forest-floor detail and a river.
• Backdrop lag repair: heavy sky artwork is generated in a background worker, keeping the interface responsive while the scene is prepared. Ground scenery is cached rather than repainting every overlapping tree on every frame.
• Smoother camera movement: the ground cache moves with the camera and refreshes when zoom, window size or station footprint changes. Superseded artwork requests are discarded, with a compatibility path for environments without workers.
• Adjustable CRT static: Appearance settings now let you change and save static strength. The default remains 100%, preserving the existing look for stations that have not changed it.
• Smoother walking animations: expanded walk cycles add intermediate frames across agent skins, with turning cadence adjusted to preserve natural foot motion. Turtle rear-facing frames and wizard staff continuity receive specific corrections.
• More consistent glass styling: task-understanding, noticed/recommendation, away-work, decision and other transient cards now match the station interface. Station announcements are more readable, and compact recommendation labels remain compact.
• Panel sizing stays saved: docked panel height survives close/reopen and reload, with appropriate handling for maximized panels and smaller windows.

Overseer creation and agent personalities

• Rebuilt Overseer setup: a full-screen, two-step identity-and-connection flow brings the character gallery, name, personality, working style, provider, model and reasoning controls together.
• Clearer connection choices: StarNet account setup has a prominent branded action; alternative providers use recognizable local logos. Model selection supports keyboard navigation, custom endpoints and saved reasoning choices.
• Six distinct personality presets: Composed, Warm, Blunt, Dry, Unhinged and Upbeat now define more consistent behavior in conversation, disagreement, uncertainty, failures and successful work.
• Detailed personality tuning: adjust warmth, humor, formality, answer length and energy; choose no, occasional or frequent profanity; control emoji and directness; or write a custom communication style. Personality tuning is separate from the audible voice.
• Visible saved overrides: customized personalities are labeled, tuning survives preset changes, and Reset to Preset clears overrides. Older personality names map to the new presets. Unhinged's language choice retains an explicit confirmation.
• More reliable setup navigation: going back preserves identity, appearance and connection entries. Replaying the tutorial preserves the current agent name. Narrow-window layout fixes prevent character and identity sections from overlapping.
• Clearer setup recovery: stale provider-specific warnings and buttons retire when the connection, agent or missing setup step changes. A connected station without a model shows CHOOSE MODEL and opens the picker instead of asking for an unrelated API key. Recovery buttons work with both mouse and keyboard.

Providers, connections and discovery

• GitHub device sign-in: connect GitHub through its device authorization flow, with a visible code, expiry/cancellation handling and useful sign-in errors. Small windows keep the code accessible. If saving a new token fails, the existing connection is preserved.
• Discover improvements: popular platforms are easier to find, with clearer guidance into each platform's connection setup.
• Field Manual connection guide: work-app connections and messaging channels have dedicated setup guidance and navigation to the relevant panel.
• Provider presentation: refreshed provider logos, aligned connection statuses and clearer StarNet account actions make connection choices easier to scan.
• Model-selection race fixed: an older model-catalog response can no longer overwrite a newer provider or model choice, including overlapping picker requests and agent changes.
• Codex-to-OpenRouter fallback fixed: saved fallback choices now switch the provider route and credential source correctly when Codex reaches its quota. Paid fallback work retains its configured spending limits and cost record. Addresses issue #12.
• Claude continuation fixed: verification reminders and later host notes retain their proper place in the conversation on OpenRouter and compatible managed connections, avoiding the unsupported assistant-prefill request that could interrupt a task with HTTP 400.
• Delegated connector access fixed: specialists retain access to connected MCP tools through the lead agent's live approval flow, including dispatched, spawned and resumed work. Denials, connector removal and renewed confirmation after external content still apply. Addresses issue #13.

Saves, conversations and recovery

• Concurrent-save protection: stale clients cannot silently overwrite newer station edits. Conflicting local changes are retained for recovery, including edits made while conflict recovery is already open.
• Startup recovery: an unsynced local cache no longer prevents startup. Save-recovery actions use the station's own controls and retain the local recovery path.
• Queued-save restart repair: completed queued writes now retain the server's acknowledged revision, preventing false conflicts when the station restarts or updates.
• Conversation conversion preserves attachments: converting a direct conversation to a group retains historical attachment references and snapshots the required files before committing. If an attachment cannot be read, the original conversation remains intact.
• Conversion retries are safe: retrying the same conversion returns the existing result; a different request cannot accidentally reuse the earlier conversion's identity.
• Duplicate restored replies fixed: reopening a completed conversation reconciles locally combined streaming text with its durable turns instead of displaying both. Uncommitted text, attachments and stopped/error markers remain available.
• Loop review isolation: reviewing or undoing a loop iteration is protected against simultaneous execution, another review, or changes to the same loop, including while Git undo is still running.

COMMS, reports and task feedback

• Structured reports in COMMS: tables, ordered and nested lists, quotations and named links now render alongside headings, emphasis and fenced code.
• Readable narrow reports: wide tables stay within their own scrolling area, and table text remains readable at narrower panel widths.
• Exact report copying: copying a message preserves its original report text and Markdown; code-block copying retains its separate behavior.
• List-format fixes: plus-sign bullets and tab-separated list markers now render consistently during normal report display and restored history.
• Clear failed-session marker: failed sessions use a flashing X without repeating the same failure label beside it.
• Truthful partial results: failed or interrupted tasks retain the text, usage and evidence already produced instead of being presented as successfully completed work.

Crew movement, furniture and station behavior

• Doorway routing repaired: agent footsteps follow traversable doorway paths instead of cutting through adjacent walls or floor boundaries.
• Continuous movement and facing: waypoint transitions remain continuous, and look-back facing is held correctly instead of snapping prematurely.
• Unreachable desk recovery: agents wait in place while real work remains active, retry routes and recover when a refit restores access. Moving a desk no longer teleports the Overseer to it.
• Correct desk-facing poses: seated crew face their desk, using an appropriate directional pose when a skin lacks matching typing art. Working agents without a seat remain standing.
• Furniture cleanup: moving, rotating or deleting furniture releases stale seating claims, offsets and leisure references. Unchanged couches keep their occupants when the station origin shifts.
• Retired crew releases seats: removing plan-derived crew frees its furniture reservation so another agent can use the seat.
• Room deletion stays coherent: intersecting furniture and belt tiles are removed in the same undo snapshot, while neighboring furniture and logical agent links are retained. Undo, redo and reload preserve the intended result.
• Approval waits are visible: crew awaiting permission stand with the waiting-for-approval state and suppress work effects. Matching responses, concurrent prompts and terminal events clear the correct wait.

Images, usage and speech

• Image costs reach run receipts: actual image charges are included in local usage and completed-run costs without charging a second time.
• Image cancellation stops publication: cancellation propagates through generation, download and processing so a later result cannot publish a file after cancellation. Up...

StarNet v0.11.1

Download

Platform	Installer
Windows (x64)	StarNet_0.11.1_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.11.1_aarch64.dmg
macOS (Intel)	StarNet_0.11.1_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.11.1

This update improves linked-account reliability, voice playback, and everyday station controls.

• Fixed managed chat using stale per-request server settings or credentials instead of the linked StarNet account. Custom providers keep their own connection settings.
• Fixed image generation rejecting a linked StarNet credits account. Ordinary wording is less likely to trigger an unnecessary image-generation requirement.
• Saved room equipment now remains available when a chat request omits placement data, and tool availability descriptions better match the actual grants.
• Spoken replies buffer complete sentences and recover audio in order, with clearer interruption diagnostics.
• Background replies preserve the conversation you are currently viewing. Inbox rows have more readable message previews.
• Fixed work ratings on resumed legacy stations, retaining the station identity through restart.
• The glass interface is now standard across COMMS, model selection and station windows. Window interactions, text readability and Appearance previews have been improved.
• Recipe Bay now offers 45 curated workflows with a simpler launch flow. Recruitment, abilities, settings, profiles, quests, channels, automation and task outputs have clearer menus and controls.
• File-change approvals and work feedback provide clearer action and run references.
• Station lighting, wall finishes and floor materials have been refreshed. New starter stations use a white shell, and prop details leave more room for browsing the build catalog.
• Global stop controls have been removed, with recovery for their legacy pause state. Individual work controls remain available.

Update through Settings > Updates > Update Center. After restarting, confirm that the installed version is 0.11.1.

Google Workspace connections remain deferred while Google verification is completed. Saved connections are retained. Google login for StarNet billing is separate.

Release verification and known limitations

Windows upgrades from 0.10.13 and 0.11.0, clean installation, and close/reopen behavior passed. Mac signing, notarization, and Intel installed-app acceptance passed.

This urgent release proceeded with the release owner's extended-soak and aggregate-readiness exceptions. Existing customer reports remain under investigation; the included fixes do not establish recovery for every affected account or device.

Complete source changes · Official signed build

StarNet v0.11.0

Download

Platform	Installer
Windows (x64)	StarNet_0.11.0_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.11.0_aarch64.dmg
macOS (Intel)	StarNet_0.11.0_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.11.0

StarNet 0.11.0 brings collaborative group conversations, more useful task starters, personal goal progression, connected-app widgets, improved live voice, a substantial station visual refresh, and extensive work on recovery, permissions, persistence and everyday usability.

This changelog covers the shipped changes since v0.10.13, grouped by their final behavior. Intermediate designs that were replaced or reverted are not advertised as additional features. The complete source comparison, including maintenance and test changes, is linked below.

Fixes

This section collects the bug fixes described throughout the feature sections below. Outstanding customer reports and platform acceptance gaps remain listed under Deferred integrations and known limitations.

• Message attachments: sent group images and files now appear on their own message, with previews and durable associations through retry and restart.
• Conversation retry: retry retains the original user request instead of substituting a subsequent question or answer; obsolete retry choices are cleared when restoring a conversation.
• Agent names and identity: cached messages refresh after roster changes, and group mentions resolve stable agent identities.
• Recipe Bay layout: recipe cards no longer collapse into unreadable narrow columns.
• Context cards: context extraction panels use the current theme instead of a mismatched flat-black surface.
• Bay labels: compact name tags stay attached to the top of each bay and shrink with station zoom, replacing the oversized floating-label behavior.
• UI readability: service cards, console text, toggle switches and balance records have clearer presentation; narrow Refit layouts keep DONE accessible.
• Station colors and geometry: UI theme changes preserve the station color grade; wall corners, window frames, hallway mouths and doorway sorting receive corrections.
• Draft preservation: setup excursions retain unsent source material within the current app session, and switching quest selections preserves unsaved evidence.
• Recommendations: saved feedback loads before generation, outdated context invalidates suggestions, and impression-count changes no longer cause unnecessary regeneration.
• Goal accounting: repeated confirmations and replaced metric values cannot award the same achievement twice; stale quest responses cannot overwrite a changed goal context.
• Routine creation: success requires scheduler readback, and lost acknowledgements preserve a recoverable draft instead of producing an unsupported success claim.
• Scheduled results: results return to the captured session, restored sessions retain actual outcomes, and routine widgets refresh after scheduler edits.
• Emergency-stop recovery: Resume Automation clears persisted routine, Night Shift and loop halts explicitly, preserving permissions and job settings and exposing partial failures.
• Loop resume races: late check or harvest results from paused work cannot interfere with a resumed iteration.
• Cancellation: queued tool dispatches and language-server startup check cancellation; group cancellation is serialized, and failed cancellation acknowledgements remain visible.
• Usage settlement: partial usage is settled once across cancellation, retry and fatal paths.
• Group questions and handoffs: question drafts and cancelled state survive restoration; duplicate handoffs are bounded, internal markers stay out of chat, and coordinator failures are surfaced.
• Voice continuity: delayed recognition cannot overwrite a newer call, repeated words survive recognition-window boundaries, captions remain steady, and interrupted replies reject late audio.
• macOS microphone packaging: signed builds include the microphone capture entitlement repair. Physical Apple Silicon microphone acceptance is still unverified.
• Widget consistency: source outages preserve the previous reading, stale updates cannot overwrite edited definitions or recreate deleted widgets, and malformed optional links do not discard valid readings.
• Connector recovery: concurrent authentication failures share reconnect work, pending calls are drained safely, expired MCP sessions reinitialize, and rejected OAuth grants no longer appear authorized.
• Provider and account state: managed model selection follows the actual provider, malformed OpenRouter tool-call history is repaired, and stale paid-account recovery responses are fenced. Historical affected-customer recovery remains unconfirmed.
• Browser and event recovery: persistent browser shutdown allows cookies to flush, reconnect suppresses duplicate events, and renderer mutation commands are not replayed.
• Backups and deliverables: appearance backups retain complete preferences, connector exports exclude secrets, import results stay visible, and workshop builds reject missing deliverable files.
• Startup and backend failures: critical module failures produce a visible banner, corrupt read stores fail explicitly, and backend crash-loop handling provides bounded restart recovery and a visible degraded reason.

Starting work, Recipes and personal recommendations

• Create a useful first draft from actual source material: client updates, meeting actions, message replies, or a custom result.
• Use pasted notes or an approved source folder; the selected source follows the request into execution.
• Keep unsent requests, samples, source mode and folder selections when visiting setup and returning during the same app session. Explicit clearing and successful launch clear the draft; reloading the app does not promise to preserve these unsent form drafts.
• Gather task context through an adaptive conversation, with follow-up questions and editable context rather than a one-size-fits-all intake form.
• Restored task conversations retire obsolete retry choices, and retry uses the original user turn rather than accidentally replacing it with a later question or answer.
• Session recommendations use current requests, previous results, active goals, learned preferences, available capabilities and your feedback.
• Recommendations explain their proposed result and supporting context; continuations require unfinished work rather than recency alone.
• New requests and goal changes invalidate outdated suggestions. Dismissals persist, feedback loads before generation, and impression-count changes do not trigger unnecessary regeneration.
• When there is insufficient personal context or the model is unavailable, three clearly general starters remain available: build a working tool, remove a recurring burden, and pressure-test a major decision.
• Selecting a starter prepares the relevant session and brief; sending remains an explicit action.
• Internal recommendation runs are excluded from unattended user-work returns and from evidence about the user's own activity.
• Recipe Bay uses standard station windows, compact catalog headers and search fields, and readable card widths that no longer collapse into narrow strips.
• Context extraction cards retain their outline while using the current station panel theme.

Personal goals, quests and Commander progression

• Register a personal goal with a success condition and actionable steps, or add a success condition to an existing plan.
• Select an active goal to focus recommendations and quest generation.
• Record real-world actions you performed yourself as well as work performed with StarNet.
• Track goal metrics and inspect the evidence behind recorded achievements and Commander points.
• Distinguish a completed plan from an achieved real-world goal. Outcome confirmation is explicit and includes evidence.
• Extend a completed plan when further work remains.
• Generate next steps using goal criteria, metrics, prior results and blocker feedback.
• Defer a quest, mark it blocked, ask for a smaller step, or resume it.
• Keep quests from other goals without allowing them to consume the focused goal's recommendation slots.
• Discard stale recommendation responses when the goal or planning context changes during generation.
• Preserve goal registration across a lost response and keep active plans when trimming history.
• Deduplicate achievement awards and metric checkpoints across retries, restarts, metric replacement and repeated confirmation.
• Preserve existing earned history without inventing retroactive points; user-reported actions do not credit an agent with doing the user's work.
• Separate Commander goal progress from crew development feedback. Capabilities are not locked behind Commander levels.
• Refresh the quest log into a Commander journal with compact typography, aligned expandable progress sections, and shared station menu materials.
• Preserve unsaved quest evidence when switching journal selections.

Finding useful work and making it recur

• Opt into bounded discovery from an explicitly selected approved folder of Markdown, text or CSV notes.
• Surface client-update opportunities with filenames, line references and excerpts from the selected material.
• Pause or remove discovery sources; sources and findings persist across restart.
• Recheck pause state during docu...

StarNet v0.10.13

Download

Platform	Installer
Windows (x64)	StarNet_0.10.13_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.10.13_aarch64.dmg
macOS (Intel)	StarNet_0.10.13_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.10.13

This is the full release list for the 214 commits shipped since v0.10.12. It is a large feature-and-reliability update, not a narrowly scoped hotfix.

LINES: a complete automation library

• The LINES shelf expanded from 4 starter layouts to 19 compiled, one-click systems.
• The library is grouped into Chains, Sorters, Crews, Quality Gates, and Flagships, with short explanations that make the progression easier to understand.
• Every Line uses the same floor compiler and routing mechanics as a Line built manually. Stamps are validated as a whole and remain one-step undoable.
• FRONT DESK provides the complete INBOX → agent → OUTBOX round trip in one row.
• ALLOWANCE DESK adds a hard per-day and per-message spending ceiling directly to the intake.
• TWO DOORS lets two independently named and budgeted intake lines share one crew member and one shipping door.
• RESEARCH LINE hands work from a researcher to a writer before shipping it.
• SHIP-OUT LOOP is the smallest useful Line: one crew member connected directly to an OUTBOX.
• ASSEMBLY LINE runs a four-stage research → analysis → writing → shipping handoff.
• SORTING OFFICE routes code work to an engineer and everything else to a generalist.
• TRIAGE DESK sorts code, research, and general work into three specialist lanes and merges the results into one OUTBOX.
• PARALLEL CREW distributes incoming jobs across three crew members.
• SECOND OPINION sends the same job to two crew members and joins both independent answers.
• RESEARCH SWARM fans one job out to three researchers, waits for every branch, and gives their combined work to an analyst.
• LOAD BALANCER alternates jobs between two desks instead of duplicating the work, then funnels both result paths into one OUTBOX.
• REVISION LOOP sends a draft through a reviewer and back to the writer until it is approved, with a three-pass ceiling.
• CODE FOUNDRY sorts code into an engineer/reviewer revision loop while general work takes a separate lane.
• FIRE ESCAPE adds a third escalation exit: approved work ships, rejected work loops, and work that exhausts its passes goes to a fixer instead of ending as a bare failure.
• THE GAUNTLET combines fan-out, join, analysis, review, and a verdict-gated revision loop.
• THE CRUCIBLE uses two separate approval gates: one for the draft and another for the polished result.
• MISSION CONTROL gives code, research, and general work distinct multi-stage routes before merging everything into one shipping door.
• THE DEEP DIVE runs a three-researcher swarm through analysis, writing, review, and revision before shipping.
• Cascaded fan-out now launches every promised branch instead of stopping after the first layer.
• Loop back lanes render and simulate from the compiled route, including their re-entry tile, so a working Line no longer looks frozen or broken.
• Loop gates gained the real third escalation lane used by FIRE ESCAPE.
• The conveyor simulator now follows the compiled done/back/escalation lanes instead of reconstructing a different route.
• Blueprint roles now include Reviewer, Analyst, and Fixer, so one-click crew suggestions match the actual stage being staffed.
• Footprint checks, routing validation, powered-belt checks, and no-room warnings were expanded across the entire 19-Line catalog.

Found work and smarter recommendations

• The new FOUND ON YOUR PROJECTS shelf can surface useful work already present in projects you explicitly blessed.
• Discovery is deterministic and local: it scans only approved roots and uses exact project text as the citation behind each finding.
• Every discovery card shows why it exists and lets you hand the cited task to a crew member without automatically starting work.
• Dismissals are arm-confirmed, durable, and denylisted so the same finding does not immediately return.
• Discovery respects pause state, cooldowns, trust boundaries, and finding expiry.
• Discovery decisions are recorded in the recommendation ledger instead of becoming invisible background behavior.
• Night Shift can cite the same project findings as evidence, but discovery never changes Night Shift scoring or grants itself access to another project.
• Completed runs now fold into a support-gated station track record.
• Recommendations, quests, and Night Shift can use literal success counts from that track record without overstating sparse history.
• Ranking uses a conservative prior until enough evidence exists, rather than treating one lucky result as certainty.
• /api/insights exposes the same evidence counts used by the recommendation system.
• Track-record evidence yields to more immediate activity when the recommendation evidence budget is full.
• Discovery refreshes when its bay opens, persisted sweep results survive restart, expired ledger rows fall away, and repeated track-record reads are memoized.
• Previously dead recommendation seams now record declines, outcomes, seed/routine answers, and expiring impressions correctly.

Connector and KEYS expansion

• Added 18 live-probed OAuth connectors: Todoist, ClickUp, Railway, Grafana, PostHog, Cloudflare Bindings, Cal.com, Fireflies, Algolia, Buildkite, Datadog, Globalping, Honeybadger, Jam, Sanity, Semgrep, Close CRM, and Ramp.
• Added a keyless OpenAI DevDocs connector.
• Added 23 specialized KEYS directory entries: Supliful, Fourthwall, ShineOn, SPOD, Teemill, Zazzle, Gumroad, Lemon Squeezy, CJ Dropshipping, Keepa, Discogs, Lob, Lulu, Duffel, Porkbun, ElevenLabs, HeyGen, Bannerbear, Shotstack, Bland, Vapi, DeepL, and Transistor.
• The new catalog entries were live-probed before inclusion instead of being added as unverified placeholders.

Billing, providers, and tool execution

• Bring-your-own-key runs on a linked station no longer require StarNet credits and are never charged against the StarNet wallet.
• Managed runs that cross a reservation now settle the real spend at that reservation instead of being refused after the work already happened.
• Billing leak-guard replays settle against the remembered real spend on every exit path.
• Gemini 3 thought signatures are preserved when function calls are replayed.
• Gemini usage is emitted before completion, matching the ordering used by other providers.
• Index-less parallel OpenAI-compatible tool calls keep distinct identities instead of collapsing into one slot.
• Name-echoing continuation deltas remain part of one tool call instead of creating duplicates.
• Durable tool results resume correctly after an interruption.
• Orphaned pristine transcript segments are no longer created during recovery.
• Short verification reads run to completion instead of being mistaken for truncated output.
• Tight output budgets retain the do-not-repeat instruction and no longer grow while trying to explain their own truncation.
• Ollama receives a realistic cold-start window for local models that take longer than 30 seconds to load.
• An MCP server's own tools named resources or prompts are no longer shadowed by StarNet's auxiliary browsing definitions.
• A clean but unexpected exit from an MCP stdio server is treated as a transport failure rather than a healthy idle state.
• LSP startup can retry after a spawn failure.
• Skill-exchange handlers return a real 400 for malformed JSON instead of leaving the socket hanging.
• Tool registry capability and schema refusals remain contained as tool errors rather than escaping the dispatcher.
• Protected mutations wait for their idempotency record before a loop advances.

Routines, schedules, and loops

• Completed one-shot routines now show truthful finished controls instead of a dead ENABLE action.
• One-shot routines can be explicitly re-armed.
• RUN NOW takes a real lease, settles its execution record, and cannot double-launch from repeated clicks.
• Pausing or deleting a routine aborts its active run and releases the associated state cleanly.
• A failure between loop context creation and process launch now settles the iteration instead of orphaning it.
• A Line advances only after its entry run is actually completed.
• One routine throwing during a scheduler tick no longer prevents the remaining routines from firing.
• Reclaimed zombie jobs record an outcome and release their concurrency slot.
• Retry anchors are cleared whenever a schedule is deliberately re-anchored.
• Transient retry backoff no longer permanently shifts an interval schedule.
• Backward system-clock changes no longer wedge an in-flight lease.
• DST spring-forward gaps that cross local midnight are detected correctly.
• Malformed and NaN schedule values are rejected or normalized instead of poisoning the scheduler.
• Impossible day/month combinations return immediately, and sparse-date searches skip by day instead of blocking the event loop minute by minute.
• Empty-prompt routine creation is refused.
• Routine creation, loop creation, and paid automation buttons have in-flight guards against double submission.
• Running badges, pause/resume state, RUN NOW state, and creation refusals now reflect the scheduler's proven state.
• Schedule guard sanitization strips GitHub-auth constructs before guarded work is launche...

StarNet v0.10.12

Download

Platform	Installer
Windows (x64)	StarNet_0.10.12_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.10.12_aarch64.dmg
macOS (Intel)	StarNet_0.10.12_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.10.12

• StarNet no longer progressively degrades into a laggy software-rendered station after several minutes or after minimize/restore. The stage and bake watchdogs now sample through a dedicated readback canvas instead of repeatedly reading the visible canvas, avoiding Chromium's software-raster fallback.
• Stage recovery no longer accumulates duplicate window/document listeners after a GPU reset. Global recovery listeners are bound once, while replacement-canvas listeners remain scoped to the canvas they belong to.
• Streaming COMMS replies now coalesce prose rendering to at most once per animation frame, and long transcript DOMs are bounded while preserving the newest messages. Complete markdown, links, bullets, ordering, and close/error flushes are retained.
• Account links can repair themselves after a reinstall when the secure device token survived but the local link record did not. Explicitly unlinked or revoked stations stay unlinked.
• Fresh installs no longer mistake normal shutdown/E-STOP bookkeeping for prior station data, avoiding a false recovery gate after a first open and close.
• Managed-model requests with an invalid or endpointless provider configuration now refuse with the real setup remedy instead of silently routing to the wrong API endpoint.
• Codex/ChatGPT conversations recover from an orphaned tool-result record instead of leaving that chat permanently stuck on a Responses API error.
• Added Higgsfield to the connector catalog with its OAuth sign-in flow, and added OAuth as an authentication option for custom HTTPS MCP servers entered by URL.
• Custom connector OAuth preserves configuration across browser consent and restart, requests provider-advertised scopes, and keeps submitted bearer values out of OAuth connector storage.
• The release train warms its browser proof before the gate begins, removing the startup race that interrupted the automatic 0.10.10 train.

This update includes the complete v0.10.11 change set plus the performance-degradation repair above. It does not claim to reconcile credits purchased under a different web account than the one linked to the station; that account-identity case remains a separate support investigation.

This focused hotfix waives the normal 48-hour RC soak because it repairs a shipped degradation affecting v0.10.10 and v0.10.11. Before publication it still requires the exact installed-desktop smoke, a sustained installed minimize/restore run, full fast/HTTP gates, the signed release train, notarized macOS artifacts, and hosted T0/G1 packaged checks.

StarNet v0.10.11

Download

Platform	Installer
Windows (x64)	StarNet_0.10.11_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.10.11_aarch64.dmg
macOS (Intel)	StarNet_0.10.11_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.10.11

• Account links can now repair themselves after a reinstall when the secure device token survived but the local link record did not. Explicitly unlinked or revoked stations stay unlinked.
• Fresh installs no longer mistake normal shutdown/E-STOP bookkeeping for prior station data, avoiding a false recovery gate after a first open and close.
• Managed-model requests with an invalid or endpointless provider configuration now refuse with the real setup remedy instead of silently routing to the wrong API endpoint.
• Codex/ChatGPT conversations recover from an orphaned tool-result record instead of leaving that chat permanently stuck on a Responses API error.
• Added Higgsfield to the connector catalog with its OAuth sign-in flow, and added OAuth as an authentication option for custom HTTPS MCP servers entered by URL.
• Custom connector OAuth preserves configuration across browser consent and restart, requests provider-advertised scopes, and keeps submitted bearer values out of OAuth connector storage.
• The release train now warms its browser proof before the gate begins, removing the startup race that interrupted the automatic 0.10.10 train.

This update does not claim to reconcile credits purchased under a different web account than the one linked to the station; that account-identity case remains a separate support investigation.

This bug-fix update waives the normal 48-hour RC soak because it repairs active account-link, model, conversation, onboarding, and connector failures. The exact installed-desktop smoke, full fast/HTTP gates, signed release train, notarized macOS artifacts, and hosted T0/G1 packaged checks are still required before publication.

StarNet v0.10.10

Download

Platform	Installer
Windows (x64)	StarNet_0.10.10_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.10.10_aarch64.dmg
macOS (Intel)	StarNet_0.10.10_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.10.10

• Fixed several StarNet account and credit edge cases: funded accounts no longer fall back to stale or zero balances, temporary reservation holds no longer trigger false out-of-credit warnings, and failed or timed-out account reads remain clearly unverified instead of inventing a balance.
• Account recovery is more dependable. Revoked or incorrect links return to a usable relink flow, and START COMPLETELY FRESH can retry browser-state clearing without quarantining the newly cleaned station again.
• The KEYS panel now distinguishes a refused or unreachable station from a genuinely empty key list, so transport failures cannot appear as saved or held credentials.
• Fixed a rare permanently black station viewport after a WebView/GPU canvas-context loss; StarNet now rebuilds the visible stage and rewires its input automatically.
• Fixed dead clicks in OUTBOX caused by tiny pointer jitter, and widened the clickable area around shipped pallet crates.
• Added tighter time bounds around account-link and balance responses so a stalled response body cannot leave the station waiting forever.

This bug-fix update waives the normal 48-hour RC soak because it repairs active account-access, viewport, and interaction failures. The exact installed-desktop smoke, full fast/HTTP gates, signed release train, and hosted T0/G1 packaged checks are still required before publication.

StarNet v0.10.9

Download

Platform	Installer
Windows (x64)	StarNet_0.10.9_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.10.9_aarch64.dmg
macOS (Intel)	StarNet_0.10.9_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.10.9

• Fixed a StarNet account-linking bug that could leave a station claiming it was linked while the account website showed no linked station.
• A station whose link was removed from the website now clears any stale balance, stops reporting “linked” or “no credits,” and offers LINK YOUR STARNET ACCOUNT directly on the Create Your Overseer screen.
• Temporary StarNet service or network failures are now shown as unavailable instead of being misreported as a zero-credit account.

This is a narrowly scoped billing-trust hotfix. The normal 48-hour RC soak is waived because v0.10.8 can block newly arriving paid users before their first agent wakes; the exact regression, full fast/HTTP gates, signed-build train, and staged T0/G1 packaged checks remain required before publication.