Changelog

All notable changes to this project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

• Tasks show their id. The one thing people actually refer to a card by — bmt-12 — was not displayed anywhere: not on the kanban card, which printed only the title and the assignee, and not in the detail view behind it. It now leads the card above the title, and leads the detail view's fact row, in mono in the same muted ink as the assignee. Every id shows, including the synthetic t-xxxx fallbacks a card gets when it arrives without one.
• Fable 5.1, GPT-6 Astra, and Gemini 3.7 Flash in the model pickers. Claude Code gets claude-fable-5-1, Codex gets gpt-6-astra, and Antigravity gets Gemini 3.7 Flash at all three reasoning levels; Cursor gets its own build of the two it carries. Every id was read out of the CLI that will be invoked with it — codex-rs/models-manager/models.json, agy models, cursor-agent models, and the installed claude binary — rather than guessed from a name. GPT-6 Astra needs Codex 0.153.1 or newer. The slug landed in that release, so an older codex on your machine will reject it when the agent spawns; update Codex first. The catalog's version bounds cover the app, not the CLI it launches, so this is a prerequisite rather than something the picker can hide for you.
• A new model no longer needs a release. The pickers now read docs/model-catalog.json on main, fetched at runtime and cached for six hours, so adding a model is one line in one file on GitHub rather than a build. The check runs at startup, not on a timer: an installed copy picks a new model up the next time it launches with a cached copy older than six hours, and an app left open does not change under you. The catalog compiled into the build stays the floor: it is what renders offline, on first launch, and whenever the remote copy is missing, unreadable, or announces a schema this build does not know. A provider present in the remote copy replaces that provider's list; a provider it does not mention keeps the built-in one, so a bad edit costs a list rather than a picker. The payload is data and never markup, and a model id is length-capped and stripped of control characters before it can reach a --model flag on a spawn command line. Same mechanism as the Settings hero card.

[0.4.6] — 2026-08-27

The release that speaks your language and updates itself. The interface runs in Chinese and Arabic, the auto-updater downloads and installs a new build end to end, fonts ship inside the app so a blocked network never leaves you on a blank window, and the way agent engines are launched is hardened.

Added

• The interface speaks Chinese and Arabic. Every string is translated (nothing falls back to English) and the terminals read right to left under Arabic. The physical-direction sweep — padding and icons that still mirror the wrong way on some screens — is tracked as the next piece of work.
• message_sent telemetry. A single anonymous event closes the activation funnel by counting the messages a person sends to an agent — a count and nothing else: no text, no length, no content of any kind. Counted at the submit, never per keystroke, and suppressed by every existing opt-out. TELEMETRY.md lists it like every other event.
• The ASK ME card renders markdown. Questions arrived with their asterisks and backticks on screen, because the card printed the raw text. It now renders the same way the file preview does: emphasis, bullets, code, tables, and links that open in the browser instead of navigating the app. The task detail's Q&A trail renders too, answers included. A single newline is still a line break, so a question written as plain text looks exactly as it did before. Raw HTML is still shown as text rather than parsed, which is what keeps agent-written markdown safe to display.
• Agents are told to format what they ask you. The orchestrator prompt and PROTOCOL.md now ask for a bold lead line, backticks around paths and commands, and bullets whenever a question has more than one option.

Fixed

• The auto-updater installs a new build end to end. The title-bar badge advances from check to available to downloading to downloaded on its own, and the action at the end restarts into the new version. This is the release that proves that path — a build's own updater is only exercised by the next release.
• Fonts ship inside the app. The renderer and the release drop no longer fetch Google Fonts at launch, so the app opens at the same speed on any network, including one where Google is blocked. The release drop can no longer white-screen while a stylesheet loads: its fonts are bundled, its content-security policy denies a remote stylesheet outright, and a loader covers the frame until it paints.
• Settings has one Save button. The Connections tab stops repeating itself; the REST API, MCP, Slack and webhook sections each keep their place.
• IME typing no longer sends early. Pressing Enter to choose a Chinese or Japanese candidate picks the word instead of firing the message with half-typed text.

Security

• Engine command launching is hardened. The name of the CLI an agent launches is validated before it is resolved against your PATH, so only a plain command name or an absolute path reaches a shell.

Thanks

16 community pull requests from 13 contributors landed in this release, one of them (#213) re-implemented rather than merged:

• #156 @gpechieu: the roster empty-write guard stays armed after a refused write
• #205 @Schopenhauer-loves-Hegel: the react-i18next multilingual UI foundation, shipping the Chinese translation
• #213 @abo123v-glitch: the Arabic/RTL terminal rendering recipe (re-implemented into this release)
• #225 @jhinzzz: Codex hive sessions are visible to usage scanners
• #242 @raifemre: the localStorage roster fallback is scoped to the hive that wrote it
• #243 @L422Y: the ASK ME question renders as markdown
• #248 @djbiz: a floor rebuild that cannot get a WebGL context retries instead of failing
• #270 @BUGHUNTER-SACHIN: /compact is not enqueued for an undeliverable agent
• #271 @HsienW: a typed hook-event payload contract
• #282 @savvaskoualis: a renamed god identity no longer reverts to "Michael"
• #284 @HundredBillion: the Settings save button moves into the modal footer
• #286 @HundredBillion: every config write is announced so Settings stops going stale
• #310 @LavaDMan: the hive-hook-node test no longer races its own stdin write
• #317 @aaroncoville: usage resets when a PTY exits
• #323 @aaroncoville: the WebGL context is released when a terminal's renderer tears down
• #339 @aaroncoville: the model catalog is brought up to date

[0.4.5] — 2026-08-22

The release that fixes the things you trusted and were quietly wrong. Cost reporting was off by more than half after a restart, semantic memory never worked on Apple Silicon, and agents could not talk to each other reliably. All three are fixed, alongside weekday scheduling, clickable paths in terminal output, one editor instead of two, and 23 community pull requests.

Added

• Weekday scheduling for triggers. A trigger can run at a time of day on chosen weekdays rather than only on a fixed interval, and the schedule is DST safe.
• Every path in terminal output is clickable. Markdown opens the preview, source and config open in the IDE, and images, archives and anything we cannot honestly render are revealed in Finder or Explorer instead of being opened. A path token is never handed to the OS "open with default app" call, so a printed installer.dmg cannot become an execution.
• Manual updates download on click, with install steps for the platform you are on. The title-bar badge is always the manual path, auto-update stays in Settings, and the badge says latest once a check confirms you are current.
• The first run after an update shows that release's page.
• A hero card at the top of Settings with your version, your plan, and a way back to the release notes.
• A 1:1 hold that tells Michael to leave an agent alone.
• Editable agent names, and Michael can choose a worker's avatar and accent when he hires.
• Gemini CLI and Cursor Agent join the engine list.
• Edit an agent from inside focus mode, next to its name.

Added

• update_applied telemetry. The app now reports, once, on the first start after its own version changes: update_applied { from_version, to_version }. Auto-update health was the one thing we had no way to see — a release is delivered in place, and the only evidence we had was installs that happened to show events on two versions. from_version is unknown for an install that predates the event, so the first release carrying it is measurable rather than silent for a cycle. Both values are version strings; nothing new about you is collected, the same opt-outs apply, and TELEMETRY.md lists it like every other event. A via property says whether the app's own updater installed it (auto), something else moved the version (manual), or there was no update log to read (unknown) — read from the update log the app already keeps, so it works for installs updating from a version released before this one. A restart the user asked for that then quietly installed nothing counts as manual, not auto, because the log names which build started next.

Fixed

• Cost is reported over the app's lifetime, not since the last restart. The telemetry counter reset on every launch while the session id stayed the same, so the floor under reported spend by 59 percent. It is now folded from the ledger, with a separate session figure alongside it.
• Semantic memory works on Apple Silicon. CoreML overflowed the quantized embedding graph, every vector came back NaN, and chroma rejected every upsert. Embeddings are pinned to CPU on macOS.
• Agent-to-agent messaging is reliable. A main-process watchdog wakes an idle worker sitting on an undrained inbox, stale nudges no longer fire at an inbox that has since been drained, mail to an id with no inbox is bounced and logged rather than dropped, the per-agent steer queue is capped, webhook dispatch is atomic, and PROTOCOL.md refreshes on boot.
• Restart to update no longer sticks when a running agent makes the app refuse to quit.
• Focus mode survives a restart, the close button no longer drops you to the sidebar, and the focused terminal refits when the roster changes.
• Terminals follow the window theme. Running TUIs are told when the theme flips, OSC colour queries are answered, and Crush and OpenCode are handed the theme at spawn.
• A partial ANSI escape carries across pty chunks instead of printing as garbage.
• Windows agent processes quit when the app does.
• Buttons in a squeezed row no longer paint over their neighbours, and the sidebar header drops its button labels before it drops the agent's name.

Changed

• One editor. The fullscreen file overlay is gone. Markdown, source and images all open in the IDE, whose git rail is now collapsed by default and carries a git mark.
• The renderer runs inside Chromium's sandbox.

Thanks

All 23 of these community pull requests landed in this release:

• #157 @gpechieu: inherited Claude Code session markers are stripped from an agent's PTY env
• #158 @gpechieu: semantic memory works on Apple Silicon again: embeddings are pinned to CPU on macOS
• #159 @gpechieu: reliable spawn, teardown and floor cards for the workers Michael hires
• #165 @rajpreetcodes: a ~ in the harness home folder resolves, so setup cannot die on ENOENT
• #171 @KrushanPatel: CONTRIBUTING.md matches the platforms the app actually supports
• #175 @rekcilyssup: a main-process watchdog wakes an idle worker sitting on an undrained inbox
• #176 @FenjuFu: Gemini CLI joins the engine list
• #177 @TTAWDTT: each agent's live context-window occupancy shows in the roster
• #178 @gpechieu: a god-hired worker gets a floor card, and it archives when the worker dies
• #179 @kdahal7: statAbs expands ~, so a path resolves the same way on every platform
• #181 @TTAWDTT: webhook dispatch goes through an atomic add, so a stale ledger cannot overwrite it
• #184 @TTAWDTT: the per-agent steer queue is capped, which bounds memory on a stalled agent
• #185 @hyperstream-pro: mail to an id with no inbox is bounced and logged instead of dropped
• #186 @BUGHUNTER-SACHIN: tests cover the Notifications and Stop idle-detection branches
• #187 @hyperstream-pro: a stale inbox nudge no longer wakes an agent against an inbox that is already empty
• #190 @swarnendu19: agent names can be edited after spin-up
• #199 @amey-op: the Antigravity queue no longer wedges for 30 seconds
• #203 @lifelmy: the Crush config env points at the agent's own directory
• #210 @chaitanyagiri: the art licence claims are true again, Modern Interiors is bought
• #214 @pontusm: Windows agent processes quit when the app does
• #219 @chaitanyagiri: engine availability is checked before Michael's engine is committed
• #226 @chaitanyagiri: the floor reports lifetime spend, not spend since the last app restart
• #227 @scy73: the renderer runs inside Chromium's sandbox

[0.4.4] — 2026-08-18

Windows agents can finally talk to each other — and the first run stops silently failing. Two bugs made the core product not work on the platform that accounts for roughly half of all downloads, and a third meant a brand-new install never started the services that carry messages between agents. Alongside them: a rebuilt dark mode, a Skills browser, a Prerequisites page, and release notes that can carry their own designed page.

Fixed

• Agent-to-agent messaging on Windows. The hive protocol reaches an agent as a multi-line command-line argument. A .cmd cannot go to CreateProcess, so any non-.exe target ran via cmd.exe /d /s /c "…", and cmd.exe cuts an argument at its first newline — taking the block that names inbox/ and outbox/ with it. Agents booted, rendered, looked healthy, and had no idea they could message anyone. Prompt-carrying spawns now decode the npm shim and launch its real interpreter with an argv array; anything undecodable falls back to the previous behaviour.
• Windows OpenCode specifically. opencode-ai's bin is a compiled binary, so npm writes an interpreter-less shim that the first fix did not model — it returned null for every Windows OpenCode install and fell back to the truncating path. Direct-executable shims are now handled, and the previously silent fallback logs the target it could not decode.
• A fresh install never started its hive services. bootstrapHiveServices() early-returns while harnessHome is null, which is exactly the state a first run boots in; onboarding then set the home without re-bootstrapping. The message router, hook server, telemetry collector and mission scheduler stayed dead for the whole session — mail never moved and agents never reported. Now bootstrapped on the null → set transition.
• The setup wizard could not be finished. ~/HarnessAgents persisted with a literal ~ and died on ENOENT: mkdir; the folder field also never pre-filled, because it read window.process.env.HOME, which is always undefined under contextIsolation. An empty folder now fails at step one instead of after step four, and the panel no longer overflows a short screen.
• "Restart & Continue" had nothing to resume. The live session id is now recorded from a second source, so continuing works even when a hook never lands.
• Dark mode was unreadable in a specific way. ink-300 measured 1.73–2.09:1 against every surface, and it is the structural token — 187 uses, 93 as 1px borders — so every control's edge was invisible and the UI read as flat grey. Now 3.4–4.0:1, on a softer ground with warm off-white text. The selected Command Center tab measured 1.55–1.87:1 (near-white on a light accent); a new --cth-on-accent token takes it to 7.0–8.5:1.
• OpenCode ran a model you might not own. It preselected a BYOK slug and silently fell back when the key was absent, while every surface kept reporting the model it had asked for.
• Terminal copy strips the CLI's quote rail and agent terminals run in UTF-8; dictation pastes what was just said; task-ledger mutations are atomic; a frozen context reading no longer re-fires /compact hourly; one odd message id no longer silences an agent's wake nudge; the cost ledger stays out of the hive's git history; a root cwd no longer resolves to the projects directory; the office floor stops rendering when nobody is looking at it.
• Agent selection is visible on every card including Michael's — it used to be drawn in each agent's own accent and was invisible on the one card that was always framed.

Added

• Skills — installed skills across Claude Code, OpenCode and Codex with scope precedence, plus a browsable catalog of 227 with search, category and publisher filters, install and uninstall. Installs are bounded and containment-checked; uninstall refuses anything that is not a SKILL.md folder inside a managed root.
• Prerequisites (Settings) — live status for uv, git, Node, MemPalace and every agent engine, with real paths, platform-correct install commands, and a button that asks Michael to fill the gaps.
• Release drops — a release body can carry an authored HTML page, rendered in a sandboxed iframe (sandbox="" + default-src 'none') as a centered modal.
• Settings hero card — version, plan, sponsor and a way to reopen the release notes, with its contents fetched from docs/hero.json so they can change without a build.
• IDE image preview (PNG/SVG/markdown embeds), agent-named title, real shortcut hints.
• The update notification says what changed, and asks for a star at most once ever.
• Grok 4.6 in the model picker.
• Dictation setup guidance behind an info mark on the voice button, including that Groq is free.
• pause and halt explain what they do on hover.
• Fullscreen gained open (terminal at the agent's cwd) and ✕ (end + archive), and its roster cards now show model, project and a context gauge.

Changed

• One card size for every agent; Michael is distinguished by surface, not by a heavier border.
• Command Center tabs wrap when docked and scroll in fullscreen; the commands tab was removed.
• Prerequisites moved out of the Command Center into Settings — it is machine-wide state, not something about the agent whose terminal you are reading.

Thanks

Community fixes in this release: @gts-47 (#129, #130, #131, #132, #133, #134, #143, #144) and @baziyer (#142).

[0.4.3] — 2026-08-13

A new brand mark: Michael's portrait replaces the "MD" tile. The logo is now the character the product is about, drawn in the app's own pixel art on the brand yellow. It is authored as pure vector (docs/logo.svg) and every raster is generated from that one source by tools/make-logo.cjs, so the site, the app and all three platform icons can no longer drift apart. Appearance only — no functional change.

Changed

• Logo replaced everywhere — build/icon.{svg,png,ico,icns}, docs/logo.svg, docs/logo.png, docs/logo-light.png, the site header and favicon, the in-app toolbar and window icon, and the README header.
• App icons are natively multi-resolution — .icns spans 16→1024 with the macOS drop shadow; .ico carries 16/32/48/64/128/256. Previously both were built around a single 1024px raster.
• Site CTA buttons are brighter by default. .btn.primary drew its fill from --accent, which also colours accent text and so must stay dark enough to read on a white page — the light theme's #E5A00D made the download button read brown. Fills now use dedicated --accent-fill / --accent-fill-hover tokens, starting at the old hover colour.
• The Product Hunt thumbnail (docs/media/ph-thumbnail-240.gif) sits on the brand yellow instead of a pale tint, matching the new mark.

Added

• tools/make-logo.cjs — generates the SVG plus every PNG, the .ico and the .icns from the sprite in src/renderer/src/scene/office/portraitArt.ts. No external image tooling.
• docs/favicon-32.png and docs/apple-touch-icon.png — native-size icons, so browsers stop downsampling a 512px portrait.

Fixed

• docs/llms.txt advertised 0.4.1 two releases after the fact. tools/check-release-links.cjs now checks it, so it cannot silently drift again.
• README version badge and status note had been stuck at 0.4.0 since two releases ago.

[0.4.2] — 2026-08-13

Anonymous usage stats — documented, opt-out, and off in forks. The project previously had zero insight into whether anyone launches the app or which features get used. This release adds a minimal, anonymous product-analytics layer (PostHog), governed by a public contract: TELEMETRY.md lists every event and property, and the code enforces that list as a hard allowlist.

Added

• Anonymous usage events (src/main/analytics.ts): first_run, app_launched, agent_spawned (engine name only), feature_used (fixed enum, once per session), and session_ended (coarse duration bucket). Common properties are app version, OS, and CPU arch — nothing else. No prompts, transcripts, file paths, repo names, or identifiers of any kind; events are PostHog anonymous events ($process_person_profile: false), keyed by a random install UUID that lives in the app's user-data dir.
• Consent surfaces: a "Share anonymous usage stats" toggle on the final onboarding step and in Settings → General (telemetryEnabled, default on = opt-out). The standard DO_NOT_TRACK env var is respected unconditionally.
• TELEMETRY.md — the complete public contract, linked from the README.

Note

The PostHog key is injected only in release CI (POSTHOG_KEY secret). Building from source or forking the repo produces a build with no key, and the entire analytics module is a no-op — a fork never sends events anywhere.

[0.4.1] — 2026-08-13

The app says what the site says. munderdiffl.in describes Munder Difflin as a clone of you that works around the clock; the app still called it a "GOD agent." This release closes that gap. Wording only — no behaviour changes.

Changed

• Michael is your clone. Onboarding refers to him as your clone throughout, and his card on the floor now carries a BOSS tag instead of GOD — he's the boss of the agents, you're still the boss of him.
• Onboarding leads with the product, not the feature list. The first screen opens on "a clone of you, working 24/7"; step 2 is "your clone's engine."
• The engine card names all ten engines. It had advertised three (Claude Code, Antigravity, Codex) since before seven more shipped — Grok, Kimi, Qwen, OpenCode, Crush, pi and Copilot are now named too.
• Site copy fixes. A misspelled "requrired" in the hero, cli → CLI, a pricing FAQ that named plans ("Basic"/"Pro") the pricing table doesn't sell, and an interactive demo card that claimed a clone ran on Cursor — which is not one of the supported engines.
• README back in sync. It listed nine engines (Qwen was missing in four places) and still reported v0.3.8.

Note

The god agent id, the hive folder layout, and message routing are unchanged. Existing hives, memory, and running agents carry over as-is; there is nothing to migrate.

[0.4.0] — 2026-08-12

The brand grew up — and the landing page with it. Munder Difflin now looks like one product everywhere: a yellow "MD" mark, matching app icons on every platform, and a rebuilt munderdiffl.in that shows the real app instead of describing it.

Added

• Real app screenshots on the landing page. The Add Agent dialog, the memory panel, and Settings → Autonomy & Budgets — captured from the actual app, not mockups.
• A live demo video in the hero. The old static home-screen shot is now a looping screen recording of the office floor with a live agent terminal.
• Clone-to-clone chat and encrypted-wire visuals with request-a-demo contact points for the Teams features.

Changed

• New app icon. The dock/taskbar icon is now the yellow "MD" tile on every platform (macOS .icns with proper margins and shadow, Windows .ico, Linux .png), matching the in-app logo and the site favicon.
• Landing page defaults to bright (light) mode. Dark stays one click away and remembered.
• Landing accent is yellow again — bright amber in light mode, gold in dark — applied across buttons, diagrams, the Pokédex-style floor sim, and the pricing cards.
• Pricing reframed around the two services — Private Cloud (a dedicated sandbox VM per clone) and Private Network (E2E-encrypted clone-to-clone wire) — with a team-size slider.
• Social previews refreshed. New Open Graph card and copy that match the current product.

[0.3.9] — 2026-08-11

Ask the app whether it's up to date. Settings → General now names the running version, says whether it is the latest, and offers one button that states what it does. Shipped now because 0.3.8 needs to reach people who already installed it.

Added

• Check for updates, in Settings. A block at the top of Settings → General that always answers "am I on the latest?": the version you're on, whether a newer one exists, and a single button — Check for updates → Download v0.4.0 → Restart to update — plus the verbatim error when a check fails. It shares the status stream, the reducer and the state machine with the toolbar chip, so the two can never disagree about what is installed; only the wording differs, because a chip that must stay quiet when everything is fine is no use to someone who opened Settings to ask.

Changed

• Fullscreen roster avatars are larger. They were rendered at 1× — an 18-pixel figure — and the tile width was free to grow past that, so a wider roster just padded the same small sprite. Portrait size now moves in half-sprite steps and the art is drawn at the tile's width.

Removed

• The usage-limit hold, entirely. Shipped in the first v0.3.8 tag; it did not release. Agents held behind a limit stayed held — the stated reset never arrived, and the manual resume now button returned them to the held state instead of draining their queue. Delivery behaves as it did in 0.3.7. rateLimit.ts, limitGate.ts, useLimitWatch.ts, the banner, the Settings section, the config keys and their IPC are all gone, along with the compaction gate that had been layered on top of them.

[0.3.8] — 2026-08-11

Memory condensation works for the first time. The harness had been reading Claude Code transcripts from a directory that has not existed for months, so the summarizer never had anything to summarize — and nothing errored, because an absent directory reads as "no transcripts yet". That is the headline fix. Behind it, a run of things that had been quietly costing tokens or hiding in plain sight: compaction firing on two schedules at once, a commit history that rendered no commit messages, and buttons whose labels were invisible in dark mode.

Added

• Triggers hub. Schedules, inbound webhooks, context rules and peer messaging now share one home in Settings, with a history of what fired and what it did.
• Collapsible panels. The IDE's git rail folds away to give the file tree its height back, and the fullscreen roster folds for a full-width terminal. Both remember the choice.
• The OpenAI key can be set where voice is explained. Settings → Voice now carries the field itself, names the model it pays for (gpt-realtime-2.1), and the disabled Talk button links straight to it.

Fixed

• Claude Code transcripts were read from a directory that has not existed for months. projectDir() built the pre-2026 project key — leading slash dropped — while Claude Code dashes every non-alphanumeric character. Nothing errored, because every caller reads an absent directory as "no transcripts yet", so memory condensation had never once succeeded: a long run of condense-aborts and zero successes, each failed attempt still writing a full backup first. The offline usage reconciler and cross-cwd session-resume read and wrote the same wrong path. Found and diagnosed by @gts-47.
• Compaction ran on two schedules at once. The hourly ops standup carried an autoCompact flag that the context trigger was supposed to have replaced, so a default install requested compaction on both cadences — and turning the trigger off left the standup compacting anyway. There is now one control, and its off-switch is honest.
• Duplicate /compact messages piled up in the queue and fired together, each answering "nothing to compact" after the first had done the work. One pending compaction per agent is now enforced in the message store, where no caller can route around it.
• The commit history is legible. It rendered through a library that positions rows at a fixed 64px regardless of the spacing it draws the graph at (rows overlapped), reserves 500px for the graph regardless of available width (text was squeezed into the remainder and wrapped), and never displayed the commit subject at all. It is now drawn directly and fits any panel width.
• Disabled buttons were unreadable, and icons vanished on dark buttons. Disabled fills swapped to a mid surface while the label kept its inverted colour — roughly 1.4:1 in dark mode. Icons were painted in the same token as a primary button's fill, so the arrow on Send was invisible whenever it was enabled.
• Two scrollers that never scrolled — the IDE's changed-file list and the per-commit file list. overflow: auto inside a capped column sizes to its content without flex: 1, so it overflowed the cap rather than reaching its own scroll threshold.
• Codex hooks stopped timing out after 1s. Codex reads timeout as seconds and normalises it with .max(1), so the timeout: 0 sentinel copied from Claude's config meant one second, not "no timeout" — every Codex worker logged a failed SessionStart hook.
• Every download link on the release page 404'd. They carried a version-pinned filename that stopped resolving four releases ago; mac downloads fell from ~118 to single digits. The release gate now refuses to ship links that don't resolve.
• A tooltip clipped by the agent dock, and the missing-key notice that overflowed the agent card.

Changed

• Scheduled auto-compact is held while any provider is rate-limited: /compact is a model call, and sending it into a capped CLI spends a rejected attempt and parks a /compact ahead of your real backlog.

[0.3.7] — 2026-08-08

Auto-update, fixed. It never ran. Not once, in any packaged build, since it shipped in v0.3.4 — and the app had no way to tell you so.

Fixed

• The native updater actually runs. electron-updater is CommonJS and exposes autoUpdater through a lazy Object.defineProperty getter, which Node's cjs-module-lexer cannot see. So await import('electron-updater') produced a namespace with no autoUpdater export — only .default.autoUpdater — and destructuring it yielded undefined. The first line of setup threw TypeError: Cannot set properties of undefined (setting 'autoDownload') into a catch that silently latched notify-only mode for the whole session. Every packaged build from v0.3.4 to v0.3.6 could therefore only ever offer "open the releases page". Invisible in development, because the whole path sits behind app.isPackaged.
• Updater failures are no longer swallowed. Every error is emitted to the renderer and appended to updater.log in the app's data folder. The old catch discarded the message, which is precisely why the bug above survived three releases.
• A single blip no longer disables updates for the session. The notify-only downgrade is per-check now, not a permanent latch, and a re-check can never clobber an already-staged update.

Added

• The toolbar version is an update control. Next to the logo it shows checking…, vX.Y.Z ready to install (click to download), live download progress, and restart to update (click to apply). With nothing pending, a click runs a manual check — the app previously only checked 30 seconds after launch and then every six hours, with no way to ask.
• update-available and download-progress are surfaced, so a multi-minute download is visible instead of looking like nothing happened. New update:download and update:current IPC: one-click download, and a reloaded window pulls the current state instead of waiting for the next tick.
• 9 tests over the update state model (src/shared/updateState.ts), covering the rule that bit here — a re-check must never wipe a staged "restart to update" — and that the underlying error reaches the UI.

Changed

• Website, README and release notes trimmed. The landing page drops the demo-video section and dark mode; the README's 48-row feature table becomes a scannable grouped list; RELEASE.md leads with the current release instead of stacking every prior one in full.

Upgrading from v0.3.5 or v0.3.6? Those builds carry the broken updater and cannot fetch this fix themselves — install v0.3.7 manually once. From v0.3.7 onward, updates arrive on their own.

[0.3.6] — 2026-08-08

A machine with nothing installed on it can now run agents. Every fix in this release is about the same failure: the app assumed the user's machine already had things on it — a shell to expand ~, a node on PATH, an npm to install with — and when it didn't, agents died with a bare exit code and no explanation. Plus a long-standing office bug where the floor went blank and never came back.

Fixed

• ~/dev/foo no longer fails with "cwd does not exist". Only a shell expands ~; Node treats it as a literal directory, so a typed ~/… path failed every existence check and the agent never spawned. ~ is now expanded once at ingestion, so the registry only ever stores an absolute cwd — which also repairs existing ~ entries that had been reading as "not absolute" forever.
• Hooks stopped silently dying with exit 127. Agent CLIs run hooks through sh -c with a bare PATH=/usr/bin:/bin:/usr/sbin:/sbin — nvm's node isn't there, so every hook payload was lost: no live status, no Stop→inbox drain, no session ids. Every hook shim now runs under the Node the app already bundles.
• node is on every agent's PATH. An MCP server declared as node ./server.js, a provider CLI that shells out, a .cjs an agent wrote itself — all died with 127 on a machine with no system node. The bundled runtime is now appended to the agent's PATH, never prepended, so anyone with their own node keeps their own version.
• The office floor comes back after losing its GPU context. Chromium caps live WebGL contexts and silently evicts the oldest — always the office, since it starts first, once enough agent terminals are open. Pixi reported nothing, so the floor simply went blank until you restarted the app. It now detects the loss and rebuilds the scene.
• God no longer messages agents that no longer exist. A live roster of the floor is pushed into the orchestrator's context on session start and on every prompt, instead of relying on it to re-read fleet.json — which is why it went stale across restarts.

Added

• Node and npm install themselves when they're missing. Selecting an engine on a machine with no Node used to print npm install -g … and run it anyway, so the user watched npm: command not found scroll past. The app now fetches the latest Node LTS straight from nodejs.org, verifies it against the official SHASUMS256.txt before anything executes, installs it visibly in that agent's terminal, and then installs the CLI. If you already have Node 20 or newer, it is left completely alone.
• An honest dead end instead of a doomed command. When no installer can succeed, the banner names the missing piece and runs nothing at all.

[0.3.5] — 2026-08-06

The queue always has an escape hatch, and the app updates itself for the first time. A fast-follow to 0.3.4: one real workflow fix, a sidebar polish pass, and the first release your installed app can pick up on its own — 0.3.4 installs get the "v0.3.5 downloaded — Restart to update" toast instead of a trip to the website.

Fixed

• Queued messages are never stuck again. Pausing floor-wide auto-delivery (the Command Center switch) used to hold every queued message with no explanation and no manual override. Each queued row now grows a send now link while the floor is paused: it moves that message to the front and bypasses only the pause gate — idle/draft/picker safety and delivery acknowledgement all still apply, so it types in the moment the terminal is genuinely free ("sending when free…" until then). The composer also says why nothing is moving: "held — delivery paused floor-wide", with the full story (and where to resume) on hover.

Changed

• Compact Command Center header. At sidebar width the old header wrapped its display-font title onto three lines, stacked "runs the floor" word-per-line, and let the two wide toolbar buttons crush everything else. Now: single-line COMMAND CENTER title + "Michael runs the floor" subtitle (both ellipsize), and the floor delivery toggle compressed to ▶ auto / ⏸ paused with the full explanation in its tooltip. The queue header's "clear all" no longer wraps either.

Notes

• First auto-updated release. 0.3.4 introduced the updater; 0.3.5 is the first version it delivers. Running 0.3.4 apps download this in the background and prompt "Restart to update" (never restarting on their own). 0.3.3 and older have no updater — grab this one from munderdiffl.in and you're on the train.

[0.3.4] — 2026-08-06

The queue you can trust, a Michael who actually knows the floor, and an IDE that shows you everything. A community release: the headline terminal/queue/roster reliability wave is by @gts-47 (Vyapak Goyal), with major fixes by @qschmick (#110, #111, #112, #114). Plus four new first-party features — voice Michael with live floor context + full app control, markdown previews (IDE and ⌘-click in any terminal), git history / branch compare / safe checkout in the IDE, and a redesigned six-tab Settings — alongside xAI Grok and Kimi Code as first-class engines, auto-update from GitHub releases, a professional type + color recalibration with full-app dark mode, and a scheduled auto-compact switch (default off).

Added (v0.3.4 feature wave)

• Talk mode grows up: live context + full control. Michael's voice session now opens with a compact per-agent floor snapshot (status, engine, context fill, breaker, inbox, in-flight tasks) and receives silent "(Floor update: …)" notes as things change mid-call — most "what's happening" questions need zero look-ups. New read tools: get_floor_state (precise live-floor JSON) and get_app_info (app version + release notes — "what's new in this version?" finally has an answer). New voice verbs behind the same safety spine: resume (the missing undo for pause/halt), auto-delivery pause/resume, tool gating, delete task, archive/unarchive, clear an agent's context (queued through every delivery gate; allowed on god behind confirm), create schedules, and change settings from a strict main-side allowlist (secrets and dangerous keys refused outright; behavior-changing keys echo old→new and require the distinct confirm word). Model bumped to gpt-realtime-2.1.
• Markdown previews everywhere agents write them. Markdown files in the IDE get a code | split | preview switch with live re-render as you type; ⌘-click any .md path an agent prints in its terminal to open a rendered preview instantly (edit toggle
  • "open in IDE" escalation included). Rendering is safe by construction for untrusted agent output — no raw-HTML pipeline exists, links never navigate the app, and remote images stay blocked.
• Git time-machine in the IDE. The left rail becomes CHANGES · HISTORY · COMPARE: a clickable commit graph (topologically ordered, all worktree branches, paginated) where picking a commit lists its files and opens per-file Monaco diffs; branch compare with ahead/behind counts and PR-style or literal modes; and guarded checkout — jump to any commit or branch, refused automatically when the tree is dirty or an agent is actively working in it. The slim git status panel also returns as a per-agent sidebar tab.
• Settings, redesigned into six tabs — General · Agents & Models · Autonomy & Budgets · Connections · Voice · Memory & Knowledge. The default agent model, autonomy mode, keep-Mac-awake, explain-things-simply, and the full circuit breaker (hard-stop included) all get real controls for the first time; dead display-only rows are gone; Danger Zone became a red row in General. Plus config truth fixes: Knowledge Graph genuinely defaults off, and the Free Flow toggle no longer shows OFF while the feature is on.

Added

• Auto-update. Packaged builds check GitHub releases on boot and every ~6 hours, download the new version in the background, and show a small "restart to update" toast — installation is always your click; the app never restarts on its own. Installs that can't self-update (the Windows portable exe) get a notify-only toast linking the release page. Settings → General → Auto-update turns the whole thing off. (Auto-update starts working for users on 0.3.4+; 0.3.3 and earlier have no updater — grab this release from the site once more.)
• xAI Grok, a first-class agent engine (grok) — worker and orchestrator-eligible. Grok's lifecycle events are Claude-compatible but camelCase on the wire; an AGENT_ID-scoped hook adapter (installGrokHooks) normalizes them into the hive's contract, so live status, guarded inbox delivery, and operator gates all work. grok [PROMPT] takes the protocol positionally and --resume continues a session. (gts-47)
• Kimi Code engine (kimi) — spawn Kimi workers with --auto autonomy and the K3/K2.7 model aliases. No hook bridge yet, so routed mail bounces to the god rather than silently dropping. (gts-47)
• Scheduled auto-compact switch — default OFF. The dedicated compact-maintenance schedule (v0.3.2) now ships disabled: scheduled /compact is opt-in. Flip it in Settings → General → Scheduled auto-compact or the Schedules tab (which keeps its interval picker and its warning when off). Existing installs keep whatever state they already chose. Scheduled compaction is now also provider-aware — each engine gets its own compact command instead of Claude's only. (founder decision + gts-47)
• Fullscreen agent roster rail. The horizontal tab bar ran out of room past a handful of agents and hid the operator controls; it's replaced by a left rail — + agent pinned at the top, god agents ungrouped above everything, workers bucketed under repository headers, restore-team and its dismiss chips pinned at the bottom. An isolated agent's cwd is its own git worktree, so a new mainRepoRoot helper follows a linked worktree back to its main checkout (cached per cwd) and groups key on the absolute repo root, so two checkouts with the same name stay separate. Notes render on the row (one line per bullet) instead of behind a hover popover, the note editor becomes a textarea so Enter makes a new bullet instead of dropping every bullet but the first, pause/halt/steer come back in fullscreen, god agents render the full Command Center, and drag-to-reorder carries over (confined to an agent's own repository group). The destructive kill button is gone.
• typing badge — see why a queue is held. A message queue held by your own unsent text on an agent's prompt used to look identical to an idle agent with nothing to do. Agent cards and the fullscreen roster now show a "your draft" badge whenever you have unsubmitted text on that agent's prompt. It's derived at render from the same check the delivery gate uses, so the badge can never disagree with the reason nothing is being delivered.
• docs/message-queue.md — the delivery contract: who may type into an agent's terminal, when, and what automation is never allowed to do to your text.
• Remote Control sessions are named after the agent (#81). claude.ai / the mobile app now shows "Michael", "Jim", … instead of <hostname>-<random>, so a floor full of RC sessions is finally tellable-apart. (gts-47)
• Roster shared between dev and a packaged build. The roster (agents + notes + queues + selection) mirrors to a file beside the hive (src/main/roster.ts), so the dev build and the installed app see the same team; localStorage remains the per-origin fallback. Restore-team also runs in parallel, fires on open, and each restorable agent gets its own ✕ dismiss. (gts-47)
• Fable 5 + Sonnet 5 in the Claude model picker (Fable 5 is the new default model). The two "default" entries — the harness's configured default vs the CLI's own — are now labeled distinctly in every picker, and every Claude option names a real model. (gts-47)
• New test suites — queue delivery, terminal automation + recovery, roster persistence, provider config/automation, codex remote, agent env, and PID release (npm run test:focused + test/proc-kill.test.cjs). (gts-47, qschmick)

Changed

• One gate for every automatic writer. The queue kept breaking because two loops each decided for themselves when it was safe to type into a terminal. The inbox-wake nudge now enqueues like scheduled /compact already did, so a single drain loop owns every "is this terminal free?" decision — idle, off cooldown, past boot grace, delivery not paused, no user draft in the way.
• Automation never destroys or closes what you own. A draft block lasts half an hour rather than a minute, a picker block half an hour rather than three minutes, and when either expires delivery simply types after whatever is on the line (the two fuse into one prompt). Wiping the line first, and sending Escape at an open picker, are both gone — deleting text is worse than garbling one prompt, and we can't verify that Escape closed anything. Both keys remain on the composer's own button, where you asked for them.
• Terminal zoom scales the whole pane. Cmd +/- moved out of PtyTerminalView into a shared subscribable module, so it now scales the message composer and the roster along with the terminal instead of leaving them pinned at sizes tuned for a 14" display.
• Changing an agent's model within the same provider resumes its session instead of starting fresh (best-effort — an agent with no recorded session still gets its model changed). The Command Center picker is now cross-provider (switch an agent between engines from one dropdown).
• Codex resume actually resumes. Codex has no --resume flag — it resumes via the codex resume subcommand — and its rollout transcript + sqlite index live together in a per-agent CODEX_HOME. Respawns now use the subcommand, and resuming by a pasted session id points CODEX_HOME at the agent home whose state DB actually indexes that session (a stray rollout copy without an index can't be opened). Codex workers also get Codex Remote: the app-server daemon is started under the agent's isolated home so the thread shows up in ChatGPT mobile. (gts-47)
• Restore team runs in parallel and the login-shell capture is memoised, so rebuilding a roster no longer serialises one shell probe per agent. Agents are added in roster order rather than completion order.

Fixed

• Blank terminal pane where typing does nothing. Three causes, which is why every single-cause fix only half-worked. (a) WebGL is now a lease taken on attach and released on detach — a browser allows only ~16 live contexts and silently discards the oldest past that cap, so restoring a team blew the cap and Chromium killed a background terminal's renderer while its pty, buffer and subscription stayed healthy. (b) requestInitialPtyRedraw latches only once the redraw actually succeeds, instead of burning the terminal's one chance on a failed fire-and-forget IPC. (c) The needs-repaint marker is cleared only after the refresh returns, so a throw no longer discards the last record that the terminal needed repainting.
• Message queue silently stopped delivering to an agent. (a) Only a bare command opens a picker — /model prompts you to choose, /model sonnet applies the argument and returns to the prompt; first-token matching latched the block on both and the submitting Enter couldn't clear it. (b) The picker block now expires like the draft block does, so a picker closed in a way we can't observe no longer wedges the queue for the rest of the session. (c) lineBuf moved onto the pool entry and is reset everywhere inputDirty is — as a closure variable it survived a draft clear, so the next keystroke recomputed the block from text that had just been deleted.
• Phantom drafts blocked delivery against text that didn't exist. inputDirty is inferred by counting keystrokes, and a TUI that swallows keys for its own UI leaves it set while the prompt is visibly empty. The prompt is now read from xterm's rendered buffer (cursor row, prompt chrome stripped) — deliberately one-directional: the screen can only clear a phantom draft, never invent one, because being wrong the other way would type over something you really wrote.
• A queued message could be typed into an open picker and acked as delivered. Clearing the input line no longer clears the picker latch — Ctrl-U kills the input line, it does not close a menu — so automation is no longer told the prompt is free while a picker still owns it.
• A model or command change died on reload. updateAgent now persists when a durable field changes (volatile run-state fields — status, action, progress, context counters — still skip the write, so a burst of terminal output doesn't rewrite storage).
• The roster stopped spawning git rev-parse on every chunk of terminal output — failed and in-flight repo lookups are cached, so an agent outside a repo is looked up once.
• Killed processes actually die (#110). Every explicit kill was a bare node-pty proc.kill() — one SIGHUP to the direct child only, so a TUI that traps SIGHUP lived on and its children (MCP servers, helper daemons) reparented to PID 1 and kept running for the machine's uptime. Every kill path now routes through ensureKilled: a grace signal, then SIGKILL of the whole process group (POSIX) / taskkill /T /F (Windows). (qschmick)
• Circuit-breaker false-positive storm on idle/compacting agents (#109 → #112). Compaction and inbox-ack token bursts no longer read as looping. (qschmick)
• Shell-capture fencing. The interactive login shell used for PATH/which capture runs the user's rc files, which are free to print (zsh's "Restored session: …" plugin chatter was being prepended to every agent's PATH). Capture output is now fenced between markers, with a multi-line sanity check before a PATH is trusted. (gts-47)
• Fullscreen surfaces notes and modals (they stacked under the overlay); an un-echoed keystroke no longer reads as an empty prompt; expired automation blocks are acted on instead of typed through. (gts-47)

Performance

• ~350× faster warm usage reads — an incremental per-file transcript cache replaces re-reading every transcript on each poll (#111). (qschmick)
• Spawns stopped freezing the app. Command resolution is memoized (with on-disk revalidation) and the login-shell PATH is captured once per session — each interactive-shell launch cost ~1s of blocked main thread, paid twice per spawn, ×N on a team restore (#114 + gts-47's equivalent, merged). (qschmick + gts-47)

[0.3.3] — 2026-07-03

An IDE on the floor, and a seventh engine. Two headliners: a built-in Monaco IDE — the VS Code editor engine in a full-window overlay, with a git CHANGES rail, side-by-side diffs vs HEAD, a file tree, editor tabs, and Cmd/Ctrl+S save — and GitHub Copilot CLI as a first-class agent engine, the project's first community-contributed provider (PR #101 by @anxkhn).

Added

• Built-in Monaco IDE panel. A title-bar IDE button toggles a full-window IDE overlay (matching the existing fullscreen-overlay pattern — the office floor, terminals, and voice UX are untouched). Left rail: a git CHANGES list (click a file → read-only side-by-side diff vs HEAD) plus the reused workspace file tree (click → edit). Right: editor tabs with dirty-state dots, save, and close; Cmd/Ctrl+S saves the active tab. The workspace root snapshots from the selected/god/first agent cwd. Monaco is fully self-hosted — electron-vite-safe bootstrap with bundled ?worker imports and loader.config({ monaco }), no CDN — themed to the harness's light palette, and all fs/git access goes through main-process IPC (git:diff + preload bridge; the renderer holds no fs/git access) (src/renderer/src/ide/*, src/main/git.ts).
• GitHub Copilot CLI agent engine (copilot, npm @github/copilot) — community-contributed. Registered as a provider preset driven in Copilot's documented non-interactive print mode: copilot -p "<prompt>" -s --allow-all-tools --no-ask-user [--model <id>], with the auto-approval flags gated by the floor auto-mode toggle like every other engine. Includes a COPILOT_MODELS picker (Claude Sonnet 4.5 default · GPT-5.4 · auto), --resume session continuity (best-effort), voice-hire (spawn) support, binary inference for pasted commands, and the official npm install -g @github/copilot offered by the missing-CLI installer. Non-hiveAware by design: print mode exits per turn and exposes no hook bridge, so canReceiveInbox is false and routed mail bounces to the GOD orchestrator instead of silently dropping (src/shared/agentProvider.ts, PR #101 — thanks @anxkhn).
• Agent-provider registry test. A self-contained, framework-free test (node test/agent-provider.test.cjs) transpiles the shared registry and asserts provider selection, the copilot command shape, model/resume passthrough, and codex non-regression.

Fixed

• IDE: no silent loss of edits typed during an in-flight save. Keystrokes entered while a save was still writing can no longer be dropped when the save completes.

[0.3.2] — 2026-06-27

Talk to Michael. The headline is Realtime Michael — a low-latency voice channel to the GOD orchestrator, running alongside the async terminal floor. Press Talk, and Michael listens, answers, and acts in real time: he reads the hive (tasks, board, memory, agents, activity) and — behind spoken echo-back confirmation for anything destructive — creates and assigns work, dispatches agents, spawns and kills workers, and steers the floor, all attributed to a distinct michael-voice actor that pings the GOD terminal. He greets you on connect, speaks task completions the moment they land ("respond when done"), and runs under a live cost meter with a hard spend cap and an idle auto-disconnect. It's bring-your-own OpenAI key: the key is decrypted main-only, minted into short-lived ephemeral session tokens, and never reaches the renderer. Plus Slack hardening (proactive posting off by default; no sends without an explicit thread), a dedicated auto-compact maintenance schedule decoupled from missions, and per-agent environment metadata.

Live verification note. The realtime voice loop is human-verified end-to-end on a real OpenAI key — connect → mic → Michael answers via the read tools, and the full destructive action path (spoken echo-back confirm → spawn / kill / dispatch → the worker appears on the floor → completion spoken back) was exercised live. It requires your own OpenAI key with Realtime API access; without one the Talk button stays visibly disabled with a "needs OpenAI key" cue.

Added

• Realtime Michael — talk to the GOD orchestrator by voice. A new low-latency realtime channel (OpenAI Realtime API over WebRTC) sits next to the async terminal. A Talk toggle (on Michael's card and in any fullscreen terminal) opens a mic session with EC/NS/AGC, semantic-VAD turn-taking + barge-in, and a device picker for both microphone and speaker. Michael runs his own persona and answers in a natural voice, with an Off → Connecting → Listening → Responding → Working state machine surfaced live on his card (src/renderer/src/realtime/*, src/main/realtime.ts, RealtimeMichaelToggle.tsx).
• BYOK ephemeral-token mint. The voice session authenticates with a short-lived ephemeral client secret minted main-side from your stored OpenAI key — the real key is decrypted main-only and never crosses IPC. The renderer's CSP allows the WebRTC SDP exchange to reach api.openai.com while keeping everything else locked down (src/main/realtime.ts, renderer index.html CSP).
• Voice action set with tiered echo-back confirmation. Michael can do things by voice — read tools (tasks / board / memory / agents / activity / cost) plus the full action set: create and assign tasks, dispatch agents, pause / steer / halt, spawn / hire, kill, and edit schedules. Every destructive verb is gated behind a spoken echo-back confirmation (a distinct confirm token, never a bare "yes"), with hard refusals for killing the GOD agent or targeting all agents at once (src/renderer/src/realtime/actions.ts, src/main/realtimeActions.ts).
• "Respond when done" completion loop. Voice-dispatched work reports back on its own: a main-process completion watcher detects when a dispatched task finishes (card → done or a done reply in the inbox) and proactively pushes the event into the live session so Michael speaks it unprompted, while a CompletionToast shows it on screen. If the session is closed, completions queue to a desktop notification and a "completions since last session" warm-start; a wait_for tool covers the block-until-done case (src/main/realtimeCompletionWatcher.ts, CompletionToast.tsx).
• michael-voice as a distinct actor. Actions taken by voice are attributed to a separate michael-voice identity in messages, the board, and the activity log, and notify the GOD PTY — so a voice-driven dispatch is auditable and never silently impersonates a worker.
• Cost guard + idle auto-disconnect for voice sessions. A live session cost HUD by the Talk toggle, a configurable spend cap that auto-disconnects when hit, and a configurable idle auto-disconnect (default 3 min, 30 s–10 min or Off) so a forgotten-open mic can't run up a bill (src/renderer/src/realtime/cost*, Settings → AI Engines).
• Greeting on connect. When a session goes live, Michael opens with a warm, rotating greeting ("Hi, what's up?", "Hey, how's it going?", …) instead of waiting in silence — best-effort and guarded so a not-yet-ready data channel never blocks a successful connect.
• Conversational read-layer. The voice read tools were reworked to actually answer hive questions: get_memory no longer dead-ends, and new agent/board tools plus an expanded persona let Michael talk through roster, tasks, and floor state naturally.
• Voice read-layer over hive messages (read/brief-only). Realtime Michael can now read message content, not just metadata: a get_messages tool returns a full message by id, one mailbox, or the latest across the floor to brief the operator. All redaction is main-side — voiceMessages() runs every subject/body through redactSecrets() before the result leaves the main process, so the renderer/voice layer only ever receives already-redacted bodies (no provider / Slack / GitHub / AWS / Google key, JWT, PEM private-key block, or Bearer token can leak) and holds zero redaction policy. Read-only: it adds no write/mutate path — voice writes still go through the separate confirm-gated action spine (src/main/hive.ts voiceMessages + redactSecrets, src/renderer/src/realtime/{tools.ts,VOICE-MESSAGE-ACCESS.md}, test/voice-messages.test.cjs).
• Talk reachable from any fullscreen terminal. The Talk toggle is no longer Michael-only chrome — it's reachable in any fullscreen terminal view (the toggle is global/session state, so it's correct everywhere), while the per-session cost HUD stays Michael-only (FullscreenTerminal.tsx).
• OpenAI Realtime key — documented and gated. Settings → AI Engines now documents the OpenAI Realtime key as its own requirement (the same OpenAI provider key, distinct from your Anthropic key; main mints a short-lived token from it per session). The Talk button shows a live enabled/disabled status and an inline "needs OpenAI key" cue when none is set, so connecting never lands on a silently-dead button (RealtimeMichaelToggle.tsx, SettingsModal.tsx, AiEnginesSettings.tsx).
• Dedicated auto-compact maintenance schedule. Auto-compaction is now a persistent, configurable maintenance mission of its own, decoupled from the standup mission it used to ride on (so editing standups can't silently drop it). It reappears disabled rather than vanishing, with a mandatory warning and a configurable interval in the Schedules tab, plus a migration (src/main/schedules*, SchedulesTab).
• Per-agent environment metadata + cwd guard. Each agent now carries queryable environment metadata with a working-directory validity guard, and a new agent-env query tool (src/main/hive.ts).

Changed

• "Voice" is now "Talk". The voice feature is renamed Talk throughout, with a redesigned navigation: the GOD card pops with a dedicated Talk line, and the worker nav cards are compacted to make room (src/renderer/src/components/*).
• Robust voice task-matching (findCard). Resolving a task by voice is now tolerant of hyphens/punctuation, phrasing, and truncation: both the spoken phrase and the stored title are normalized, candidates are scored (exact / prefix / token-coverage / substring), and close matches trigger a spoken "which one?" disambiguation instead of silently mutating the wrong card (src/main/realtimeActions.ts, test/realtime-findcard.test.cjs).

Fixed

• Fullscreen agent modal now opens above the fullscreen view. The Add-Agent modal launched from the in-fullscreen "+ agent" button rendered behind the fullscreen terminal (z-index 100 vs 250) and was non-interactive. It's lifted to the dialog tier (300) so it's on top and clickable, and the fullscreen Esc handler now closes the modal first instead of exiting fullscreen underneath it (AddAgentModal.tsx, FullscreenTerminal.tsx).
• Voice get_memory no longer dead-ends. The conversational read tools were fixed so memory lookups return usable answers instead of stalling the turn.

Security

• BYOK voice secret invariant. The real OpenAI key is encrypted at rest and decrypted main-only; the renderer only ever sees a short-lived ephemeral client secret minted per session. The key never crosses IPC and is never logged. The voice read-layer reports tokens, not dollars (de-monetized chrome), and every destructive voice action is held behind spoken echo-back confirmation with hard refusals for killing the GOD agent or targeting all agents at once.
• Slack: proactive posting off by default + explicit-thread guard. App/voice Slack sends are now off by default behind a config flag + Settings toggle, and a request with no explicit channel+thread is refused rather than guessed — closing an unattended-broadcast path (src/main/slack.ts, Settings).

[0.3.1] — 2026-06-22

Three more coding CLIs join the floor — OpenCode, Crush, and pi.dev — each usable as a worker and as Michael, with bring-your-own keys + local LLMs. Plus two reliability fixes: the sleep-frozen message router and Codex workers' filesystem permissions.

Live verification note. The three engines are wired end-to-end and selectable as god, and their architecture (preset + bridge + payload contract) was reviewed line-by-line. Their bridges' runtime behavior needs real model calls, so the following are on-device checks pending BYOK keys / a local LLM (not runtime-proven here):

1. each bridge's turn-end signal actually fires — OpenCode session.idle, pi agent_end, Crush's proxy-synthesized Stop — flipping the agent to idle;
2. OpenCode local-LLM happy path: pick local/<id> with a base-URL set and confirm a turn completes (the injected config now registers the selected model id);
3. Crush routes through the proxy on an OpenAI-wire model (the default god is now openai/gpt-4o) and Crush honors the partial base_url override;
4. the auto-mode gate holds (no permission:allow / --yolo when the floor toggle is off).

Crucially, mail delivery does not depend on those signals: a new provider-agnostic PTY-quiescence idle fallback flips any silent-but-pinned-working agent to idle, so the provider-agnostic idle inbox-wake nudge drains a god even if a bridge's turn-end signal never fires. That backstop is the safety net under shipping all three as canReceiveInbox:true.

Added

• Three new selectable engines: OpenCode · Crush · pi.dev. Each lands as a declarative AgentProviderPreset and appears automatically in the Add-Agent picker (worker) and the god engine picker (orchestrator). Each gets a bridge for live status + turn-end inbox-drain:

  Engine	Identity	Bridge	Notes
  OpenCode	opencode (anomalyco/opencode, TS)	native plugin (session.idle)	bundled per-agent plugin; no traffic interception; auto-approve via gated OPENCODE_CONFIG_CONTENT
  Crush	crush (charmbracelet, Go)	proxy (qwen-tier)	per-agent CRUSH_GLOBAL_CONFIG routes traffic through the loopback sidecar (Crush has no base-URL env)
  pi.dev	pi (earendil-works)	hooks (bundled extension)	pi.on(event) → HIVE_SOCK; extension auto-approves tools only when the floor is in auto mode

• BYOK + local-LLM config UI (Settings → AI Engines). A new per-provider config surface: API keys for the backend model-providers (Anthropic / OpenAI / Google / OpenRouter / Groq) stored write-only in the encrypted secret broker (never read back to the renderer; materialized main-only at spawn), plus per-engine local base-URL + default-model fields (HarnessConfig.providerBaseUrls / providerDefaultModels). Pi/OpenCode/Crush/Qwen pick up the keys + endpoints at spawn; auto-mode stays gated behind the floor toggle, and each engine runs unsandboxed in auto mode (surfaced as a caveat).

• Provider-agnostic idle backstop (PTY-quiescence fallback). A floor-wide check flips any agent pinned working with no terminal output for a short window back to idle — so the idle inbox-wake nudge can always drain a non-Claude god even if its bridge's turn-end signal (Stop / session.idle / agent_end) never fires. This is the safety net under shipping all three engines as god-eligible (canReceiveInbox:true) while their bridges await on-device verification (src/renderer/src/hooks/useHive.ts).

Fixed

• Codex hive workers get full filesystem + auto-approval from spawn (parity with Claude). A Codex-engine agent in auto mode launched with -a never -s workspace-write, whose sandbox scopes writes to the PTY cwd (the user's project). But a hive worker must also write to its agent folder at <harnessHome>/hive/agents/<id>/ (move inbox/ → .done/, append memory.md, drop outbox JSON, write deliverables) — a different path tree from cwd, which workspace-write blocked. So a freshly spawned Codex worker couldn't complete HIVE PROTOCOL housekeeping and reported "it does not have permissions … grant write permission to the agent folder." Codex's auto-mode flag is now --dangerously-bypass-approvals-and-sandbox — the documented equivalent of Claude's bypassPermissions / Antigravity's --dangerously-skip-permissions (skip all approval prompts and drop the OS sandbox), so a Codex worker has the same filesystem access and auto-approval as a Claude worker from the get-go (src/shared/agentProvider.ts; reference/copy updated in src/shared/codexCommands.ts, OnboardingWizard.tsx, renderer/store/config.ts). Claude/agy/antigravity behavior is unchanged.
• Re-arm the hive message router on wake (god→worker delivery survives sleep). The outbox→inbox router is a setInterval (hive.routeOnce every ~1.5s) that, like the always-on beats, freezes during true macOS system sleep. onSystemResume() already re-armed the mission scheduler, the fleet/breaker beats, and keep-awake on powerMonitor resume/unlock-screen — but it never re-armed the router. So after a long sleep (e.g. laptop closed overnight) the scheduler→god path recovered while every agent's outbox silently stopped draining: god→worker, worker↔worker, and broadcast mail piled up undelivered, and no message event was logged. The resume handler now re-arms the router (clear-then-set, idempotent) and immediately drains the accumulated backlog instead of waiting for the first post-wake tick; the renderer's idle inbox-wake nudge then wakes each parked recipient once its mail lands (src/main/index.ts). Verified by scripts/verify-keepalive-catchup.mjs (now also reproduces the pre-fix backlog stall and proves the re-arm + flush).
• Open-source model quick-picks + local-setup guides in Add-Agent. Hiring a worker on a local-capable CLI engine (OpenCode/Crush/pi.dev) now shows curated OSS-model quick-picks — a Local bucket (Mac-runnable Ollama tags: gpt-oss 20B/120B, Qwen3 30B-A3B/Coder, DeepSeek-R1 32B, Mistral Small, GLM-4.7-Flash, Llama 3.3 70B) and a third-party OSS provider bucket (BYOK: gpt-oss/Llama via Groq, DeepSeek-V4-Flash, GLM-4.6, Kimi K2.6, Qwen3-Coder via OpenRouter). Picking one fills the engine-correct slug (OpenCode local/<tag>, Crush/pi ollama/<tag>; provider slugs identical across engines) and rebuilds the command. Slugs are transcribed from a verified catalog — bleeding-edge frontier models are intentionally left out of code defaults. The Add-Agent help line and Settings → AI Engines local-setup area now hyperlink two how-to guides (run on open models · set up on a Mac Mini) (src/shared/ossModels.ts, AddAgentModal.tsx, AiEnginesSettings.tsx).
• Crush no longer dies with Unknown command on spawn (the hive protocol now reaches it). A Crush worker was launched as crush --model <m> --yolo "You are …(the whole hive protocol)" — the protocol passed as a positional arg. But bare crush is an interactive Bubble Tea TUI on a Cobra root command, which reads the first positional as a subcommand, so it aborted with unknown command "You are…"; the protocol never reached the model, the worker never learned it was a hive agent, and the PTY died. Crush has no --prompt flag and crush run is one-shot, so the protocol is now typed into the TUI instead: a new preset capability seedDelivery:'type-into-tui' makes the spawn drop the positional (crush [--model m] [--yolo]) and hand the protocol back as a seedPrompt, which the renderer types in as the worker's first turn after a boot-grace — through the same per-pty write-chain as the inbox-wake nudge, so the seed and a nudge can never jam onto one line. Covers fresh Crush spawns, restores, and Crush-as-Michael (src/shared/agentProvider.ts, src/main/hive.ts, src/main/index.ts, src/preload/index.ts, src/renderer/src/hooks/useHive.ts, AddAgentModal.tsx, AgentStrip.tsx, store.ts).
• Auto restart-and-continue after a first-time engine-CLI install (no dead-end). When an agent's engine binary (OpenCode/Crush/pi.dev/Codex/…) wasn't installed, the missing-CLI short-circuit ran the provider's installer in the PTY, then printed "click restart & continue to launch the agent" — but no such button exists for a not-yet-started agent, so the PTY just sat at process exited (code 0) and the agent dead-ended. Now, on a clean install exit, the PTY-exit handler auto restart-and-continues: it re-runs the same spawn into the same pty/window (carrying a noAutoInstall flag) so the freshly-installed CLI launches with no user click, and the renderer re-arms that terminal in place (clears the "process exited" line, re-enables input) via a new pty:relaunch signal. Provider-agnostic (every engine's installer path) and idempotent by construction — noAutoInstall guarantees the installer can never fire twice, and providers with no bundled installer (manual-hint-only) are never armed for relaunch. The install banner copy is now honest ("Installed — launching the agent…") (src/main/index.ts, src/main/pty.ts, src/preload/index.ts, src/renderer/src/components/terminalPool.ts).

[0.3.0] — 2026-06-21

A platform release: the floor stops being Claude-shaped. Selectable agent engines make every hire — and Michael himself — a pluggable engine (Claude Code / Antigravity / Codex / local providers), each with its own per-hire skills + MCP catalog behind a consent UI. A new integrations registry + loopback secret broker turns "connect a service" into a write-only, registry-driven Settings flow. Michael can now spawn an ephemeral worker straight from Slack — reply, then tear it down safely with worktree GC and token caps — surfaced in a new Workers tab. Plus temporal date-range skills and a worker capability catalog, a Provider / Hive picker in onboarding and add-agent, the Agent Gallery (the rebranded Hiring Fair) with six off-the-shelf hires, feature-aware onboarding, and wake-reliability hardening. Everything from v0.2.8 and earlier is included.

Added

• Selectable agent engines + per-hire capabilities. A new engine abstraction (agentProvider + an mcpCatalog, mirrored across a 3-file config) makes the runtime behind each agent pluggable — Claude Code, Antigravity, Codex, or a local provider (a claw/qwen backend proxy bridge with default-MCP merge). Each hire carries its own manifest of allowed skills + MCP servers (a default-deny allowlist over the catalog), with bundled skills shipped via Electron extraResources (resources/skills → <resources>/skills) and a consent UI that surfaces every skill/MCP a hire wants before it can use it — untrusted hire input is reviewed, never auto-granted.
• Swappable Michael (god) engine. The orchestrator is no longer hard-wired to one CLI: useHive gains an engine-spawn path, Onboarding gains an engine picker for Michael, and a change-engine flow lets you re-home the god orchestrator onto a different engine without rebuilding the floor.
• Integrations registry + loopback secret broker. A declarative integrations registry (src/shared/integrations.ts) plus a loopback secret broker (src/main/integrationBroker.ts): secrets are write-only (set once, never read back into the renderer) and reached only through the broker over loopback. A registry-driven Settings UI (IntegrationsRegistry) renders each integration's config form from the spec — conformed to registry spec v1 — and a first wave of declarative templates (the canonical schema + initial YC-style templates) ships in the registry. The integrations:* surface is exposed to the renderer through a dedicated preload bridge.
• God-triggered ephemeral Slack worker loop. Michael can now spawn an isolated worker directly in response to a Slack request — the worker does the work, posts its reply back into the thread, and is then torn down safely. Lifecycle hardening adds worktree garbage collection, token-cap wiring per spawned worker, and a teardown-safety gate that refuses to auto-discard a worker's unintegrated work. The pty:spawn IPC handler was refactored into a reusable spawnAgentCore that underpins worker spawning, and a new Workers tab surfaces live ephemeral workers in the UI.
• Temporal date-range skills + worker capability catalog. A family of date-range skills (today / yesterday / thisWeek / lastWeek / thisMonth / thisQuarter / thisYear / lastMonth / lastQuarter / lastYear / last7Days / last30Days … plus an arbitrary-range temporal resolver backed by temporal/when.mjs) resolve a named window to concrete ISO dates without hand-math. A worker capability catalog lets each spawned worker read exactly which skills and brokered integrations it has and how to call them.
• Provider / Hive picker UI. A new HivePicker component plus a ProviderLogo set (real provider logos) appear in onboarding and the add-agent flow, so choosing the engine/provider for a hire is a first-class, visual step instead of a free-text command.
• Agent Gallery + six off-the-shelf hires. The community gallery is rebranded from The Hiring Fair to the Agent Gallery, and ships six ready-made, off-the-shelf hires you can browse, review, and spawn.
• Feature-aware onboarding + a permissions & reliability step. First-run onboarding now adapts to the features you have available and adds an explicit permissions & reliability step.
• Visible engine-CLI installer. When the engine binary for a chosen provider is missing, the installer now runs visibly instead of failing silently, so a first-time setup self-heals.

Changed

• The Hiring Fair → Agent Gallery. The gallery is renamed throughout (landing page, in-app links, copy) to Agent Gallery; existing hire links and the /hires/ path continue to work.
• Add-Agent config IA rework + Command Center UX fixes. The Add-Agent modal's configuration is reorganized around the new engine/capability model, with assorted Command Center UX cleanups.
• VDE prototype (experimental). An experimental Virtual Desktop Environment prototype lands behind the scenes, with a Groq chat-completion module (src/main/groq.ts) powering its AI assist.

Fixed

• Orchestrator delegates opportunistically to existing agents. Michael now checks the live roster (active agents in registry.json + their state in fleet.json) before spawning, and prefers routing a task to an existing agent that already fits — above all when the request names one ("ask Pam…", "have Jim…") — instead of reflexively creating a new agent; he only spawns a fresh one when no existing agent is a sensible fit, and says that he checked. Encoded in both the floor orchestrator prompt and the Slack autonomous-request protocol (src/main/hive.ts, src/main/index.ts).
• Auto-revive wedged terminals on wake. A terminal that wedged while the machine slept is now detected and auto-revived when the machine wakes, instead of sitting dead until manually restarted.
• Catch up missed schedules on wake (power keep-alive hardening). Scheduled missions whose fire time elapsed while the machine was asleep are now caught up on wake rather than silently skipped (verified by scripts/verify-keepalive-catchup.mjs).
• Worker stale-done guard. A worker is now released only on a done authored after it was spawned, so a stale done from a prior life can no longer prematurely release or tear down a live worker.
• Floor: "add agent" button stays on one line. The add-agent button no longer wraps as the roster fills.
• Voice button disabled with a tooltip when the Groq key is missing, instead of failing on click.

Security

• Confined the integrations:test probe path. The connectivity-test path for integrations is now constrained so it can't be turned into a secret-exfiltration or SSRF primitive — a brokered, bounded probe rather than an arbitrary outbound request driven by registry/secret input.

[0.2.8] — 2026-06-15

A feature release: shareable hires — package a role-configured agent as a portable manifest, share it as a file or host it in a gallery, and import it into any office with one click. Plus The Hiring Fair, a community gallery of ready-made roles, and a hardened, untrusted-input import pipeline.

Added

• Shareable hires (#70, #71). A portable munder-difflin/hire@1 JSON manifest describing a role-configured agent — name, sprite, provider, model, command flags, goal, capability tags, token budget. Two import paths, one pipeline: a munderdifflin://hire?src=<https-manifest-url> deep link (fetched and validated in the main process, queued, then pulled by the renderer on mount) and an import hire… button in the Add-Agent modal that reads a local manifest file. Either way the manifest only pre-fills the Add-Agent modal behind an "imported" banner; spawning stays an explicit human click — import never auto-spawns. Protocol registration ships for all three platforms (macOS open-url, Windows/Linux single-instance lock + cold-start argv forwarding), and packaged builds register the scheme via electron-builder.yml.
• The Hiring Fair — community gallery at munderdiffl.in/hires (docs/hires/, static, no build step, served by the existing GitHub Pages setup). Seed roles drawn from the cast (Pam writes docs, Dwight enforces QA, Jim reviews PRs, Creed audits security, Angela audits the office's own token spend, Stanley does the migrations nobody wants), each with a Claude Code / Antigravity / Codex provider toggle (per-provider variants generated from one base manifest), function filters matching the landing page, and a client-side validator identical to the app's alongside a JSON schema (docs/hires/spec/). Model suggestions are data-driven (docs/hires/models.json), so new models are a one-line update.

Security

• A hire manifest is untrusted input — defense in depth. No auto-spawn and no executable field: provider: "custom" is rejected and the binary always comes from the user's local provider preset. Embedded CLI flags are gated by a default-deny allowlist (SAFE_FLAG_NAMES) — only known-harmless flags pass, nothing system-prompt/settings-related — replacing an earlier denylist that drifted as each CLI added flags. model is constrained to a safe charset (MODEL_RE), and a command-line quoter neutralizes cmd.exe metacharacters (& | ^ < > ( ) % !) on every spawn path — closing a Windows command-injection class (PoC "model":"x&calc"). The manifest fetch is https-only, manual-redirect with per-hop re-validation (kills redirect SSRF into 127.0.0.1 / 169.254.169.254, including an IPv6-bracket bypass), streamed with a 64 KB byte cap (no trusting content-length), a 10s timeout, and ≤5 hops. The dependency-free validator (src/shared/hire.ts) is shared by the main process, the renderer, the gallery (docs/hires/validator.js), and the JSON schema, so all four stay in sync.

[0.2.7] — 2026-06-13

A feature release: talk to your agents with your voice, an opt-in enterprise Knowledge Graph, multi-window "floors", a richer message composer with file/image attachments, the groundwork for TV-show office themes, and a redesigned landing page — plus composer and fullscreen polish.

Added

• Free Flow voice dictation → message queue (now on by default). Hold Option to talk; your speech is transcribed by Groq Whisper (whisper-large-v3-turbo) straight into the message composer. Gated on a Groq API key, which is encrypted at rest.
• Enterprise Knowledge Graph v1 (now on by default). A multimodal store of your own documents / policies / business context, with a CLI agents can query for ranked passages and full documents — so company-specific facts come from your data instead of guesses.
• Multi-window "floors" (now on by default). Open isolated office windows, each with its own set of agents and per-PTY routing.
• Rich message composer — file & image attachments. Attach files/images (via a "files" button or paste-to-attach), shown as removable chips above a taller, resizable input; you can send with attachments alone.
• Restore agent sessions across restart, with Restart & Continue (#78). Agents reattach their prior Claude conversation after an app restart: Michael resumes his session (the orientation prompt is skipped on a genuine resume), and a restored worker re-enters its existing worktree instead of re-isolating, so uncommitted work isn't lost. The recorded session transcript is seeded into the target cwd before --resume attaches (and --resume is only used when the transcript is actually present, so there are no broken resumes against a missing id), and the pooled terminal soft-resets in place — staying live and typeable across a model change or respawn, redrawn at its real fit-derived grid. A per-agent Restart & Continue button respawns the session on the same model with resume to redraw a garbled terminal, and Add Agent gains a "resume session" field that reattaches by session id (auto-filling the folder, falling back to a fresh session if the id isn't found).
• Drag a file onto a terminal to inject its path (#79). Dropping a file (an image, etc.) onto an agent's terminal now writes its absolute, shell-escaped path into the session — so Claude Code detects the image path in the prompt and attaches it — instead of Electron navigating to the dropped file:// URL. Backed by webUtils.getPathForFile exposed from preload (Electron 32 removed File.path); only file drags are intercepted, so text/selection drags still fall through to xterm.
• TV-show office themes — infrastructure (behind a flag, off by default). A theme abstraction (ThemeConfig + registry/loader), a Settings theme picker with a destructive switch-flow, and the first themed map (Brooklyn-99 precinct). Ships dark via the tvShowOffices flag while the remaining maps land.
• Live GitHub star count next to the Star buttons on the landing page.

Changed

• Composer redesign. A full-width input above a single tidy control bar (Delegate · Attach · voice · Send) — no dead space from a stacked side column.
• Landing page redesign. Bento layout for the #features and #why sections with new SVG illustrations; the #claude section refreshed for v0.2.7.

Fixed

• Fullscreen tab bar no longer clipped. The fullscreen terminal's tab bar is un-clipped.
• Slack double-ack (#). A single Slack message delivered as both app_mention and message.* is now de-duplicated by channel:ts, so it's handled exactly once.

[0.2.6] — 2026-06-10

A polish + reliability patch: the agent terminal renders correctly the moment it opens, npm run dev no longer crashes on a missing sidecar, a Windows ConPTY crash is guarded, the wall clock becomes a clickable closing-time control, the ASK ME board reads in the memory font, and the Slack file download is host-pinned.

Added

• The office clock is interactive (#64). The clock on the wall reads the real time, and clicking it opens the closing-time (graceful shutdown) flow.

Fixed

• Terminal no longer renders oversized/clipped when an agent boots. xterm used to fit before its container had a real size and cached the character-cell metrics from before the web font (VT323) loaded, so the welcome banner overflowed and was clipped until you manually resized. The view now waits for a real size, re-measures and re-rasters the WebGL glyph atlas after the font loads, and lets the ResizeObserver drive the first fit — so it fits immediately on boot.
• npm run dev no longer crashes on the missing Slack-trigger sidecar (#67). The .cjs sidecar copy now runs as a vite writeBundle plugin, so both dev and build emit out/main/slack-trigger.cjs from one place. (v0.2.5 fixed only the packaged-build path, so a fresh clone's npm run dev still died at boot.)
• Windows: node-pty ConPTY AttachConsole crash guarded (#65). Companion to the Antigravity provider work — the main process no longer crashes when ConPTY fails to attach a console.
• ASK ME reads in the memory font (#63). The ASK ME board now uses VT323 instead of the chunky Pixelify Sans, matching the rest of the memory surfaces.

Security

• downloadSlackFile() host-pinned to Slack. The Slack bot token is now only ever sent to slack.com / *.slack.com hosts — a defense-in-depth guard before the Authorization: Bearer header is attached (the URL is already Slack-issued + HMAC-verified, so this hardens against a future redirect/parsing change).

[0.2.5] — 2026-06-10

A reliability + reach patch: a Windows-terminal regression fix, an agent-lifecycle cleanup that ends the breaker inbox-flood, Slack requests that actually reply with substance, a delegate toggle, six new tutorials/blogs, and an enriched landing diagram.

Added

• Delegate-to-agents toggle. A toggle switch above the Send button in the god orchestrator's composer prepends a delegation instruction so a request fans out to available agents (and is handled one-by-one if none are free). God-only, default off.
• AUTONOMOUS REQUEST PROTOCOL for Slack-origin requests. Inbound Slack requests now run fully autonomously: god routes the request to the most-relevant agent, that agent does the work and posts its substantive result back into the Slack thread itself, then reports to god. It pauses only for high-severity actions (pushing to main, spawning infrastructure/paid services, deleting files it didn't create), and any decision it needs is asked as a numbered-options reply in the thread and correlated back by thread_ts.
• Six new tutorials & blog posts — webhook setup, the full Slack setup, deploying an automated PR-reviewer agent, deploying a blog-writer agent, why CLI agents are so powerful (and how the hive cuts token use), and why a mixed-capability swarm beats a clone army.
• Enriched "how it works" landing diagram — Slack / Webhook / Schedule triggers feeding the orchestrator, plus a band showing each agent in its own isolated local git worktree.

Fixed

• Windows agent terminals no longer die on Program-Files installs (#55). cmd.exe is now invoked with a properly double-quoted command line (/d /s /c), so a Claude/node path containing a space (e.g. C:\Program Files\…) launches instead of splitting on the space.
• Orphaned-agent lifecycle cluster (#56/#57/#58). A stale agent entry with no live terminal no longer (a) re-writes a frozen cost-ledger row every ~30s, (b) trips the circuit breaker into an unclearable inbox-flood to the orchestrator, or (c) lingers un-archived — a startup migration archives no-PTY entries and the breaker now skips assistant/orphaned shells.
• Slack replies are real answers, not empty confirmations. Worker agents post the actual outcome into the thread; the orchestrator's auto-summary is now a fallback that never posts a bare "✅" with no content and skips threads already answered directly.

Removed

• Reverted the unfinished compact-protocol feature — it was only half-wired (main-side committed, renderer-side unfinished) and broke the web typecheck. It will return fully wired in a later release.

[0.2.4] — 2026-06-09

A multi-provider patch: Codex graduates to full hive parity via a native lifecycle-hook bridge, the god orchestrator opens to its terminal by default, and a handful of resilience fixes land.

Added

• Codex lifecycle-hook bridge — full hive parity. Codex now joins the hive as a first-class, hive-aware provider: a native lifecycle-hook bridge maps Codex's events into the existing hook pipeline (live status + inbox-drain + outbox routing), and agy/codex dispatch is unified behind one path. Verified running hive-aware in bypass-permissions mode. (#47, #54)
• Codex hook discovery via config.toml [hooks]. The bridge registers through Codex's config.toml [hooks] surface rather than a bare hooks.json, matching how Codex actually discovers lifecycle hooks.

Changed

• God orchestrator opens to the Terminal sidebar by default. Selecting the god agent no longer reopens a stale "ASK ME" tab — a leftover command-center tab request is cleared on select, so the panel mounts to its terminal default. The ASK ME tab is still one click away.
• Landing + blog refreshed for multi-provider. The landing page now presents Claude Code, Antigravity (Gemini), and OpenAI Codex as equal first-class providers (with a one-line mobile-friendly badge), and a grand v0.2.4 launch post + technical walkthrough replace the v0.2.3 posts.

Fixed

• Slack/webhook tunnel no longer crashes at load. tunnelmole is ESM-only; a static import in the CommonJS-bundled Electron main process threw ERR_REQUIRE_ESM. It's now loaded via a dynamic import() inside openTunnel(), so the public ingress actually starts.
• Heartbeat re-engages the god on an unread actionable inbox — not only when the floor is quiet — so worker/human mail is drained promptly.
• Slack done-summary stops retrying on terminal errors. A permanently-failing post (e.g. the bot token missing chat:write → missing_scope) is now recorded and logged once instead of retrying every 5s and flooding the console; transient errors still retry.

[0.2.3] — 2026-06-09

A multi-provider release: the floor is no longer Claude-only. Antigravity (Gemini) and Codex agents become first-class hive participants, schedules get their own tab, and the Slack / webhook ingress is moved off the flaky public tunnel.

Added

• First-class Antigravity (Gemini / agy) provider. A worker can now run the Antigravity CLI as a full hive participant. Because agy has no Claude-style --append-system-prompt/--settings hooks, the hive identity + protocol ride in as the session's initial prompt, and a native agy-hook bridge normalizes Antigravity's lifecycle events into the existing hook pipeline so a Gemini worker gets the same live status + inbox-drain as Claude — on the subscription, no API key. (#54)
• Schedules tab. Recurring auto-dispatched missions (and the adaptive heartbeat) get their own Command-Center tab instead of an inline section. (#50)
• Terminal work-order handoff for hookless providers. A provider with no inbox-drain path now receives hive mail as a WORK ORDER FROM HIVE typed into its terminal, falling back to a god-bounce only if the renderer is unavailable. (#53)

Fixed

• Codex agents now follow the hive protocol and message back. Codex spawned without the hive protocol or any hook, so it never read its inbox or wrote its outbox. Codex is now a non-hive-aware-but-inbox-capable provider: the protocol is injected as its initial (positional) prompt, its outbox is drained provider-agnostically by the router, and inbox mail reaches it via the renderer's idle inbox-wake nudge. Codex and Antigravity coexist in the provider union. (#47, #54)
• Slack + webhook public URL no longer silently breaks. The ingress used localtunnel/loca.lt, which now serves a browser interstitial that fails Slack's url_verification POST and breaks saved webhook URLs. Both slack.ts and webhook.ts now use tunnelmole (MIT, POSTs pass straight through), and a failed tunnel surfaces a real error instead of a silent "started" with no URL.

[0.2.2] — 2026-06-07

A community polish release — almost entirely the work of @Gulum: a live context-window gauge on every agent card, sharper terminals, correct Windows metering, and dispatch that always routes through the god.

Added

• Live context-window gauge on each agent card. A Claude Code statusLine pushes the session's exact token count and real context-window size after every response, so each agent card shows a precise live fuel gauge drawn from Claude Code itself instead of a transcript estimate. The gauge also zeroes the instant you send /clear, rather than briefly showing the previous session's full bar until the next response. (Thanks @Gulum — #12, closes #11.)
• Per-session terminal theme toggle + Unicode 11 emoji widths. Each terminal session can now switch its Claude theme independently, and emoji column widths follow Unicode 11 so wide glyphs stop nudging the cursor out of alignment. (The WebGL renderer, copy/paste, and minimumContrastRatio from v0.2.0 are kept as-is.) (Thanks @Gulum — #26.)
• All human dispatch flows through the god. Every Command Center dispatch now mails the god (Task from the human) instead of writing straight into a worker's inbox; the worker dropdown becomes a suggested owner (Michael still decides), so nothing skips the orchestrator. (Thanks @Gulum — #45, fixes #44.)
• Dedicated context-window row on the monitor tab. The Floor monitor's cumulative budget bar was being misread as a per-agent context gauge; a separate ctx context-window row now sits alongside it, so the live context window and the cumulative budget aren't confused. (Thanks @Gulum — #46.)

Fixed

• Windows usage meter no longer reads 0/0. The transcript reconciler built the per-project directory name with the POSIX rule, but Claude Code on Windows encodes every non-alphanumeric character (including the drive colon), so the meter never found the transcript and always read 0 tokens / $0.00. (Thanks @Gulum — #34, fixes #10.)
• Send-only assistant mail no longer black-holes. Direct mail to the send-only prep assistant landed in an inbox nothing reads; the router now bounces it to the god (subject prefixed [bounced …]) instead of dropping it. (Thanks @Gulum — #33, fixes #32.)
• Boot banner no longer stacks in scrollback. tryFit() fired resizePty on every fit even when the dimensions were unchanged; redundant resizes are now skipped, so the boot banner stops re-stacking in the terminal history. (Thanks @Gulum — #8.)
• Visible text-select cursor on the cream theme. The hovering I-beam (an OS cursor that CSS color hints can't touch, drawn white by several Windows schemes) is now an inked I-beam with a halo, so it stays visible over the light terminal. (Thanks @Gulum — #39.)

Acknowledgements

This release is almost entirely the work of @Gulum — thank you. Maintained by @chaitanyagiri.

[0.2.1] — 2026-06-07

A small follow-up to v0.2.0 that makes the scheduler considerate of agents that are mid-task.

Changed

• Scheduled auto-compaction is queued, not forced. The hourly ops-standup's terminal compaction is now enqueued per agent and delivered only when that agent is idle (deduped — at most one /compact pending at a time), so it compacts between steps instead of jamming a working terminal mid-step. The standup prompt now asks each agent to summarise its current task and next step, then resume from the same point after compacting.
• Heartbeat is inbox-driven. The floor heartbeat (reengageGod) no longer types directly into Michael's terminal; it drops its digest in his inbox, which the busy-aware inbox-wake delivers once he's idle.

Docs

• Expanded the README roadmap (chat integrations, pluggable agent CLIs, realtime Michael).

[0.2.0] — 2026-06-07

The observability and control release. v0.2.0 makes the fleet visible and keeps it in check — and it's a community release in the most literal sense: most of the work below came from external contributors. Huge thanks to everyone credited.

Added

• Command Center overhaul. Michael's control surface was reworked into the place you actually run the floor from — the roster, dispatch, schedules, memory, and activity views now carry the new live signals (token budgets, telemetry, breaker state) without becoming a wall of numbers.
• Per-agent token budgets + live fleet monitoring. Every agent carries a token budget, and the floor monitors consumption live so a single agent can't quietly run the bill up.
• Live OTel telemetry collector + per-model cost. A built-in OpenTelemetry collector and a UsageProvider seam feed real usage in, with per-model cost attribution (interim transcript-backed stub behind the seam to start).
• Fleet grid + per-agent tool-span waterfall. A live grid of the whole fleet, plus a per-agent tool-span waterfall that shows what an agent spent its turn doing — which tool calls ran, in what order, for how long.
• Agent-card context-window gauge. The agent card's progress bar is repurposed into a context-window gauge so you can see at a glance how close each agent is to filling its context. (Thanks @Gulum — #12.)
• Circuit breaker. A steer → constrain → stop ladder plus a cost/runaway guard, fed by hook signals (repeated identical tool calls), an onApiError seam for error-storm trips, and budget config.
• Scheduler heartbeat. A heartbeat beat tracks each agent's last output (quiet/idle signals), enforces circuit-breaker policy, and adds spawn guardrails; the SCHEDULES view shows the heartbeat row plus last-fired / next-fired times. (Thanks @albozes — #2.)
• Human-in-the-loop, mid-run. A HITL gate, mid-run steer, and graceful stop all delivered through hook returns — approve, redirect, or cleanly halt an agent mid-turn instead of yanking it.
• Durable SQLite persistence (Phase A). A SQLite durable store persists window bounds and history, alongside a persisted session_id and a durable cost ledger (cost-ledger.jsonl) so cost and provenance survive restarts.
• MemoryReflector — memory condensation. The janitor's missing condense half: a reflector that condenses memory instead of only mining it, keeping the semantic store lean.
• Configurable hive/memory home folder. Point the hive and memory home at a folder of your choosing, with a safe move that relocates existing data.
• One-click "Restore team." After a harness restart, a single click brings back the last session's workers — no more re-adding agents one by one. (Thanks @Gulum — #16.)
• Delete scheduled missions. Scheduled missions now have a delete button. (Thanks @Gulum — #9.)
• New avatar states. A compacting state (on PreCompact) and a looping state (when the breaker engages) so the floor reflects what the control layer is doing.

Fixed

• Terminal contrast + HiDPI legibility. A minimumContrastRatio floor and a tuned light palette (including dual fg/bg-legible green/yellow) keep text legible on coloured backgrounds across both the new and legacy terminal views.
• Crisp, readable floor text. A HiDPI canvas, bold bubbles, and a walk-flicker fix sharpen office-floor text; thought-cloud text now stays 1:1 when the window shrinks. (Thanks @Gulum — #20.)
• Terminal no longer jumps to the top of history on first scroll, and the viewport dead zone is gone (re-sync routed through xterm, not the DOM). (Thanks @Gulum — #8.)
• Windows: keep the hive running behind the lock screen. The hive no longer freezes when Windows locks — keep-awake plus no throttling. (Thanks @Gulum — #18.)
• Live agent statuses and composer-draft fixes for the message-queue composer. (Thanks @Gulum — #7, #27, #28.)
• Palace writer-lock serialization and Windows named-pipe + mempalace detection so the hook server and semantic memory work on Windows. (Thanks @Xileck.)
• Per-PTY input serialization so the boot sequence can't jam mid-spawn; restored +x on the node-pty spawn-helper so agents can spawn.
• GOD orchestration tabs are now scrollable; the title-bar settings button is a clear gear chip.
• Global defaultModel wins over role tier (an explicit per-agent pick still wins); cost-ledger row is fully snake_case for a 1:1 SQLite migration (#4).

Acknowledgements

Reported / requested by the community: @JLAD75 (Windows hive router / hooks.sock — #1), @billrehm (Windows GOD-spawn error 193 — #22), @darrensheffield (uv-not-installed assumption — #30; macOS Gatekeeper — #29), @pdurlej (first-class Codex CLI provider request — #21), @wild-gobatz (agents showing idle until clicked — #3). Maintained by @chaitanyagiri.

[0.1.9] — 2026-06-06

Added

• Hourly ops standup. A built-in scheduled mission (enabled by default) where the GOD orchestrator reviews every agent — who's doing what, whether tasks are on track, and whether agents are still running — and compacts each terminal's context on the same hourly cadence to keep agents lean. Toggle it in the Command Center; a one-time migration seeds it into existing installs (and won't re-add it once deleted).

Fixed

• Agents exited on their own at a "Bypass Permissions mode" prompt. Agents spawn with --permission-mode bypassPermissions, which on a fresh machine shows a one-time interactive "WARNING: Bypass Permissions mode … 1. No, exit / 2. Yes, I accept" prompt the terminal couldn't answer, so the agent exited code 1 within seconds. The harness now idempotently pre-accepts Claude Code's dangerous-mode warning (skipDangerousModePermissionPrompt / skipAutoPermissionPrompt) and per-folder trust before each spawn.
• Blog cards. The colored thumbnail tile sat flush against each card's bold border; it's now inset with padding for breathing room (desktop + mobile).
• Windows: hook server + semantic memory. The hook server now binds a named pipe on Windows (where Node IPC isn't a filesystem socket), and mempalace is detected via where + the standard .exe install locations so semantic memory works on Windows. POSIX behavior unchanged. (Thanks @Xileck — #4.)
• Palace mining writer-lock collisions. Mining is now serialized — a single writer at a time with a re-entrancy guard — fixing "palace is held by PID …" failures when multiple agents mined concurrently. (Thanks @Xileck — #5.)
• MemPalace index noise. Each agent's .gitignore is ensured before mining so settings.json, the cursor file, and raw inbox/outbox message JSON stay out of the semantic index — keeping recall and wake-up focused on real memory.

[0.1.8] — 2026-06-05

Fixed

• Windows agent spawn. Launching an agent on Windows failed with cannot create process, error code: 2 (ENOENT) because binary and PATH resolution were Unix-only (SHELL//bin/zsh, -ilc, which, and Unix-only fallback paths). Windows now resolves claude via where, checks the standard Windows install locations (%APPDATA%\npm\claude.cmd, %LOCALAPPDATA%\Programs\claude, %USERPROFILE%\.claude\local), uses the process PATH directly (no login-shell probe), and recognizes Windows-style (\) absolute paths. macOS and Linux resolution is unchanged; the Unix fallbacks now also include ~/.volta/bin.

[0.1.7] — 2026-06-04

Added

• Slack → Michael's queue. A new Slack integration (Settings → Slack) pipes a channel's messages straight into Michael's message queue — paste a message in Slack and it lands in his queue exactly as if you'd typed it. Off by default; every request is verified with your Slack signing secret (HMAC + 5-minute replay guard) before it's accepted, and a localtunnel exposes the local webhook for Slack's Event Subscriptions.

Changed

• Approvals are now native. The in-app approvals queue/panel is removed in favor of native Claude Code human-in-the-loop prompts. A to:"human" decision now reaches you through Michael's session and native permission prompts — approvable from your phone via /remote-control — and Michael boots straight into running the floor.

Fixed

• The floating approvals panel could re-queue an item when you approved it (resolveApproval re-routed the message back into the queue). Moving to native HITL removes the panel and the bug.

[0.1.6] — 2026-06-04

Added

• Per-agent git worktrees. A 'Git isolation' toggle in Add Agent auto-provisions a dedicated worktree (<harnessHome>/worktrees/<agentId>/) on spawn and tears it down on kill. Agents on the same repo never collide on branches.
• Task kanban with dependencies. A Tasks tab in the Command Center renders a full kanban board (todo / doing / blocked / done). Each task carries an assignee, a dependsOn[] list, priority, and description — and persists in hive/tasks.json via a new hive:writeTasks IPC channel.
• Scheduled missions. A Schedules section in the Floor tab lets you define recurring auto-dispatch missions (label, interval, target agent, body). The main process fires each on a setInterval, stamps lastFiredAt, and persists the list in config.
• Real token & cost telemetry. The Activity tab reads ~/.claude/projects/ JSONL transcripts — the same files Claude Code writes — and displays actual input/output/cache token counts and estimated USD cost per agent per model. No more proxy tool-call counts.
• Global hive text search. Full-text search across board.md, tasks.json, and all agent memory.md files, available in the Memory tab alongside MemPalace semantic search.
• Threaded chat. A Messages tab in each agent's sidebar renders every hive message grouped by conversation with full reply chains and an inline reply form.
• Memory graph. A visual graph in the Command Center Memory tab maps agents and their knowledge relationships.
• GitHub issue ingestion. An Issues section in the Floor tab pulls open issues from any registered repo via gh issue list and lets you assign them to any agent with one click.
• CI status watcher. A CI Status section in the Activity tab polls gh run list for every registered repo and shows live pass/fail/in-progress status for GitHub Actions runs.
• Desktop notifications. Native OS notifications fire when an agent finishes a task or is waiting for your input. Toggle in Settings.
• Agent archival. Closing an agent's tab archives it (memory + history intact) rather than deleting it permanently.

Fixed

• Scheduler now honors lastFiredAt on config reload — missions don't double-fire after a save.
• PTY lifecycle teardown runs on natural process exit as well as explicit kill, so worktrees are cleaned up reliably.
• Task IDs fall back to a stable UUID when the title is empty; writeTasks IPC validates its input.

[0.1.5] — 2026-06-04

Added

• Dwight, Michael's prep assistant. A persistent, visible assistant agent (Sonnet, 1M context) spawns on startup. A global enrich toggle routes Michael's queued prompts through Dwight first — he gathers repo context and rewrites the prompt, then forwards it to Michael through the hive; toggle it off and prompts go straight to Michael.
• Michael's Command Center. His sidebar becomes a control surface with Terminal, Floor (agent roster + per-agent model selector with safe restart, a dispatch box, and working dirs), Memory (MemPalace search + per-agent memory), and Activity (live log feed + board + usage proxy), plus a copyable Claude command handbook.
• Per-agent model selection — a model picker in Add Agent, a shared model list, and a message-queue composer with the enrich toggle.
• Getting-started tutorial on the blog (canonical install + first-run walkthrough), with Blog/tutorial CTAs and a redesigned "How it works" section on the landing page.

Fixed

• Agents no longer read "idle" while still working — a Stop blocked mid-turn now reports blocked so the UI keeps the agent in its working state.
• Long agent thought/tool labels now word-wrap inside their cards instead of overflowing the bubble horizontally (Pixi word-wrap with a raw-length cap so a pathological string can't grow a runaway-tall card).
• Switching agent terminals now lands at the latest output, while resizes preserve scroll position; the idle action label no longer echoes the "idle" badge.

[0.1.4] — 2026-06-04

Added

• Signed macOS builds. The app now ships with a hardened-runtime Developer ID signature (notarization is attempted in CI and stapled when it succeeds; the build is best-effort, so a notarization hiccup never blocks a release). Because macOS binds a folder-access (TCC) grant to a stable code signature, you're now prompted for Documents/Desktop/Downloads access once instead of on every agent action. Usage-description strings explain each prompt. Signing/notarization run in CI only when Apple credentials are present, so contributor builds stay unsigned and green.
• Blog at /blog — an Eleventy-generated static blog sharing the landing page's neo-brutalist design system, seeded with the first posts on long-term memory, multi-agent harnesses, and MemPalace, plus tag/topic indexes and an RSS feed.
• On-site SEO/AEO metadata — JSON-LD, robots.txt, a root sitemap.xml, and richer link-unfurl/meta tags across the site.

[0.1.3] — 2026-06-01

Added

• Settings panel (title-bar gear) with a Reset & start over action that wipes Michael's memories, the entire hive (every agent, message, task, and the board), and the semantic-memory palace, then relaunches the app into onboarding.
• Boot loader ("clocking in") shown while the GOD agent initializes, so returning users no longer see the empty "add agent" screen during startup.

Fixed

• Crash dialog on quit caused by sending IPC to an already-destroyed window during teardown; all renderer sends are now destroyed-safe and shutdown steps are best-effort.
• Michael no longer marches to the door flagged "needs you" right after finishing a turn — idle "waiting for input" notifications now let him linger at his desk instead of escalating as a blocked/needs-action state.

[Brand & rename]

Added

• Brand identity: Munder Difflin — logo (docs/logo.svg), square mark (docs/logo-mark.svg), and hero banner (docs/banner.svg).
• Landing page at docs/index.html (GitHub Pages–ready).
• In-app branding: window title, boot screen, title-bar MD badge, and fullscreen header captions.
• Open-source community files: SECURITY.md, CHANGELOG.md, issue/PR templates, and a CI workflow.

Changed

• Renamed the project from Claude Terminal Harness to Munder Difflin across the README, docs (SPEC.md, DESIGN.md, HIVE.md), package.json, and the app UI.

0.1.0 — 2026

Initial working prototype.

Added

• Electron + React + TypeScript shell (electron-vite).
• Real terminals via node-pty, rendered with xterm.js; multi-agent spawn/write/ resize/kill over typed IPC (window.cth).
• Pixi.js office floor: Tiled map, camera, recolored cast, pathfinding, seat assignment, tool bubbles, and message envelopes.
• The hive: on-disk multi-agent layer (hive.ts), hook server + cth-hook shim and Stop-loop (hooks.ts), and a semantic memory layer (memory.ts).
• GOD orchestrator agent, approvals queue, and memory search panel.
• Sandboxed file browser + CodeMirror editor and a git tab (status, log, branches, commit graph).
• Onboarding wizard, safe-quit guard, and a tokenized SNES/Animal-Crossing design system.