Closing an agent terminal can no longer kill an unrelated program, the terminal font is fixed on Electron 45, and Guppy makes no font requests. Rollback: 1.1.105 (reinstall).
Fixed
Closing a terminal kills only that terminal's processes. On Windows, closing or restarting an agent terminal killed its processes by process id. If the terminal's shell had already exited, Windows could have given that id to another program, and the kill could reach it. Guppy now holds the terminal's own process from the moment it starts (node-pty hands over its process handle), so the id cannot be reused while any kill might still name it. Every kill path checks that the id still belongs to that terminal, including node-pty's own clean-up after a close.
A terminal closed very early no longer fails. Closing an agent terminal before its first output could raise an error, or try to connect to a pipe that was already gone (ENOENT). Both cases now close quietly.
Terminal text is no longer clipped or overlapped. The 1.1.105 terminal font was measured wrongly on Electron 45. The terminal now uses the bundled Martian Mono (then Consolas, then monospace), with whole-pixel cells that are never narrower than a character.
A floor can no longer lose its team on disk. A first write that held only god could replace the saved team. It is now refused.
A second Guppy that lost the instance lock writes nothing, at start or on quit, and does not run the startup orphan sweep.
Claims: the derived-claims writer stops on a read-only ledger instead of retrying, and a failed background worker no longer stalls the queue behind it.
Added
Codex agents finish what they started. When a Codex agent ends its turn while a request is still owed, Guppy continues it (at most 8 times per chain), and a Codex continuation waits up to 5 minutes.
The Blob looks around. Its eyes follow the cursor, a hovered ring slot, and the text field while you dictate. The overlay, the tray icon and notifications share one mood, the tray is idle at rest, and Blob faces can be saved as images.
Changed
Every UI font is bundled. Guppy no longer requests fonts from the internet.
Faster main process on big floors. Watch and integrity checks, and the router's catch-up for missed changes, now run asynchronously instead of blocking the main thread.
Installing from source rebuilds only node-pty for Electron and patches node-pty during install (npm ci). A worktree check (tools/worktree-setup.cjs --check) names a missing or unpatched native build before the tests run.
Please check (live, on this release)
Terminals: open and close several agent terminals, including one closed straight after opening. Confirm that nothing else on the PC closes and no error appears.
Terminal font: confirm that the terminal text is sharp and that no characters overlap.
Codex: give a Codex agent a build request and confirm that it carries on past planning.
Guppy no longer freezes at startup on a large hive, runs on Electron 45, and the claims ledger can be created from Settings. Rollback: 1.1.103 (reinstall). 1.1.104 could freeze at startup on a large hive.
Fixed
Startup no longer freezes on a large hive. 1.1.104 could freeze for many seconds at start on a hive with tens of thousands of handled messages and a large claims ledger. It re-read every handled message and the whole ledger on the main thread, in long blocks. Startup now reads only what changed since the last run. The claims work (claim status, the search index, the legacy import and the derived claims) runs in short slices or in a background worker thread. If the worker cannot run, Guppy does the same work the old way and logs it. On a test hive ten times the size of the one that froze, no pause after startup was longer than about 0.6 s.
Large floors stay responsive. Messages and Inbox read only new mail, in short slices off the main thread, and Threads shows "Loading conversations…" until the first answer. The task list is one cached view. The usage check looks only at the transcripts and Codex rollouts that changed.
Plans with no 5-hour limit now show their weekly capacity instead of "Capacity unknown". Some Claude plans report only a weekly limit. The capacity strip used to look for a 5-hour figure, find none and say "Capacity unknown". Now it shows the weekly figure, with its pie, in the strip, the provider details and the agent usage line. If the plan reports a 5-hour limit again, the strip goes back to showing it. Before the first live reading after a restart, it still says unknown.
The agent roster no longer redraws itself every 5 seconds. Its background check for task badges and waiting messages redrew the whole roster strip on every check, even when nothing had changed. Now it redraws only when the answer changes.
Added
Settings > General > Startup health. A read-only line with the result of the last start. For the first 2 minutes after the window opens, Guppy watches its own main thread. Any pause of a quarter of a second or more counts as a slow moment and is written to the log with what was running.
Create the claims ledger from Settings. Settings > Memory & Knowledge > Claim ledger > "Create claims ledger" imports each agent's legacy memory, one idle agent at a time. Busy agents are recorded and retried on the next request, and nothing is deleted. Progress is shown while it runs.
Settings > Blob. The Blob has its own tab, with a link to a five-minute set-up guide, the Groq API key (shared with the Voice tab) and the overlay settings.
Changed
Guppy runs on Electron 45 (was 32). Terminal paste, the clipboard and the native modules were moved to it.
Building from source needs Node 24 LTS (.nvmrc, engines and CI). Node 20 is end of life. Settings > Prerequisites > "Build from source" checks for Node 24.
Security
The file guard checks where an open file really is. It compares the opened file's actual location with the allowed folders, not only the requested path. It refuses files with more than one hard link, and it compares paths folder by folder, so a name that starts with ".." is no longer refused by mistake.
The installed app's Electron fuses are set at build time. The app loads only from its checked app archive, and the release build does not accept Node inspector arguments.
Please check (live, on this release)
Startup: start Guppy on your own hive and confirm the window answers at once. Then open Settings > General > Startup health after 2 minutes and check that it reports no slow moments.
Claims ledger: in Settings > Memory & Knowledge > Claim ledger, run "Create claims ledger" and confirm the progress line finishes and agents' memory still answers.
Terminal paste: right-click paste and Ctrl+V in an agent terminal.
1.1.104 Not recommended
Note: 1.1.104 is not recommended. It could freeze at startup on a large hive. 1.1.105 fixed this: if you are on 1.1.104, update to 1.1.106.
The Guppy Blob: an on-screen helper with hold-to-talk dictation into any app, Guppy can keep running in the tray, and Settings checks what building from source needs. Rollback: 1.1.103 (reinstall).
Added
The Guppy Blob. An optional always-on-top blob you can drag anywhere. Click it for a ring of actions: record message, invisible and close; screenshot, record meeting and screen are shown as coming soon. Turn it on in Settings > General > Blob, where you can also set its show/hide hotkey and hide the Blob, or Guppy's own windows, from screen shares. It stays on screen when a monitor is removed or the layout or DPI changes, and it recovers once if its window crashes. Off by default.
Hold-to-talk dictation (Windows). Hold Ctrl+Alt+Space in any app and the Blob shows that it is listening. Release, and the transcript (Groq speech-to-text, using the Groq key in Settings > Voice) is typed into the field you were in. If you switched windows before it finished, nothing is typed and the Blob says so. A microphone error shows at once. Clips run from 0.3 s to 120 s, up to 25 MB. Off by default (Settings > General > Blob).
Keep Guppy running in the tray on close (Windows). With Settings > General > "Keep running in the tray when closed" on, closing the window hides Guppy in the notification tray, and the primary window stays reachable from the tray (Open Guppy, Show Blob, Quit Guppy). Agents keep running and using tokens while hidden, and the PC does not sleep on its own until you quit Guppy from the tray. Off by default.
Settings > Prerequisites: "Build from source". An opt-in group that checks what this PC needs to build Guppy itself: Node 20, a Python that node-gyp can find, and the VS 2022 C++ build tools with Spectre-mitigated libraries. Each missing item shows the exact command to copy; nothing installs on its own. The essential prerequisites are unchanged.
Changed
Building from source is pinned to Node 20, matching CI (.nvmrc and engines). Node 26's PGO build broke the native modules (better-sqlite3).
Docs: the README, the MIT LICENSE and the release, telemetry and contribution docs from main are carried into this release.
Fixed
Test-only: the path-containment tests run on Windows without the symlink privilege (they cover junctions), and several wiring pins were fixed so the suite passes on a clean install.
Please check (live, on this release)
Tray: turn on "Keep running in the tray when closed", close the window, and confirm Guppy stays in the tray. Open Guppy from the tray (click and menu), use Show Blob, then Quit Guppy from the tray and confirm it exits.
Dictation: with a Groq key in Settings > Voice and dictation on, hold Ctrl+Alt+Space in another app, speak, release, and confirm the text is typed into the focused field.
Blob overlay: enable it, drag it, open the ring, toggle invisible, and check hide-from-shares.
New floors get the claims ledger with no setup, search sees a first claims sync while it runs, and replies no longer race the startup mail catch-up. Rollback: 1.1.102 (reinstall).
Changed
The claims ledger is on by default for new floors. A floor created on 1.1.103 builds its mail-cards index and its derived claims at first start, with no manual step, and keeps both current. If the floor's project folder is a git repo, its commits feed the ledger too. A plain folder works from mail alone, and its commits are picked up once the folder becomes a repo. Settings shows the repo the ledger reads. Floors made before 1.1.103 keep their settings. At every start, each floor indexes any mail it has not indexed yet, so the first start after the upgrade can take longer on a floor with a lot of history.
memory get <key|card-id> returns the current value of a key or a card as a field. After it come the competing values and the superseded ones, each with what replaced it. memory search adds --exact and --newest, and memory history accepts the ids that get prints. The memory tools are also offered over MCP.
Agents start one fresh conversation after this install. The protocol text injected into sessions now names memory get, so each affected session starts a new conversation the first time it starts on 1.1.103. Identity, memory, inbox and ledger are kept.
Fixed
A floor's first claims sync is searchable while it runs. A first sync (or one after a reset) used to apply every derived claim in a single step at the end, so search found none of them for minutes. It now applies whole claims in slices, and search sees the first slice within seconds. A sync that stops part-way resumes where it left off and embeds only what is missing.
Replies sent during the startup mail catch-up are linked, not bounced. A reply that arrived while the catch-up was still reading the hive could miss its parent. It is now delivered as written and linked once the catch-up settles, with rows kept in arrival order.
The derived claims link commits, done mails and card status more carefully. Fixes cover merge and cherry-pick commits, branches cut after a merge, a commit subject naming another card, superseded cards, and slow backfills on large boards. Accuracy validation is in progress; the last sealed test did not meet its 0.90 bound.
Heavy jobs are held to half the PC, closing a terminal ends everything it started, and Codex agents keep out of your own Codex folder. Rollback: 1.1.101 (reinstall).
Changed
Heavy jobs are hard-capped at 50% of the PC's CPU, on by default. The Windows option that limited heavy jobs to half the processors is replaced by a hard cap at 50% of the whole machine's CPU rate. Every processor stays visible, so tools that size themselves from the processor count are no longer squeezed onto half of them. The new Heavy jobs CPU cap switch in Settings is on by default and applies at once to jobs already running; turning on the old half-processors option carries over as on. If a heavy job runs inside another job that refuses the cap, its terminal says that the cap was not applied, and the command still runs.
The Settings hero card follows the current release. The card and its "full changelog" link read from the latest release (or the repository's default branch), found when the card refreshes, instead of a fixed branch that was no longer kept up to date. Offline, it shows its last copy or the built-in text. A failed refresh is not retried for an hour, and redirects are followed over https only.
Fixed
Closing an agent's terminal ends every process it started. On Windows, closing a terminal shut its console first, so processes that ran outside that console (such as Antigravity's inner process, or a detached helper) kept running. The app now ends the whole process tree first, and tries again if anything survives, before it closes the console. Quitting the app finishes any close still in progress.
Codex agents no longer point into your own Codex folder. Each Codex agent's settings are copied from your own Codex config. In regular builds, the copy kept paths into your own ~/.codex (the Node REPL MCP server's home and trusted paths, and the plugin marketplace source). Every such path now points into the agent's own Codex home, in every build.
Settings changes no longer undo each other, the main process reads its files far less, and test runs that open windows wait their turn. Rollback: 1.1.100 (reinstall).
Changed
Config, registry and task reads are served from memory. The app kept re-reading config.json, the agent registry and tasks.json from disk many times a minute (about 2,900 config reads a minute during a heavy run). Each is now read once and kept until the file changes: the app's own writes refresh it at once, and an edit from outside the app is picked up within a second. Every change that reads, edits and writes a file still reads it fresh, so an edit made meanwhile is kept.
A filtered test run that opens app windows waits for the heavy slot. A test run filtered to a few files is now judged by the files it picks, as running those files directly is. If any of them starts the app's test window (directly or through a helper), the run waits for the heavy slot. Before, only a filter matching 20 or more files did. This applies to npm test, npm run test* and the test runner itself, called by any path, including a network share.
Fixed
Turning on a second MCP server no longer turns off the first. In one visit to the settings panel, enabling a second MCP server reverted the first one. Each toggle now changes only its own server.
Toggling the knowledge graph keeps its folder. Turning the knowledge graph on or off no longer erases its root folder setting.
The auto-compact toggle and Schedules no longer undo each other. Changing auto-compact in Settings, or a schedule in Schedules, could put back the other's earlier state, or a schedule's last-run time. Each now changes only its own schedule. Registering or removing a project likewise changes only that project, so a project added meanwhile is kept.
The heavy-job lock judges commands correctly, heavy jobs hand over cleanly, and a new floor starts empty. Rollback: 1.1.99 (reinstall).
Added
Name the next heavy-job holder (heavy next <agent>). The agent holding the heavy slot, an agent with the slot reserved, or god can say who goes next, and the slot passes to that agent when the job ends.
Changed
A broad test filter counts as the full test suite. A test run filtered by a word that matches 20 or more test files now waits for the heavy slot. This applies whether the filter is given to npm test (including with --prefix) or to the test runner directly, and from any folder. Before, a broad filter ran light next to another agent's suite.
The lock sees through more ways of starting a heavy command. It now catches a node script that starts the suite, a PowerShell loop over a list of commands, and a command wrapped in brackets or backticks. It also catches a command kept in a variable and a script written just before it runs. It knows the shell a Codex agent's command runs in.
A running heavy job keeps its slot. A job the app last saw running is renewed at the hour instead of being freed while it still runs. A reserved slot that lapses puts its owner back in the queue once. heavy status says where the asking agent stands.
heavy release from the holder works. When the held job has ended, the slot is freed and the next agent is let in at once. While the job still runs, the release is refused and the job is named. Before, it answered "nothing to release" while the slot showed as held.
A release keeps every earlier release's notes. Building a release now fails if a released section of this changelog was dropped or edited, as happened to the 1.1.98 heading in the first 1.1.99 build.
Fixed
A new floor starts with only its own god. A new floor no longer brings back the last floor's agents. Any agents it set aside are saved in the new floor's roster-backups/ folder. A fresh home from Settings now keeps your display preferences, such as theme and terminal font.
Light commands are no longer held as heavy. A script under /tmp in Git Bash, a PowerShell line that only sets a variable, and a loop variable in a PowerShell foreach no longer wait for the heavy slot. A command that runs a single test file, or adds a git worktree, stays light.
No duplicate turn after a Codex stop. A message a Codex agent was shown when it stopped now counts as delivered, so it is not typed into the agent a second time.
The window stays smooth while agents work, a new floor knows its tools, and fewer wakes get stuck. Rollback: 1.1.98 (reinstall).
Added
A new floor knows the floor tools (NEW-FLOOR-KNOWHOW). The per-agent compaction window, the claims ledger, the fleet settings and the floor commands are now written into what Guppy ships (the protocol, the command list and the coordinator's start-up notes), so a fresh floor no longer depends on what older agents happen to remember.
God can settle a typed-over wake (GTC term rule). When an agent is held because a wake may have been typed over, the coordinating agent can now rule "send again" or "already handled" from the terminal, instead of the agent waiting until you click. The ruling names the exact hold it answers, so it cannot settle a newer one by mistake.
A test run started by a Codex agent cannot be started twice. The related-tests runner keeps a status file and refuses a second copy while the first is alive, and Codex agents are told that their 30-second command limit does not mean the run was killed.
Changed
Saving agent activity no longer freezes the window (UI-CRAWL). The activity and capacity records are written in the background instead of on the window's own thread, and closing the app still saves the latest copy.
Quiet mail waits until it matters (QUIET-HOLD-ACT). Only unsolicited notes are held back; a reply someone asked for, or a message that relays your words, wakes the agent at once.
Codex session files stay small (CODEX-ROLLOUT-SIZE). A long Codex session's log file is rotated safely instead of growing past 100 MB, which also stops the antivirus rescanning it on every line.
Fixed
Agents wrongly shown as typed over. A wake sent right after the app starts is no longer judged typed over when there was no screen to look at, the hold clears once the mail is delivered, and a key you press always wins over the app's own check.
Read-only log scans no longer wait behind a heavy job. A path held in a shell variable, or a plain folder, is no longer mistaken for a program that needs the heavy slot.
Test runs leave no PowerShell files behind. PowerShell started by a test gets a temp folder of its own, and folders left by a killed run are cleaned up later.
Agent mail arrives without anyone typing, heavy jobs share the machine fairly, and cost figures are right. Rollback: 1.1.97 (reinstall).
Added
A stall watcher records every app freeze (STALL-WATCH). Each time the window stalls, the app writes down how long and what it was doing, so the typing lag can be traced to its real cause.
An alert when an agent's wakes keep being refused (WAKE-REFUSAL-STUCK). Mail that cannot be delivered for a long time now raises a notice instead of waiting in silence.
Unexpected agent exits are logged (PTY-EXIT). The exit record keeps the last screen lines, with tokens, keys and PEM blocks blanked out.
A crashed app window recovers by itself (ELECTRON-VIEW-SELF-HEAL).
Changed
Heavy jobs take turns in a real queue. A finished job frees its slot at once (it used to stay locked for 30 to 40 seconds or until a timeout), waiting jobs are no longer dropped after an hour, and short commands are no longer mistaken for heavy ones.
The task board is split into smaller files (TASKS-JSON-SPLIT). A small change no longer rewrites one 1.45 MB file.
Fewer redraws. Parts of the window no longer redraw when nothing changed.
Hooks never call a paid model. The hidden background memory tidy-up (about $0.30 to $0.70 a day) is gone, and the app enforces that hooks run without a model.
Old release work folders are pruned (npm run cut:prune), and test runs remove their temporary files. The folders had grown by about 21 GB a day.
Only the coordinator marks a card done, and only once its work is merged.
Installer builds run through one script (npm run cut:win) that installs, builds and checks the native modules the same way every time.
Fixed
Wakes typed but never sent. A wake left sitting in an agent's message box is now sent or cleared.
Mail held behind a stale "loading" label. A resumed Codex agent whose screen kept an old model: loading header now takes mail once its own status line shows.
Dwight stopping mid-task after sending an in-between update.
Cost figures. Missing price rows made some costs about 3 times wrong, and the fleet view's dollar totals read too low.
The Slack and webhook listeners listen only on this PC, no longer on the whole network, and internal hooks no longer carry access tokens they do not need.
Codex agents no longer drop typed messages or mouse clicks, and show their activity on the fleet view.
Messages with a blank subject, wakes sent to the wrong group, and a loop guard that did not count.
A cancelled test run always restores the code it changed, and the suite passes with the half-cores setting.
Your message box types into the agent's terminal when the agent is ready; agent wakes can too. Rollback: 1.1.96 (reinstall), or per feature by config (below).
Added
The message box sends through the agent's terminal (GTC M3). A message you queue for an agent is typed into its terminal once the agent's own hooks say it is idle (never a guess from the screen), with nobody typing and no picker or draft in the way. It is confirmed by the agent's own prompt hook. If the text does not reach the prompt row, nothing is sent and the item is dropped with a note; Send-now skips only the idle wait. Config messageBoxPath: 'owner' keeps the old path for this release.
Agent wakes through hook readiness and a confirmed typed send (GTC M4), opt-in. With config wakePath: 'term', inbox wakes and boot prompts are typed only when main's hook state says the agent is idle, and count as delivered only on the provider's own confirm (Claude's prompt hook, Codex's sentinel, Antigravity's PreInvocation); nothing is retried blindly. stopContinue per provider lets a Stop hook carry waiting mail into the same turn. The screen check only shadows these sends and decides nothing. The default stays wakePath: 'owner', the 1.1.96 path.
Release cuts gate on identity and on every mutant runner. Each cut now runs scripts/verify-packaged-identity.cjs on the built app and requires every test/tools/*-mutants.cjs runner to kill all its mutants at the cut commit.
Changed
Test runs start without a load burst (LOAD-BURST). The test runner lets at most a quarter of the cores' worth of test files load at once and records each file's start. The first minute's median load time halves (2.4 s to 1.2 s; p95 12.4 s to 5.8 s) at the same wall time. Three fixed sleeps in the test rigs became waits on the real event, with failures that name what they waited for.
A loop over a test glob is a heavy job (HGL196).for f in test/*.test.cjs; do node --test "$f"; done, a computed file list or many repeated runs counts as a suite and takes a slot. Writing the file names out keeps a short loop light.
Fixed
Installers no longer carry native build leftovers (BLU). 91 node-pty build files (MSBuild projects, linker intermediates and logs) shipped in 1.1.95 and 1.1.96, 30 of them holding the build machine's user folder. Only node-pty's 5 runtime files ship now.
An Antigravity Stop in the middle of a tool chain keeps the tab working. The hook stream to the window dropped fully_idle, so the tab went idle while the agent was still running.
Four guarantees GTC M4 had broken are back. A blocked Stop runs the compact-health watcher and the session observers again (autocompact pilot). Idle still means idle, quiet mail stays quiet, and every boot-submit outcome keeps its row, now pinned on the new lines.
Mutant runners that had gone inert kill again. inbox-wake (5 mutants inert since about 2026-09-24) is 36/36, the Antigravity runner is 32/32 again, and the capacity runner (6 mutants inert since before 1.1.95) kills all 198. Each re-anchored mutant breaks the same behaviour on its current line.
Heavy jobs are counted where they really run, and god's screen reading survives a busy start. Rollback: 1.1.95 (reinstall).
Added
Test processes ask for a heavy-job slot themselves. Every agent's Node processes load a small preload that, in a test process, checks with the app before any test runs. A test run the agent's own command shows (node --test of named files) stays light. A run hidden inside a script, an alias or a computed command counts as a heavy job: it takes a slot, shares its agent's slot, or waits in the queue. If the app cannot be reached, the run goes ahead and one log row says so.
Antigravity usage choices on the Floor tab: 3P 5H, 3P Weekly, Gemini 5H and Gemini Weekly beside the budget. The line names the family, and clicking it opens that family's detail.
Changed
The hidden screen copy survives a busy app start. On 1.1.95 it switched itself off about 6 seconds after start, so god's term read found no screen. Its cost is now measured, not estimated, and judged only over a full window, never in the first 30 seconds or 20 seconds after an agent starts. If it does switch off, it comes back after 1, 2, 4 and up to 30 minutes.
The screen check can read the hidden copy. The setting screenGateSource (default renderer) can be set to mirror. If the copy cannot answer, the check falls back to the visible terminal and logs why.
Commands that start with a shell keyword are classified (if, then, do, { and similar), so a heavy command inside them is counted.
Memory notes answer faster under disk load. The per-line disk flush runs off the request path. A note is still acknowledged only after it is on disk.
Fixed
File requests from the window are limited to agent folders. The app's file list, read and write requests refuse any folder that is not an agent folder it knows.
npm ci builds inside agent shells. node-pty's winpty build called two batch files by a bare name, which fails when NoDefaultCurrentDirectoryInExePath is set (as in Claude Code's shells). Both calls are now qualified, and the install checks it.
A release has one path: the verified local build. The GitHub Release workflow no longer publishes on a tag push. It runs only by hand, as a dry build.
God can read and drive any agent's terminal. Rollback: 1.1.94 (reinstall).
Added
term: god's direct channel to agent terminals. The coordinating agent (god) can read an agent's live screen (term read), type text (term type), type and press Enter (term send) and press named keys (term keys). Only god can use it. It is checked on every request, and the command and its address are given to god's terminal alone. Each action is logged with its text clipped and the screen before and after. Text is sanitised before it is written: escape sequences are stripped, and other control characters are refused. Text that could open a menu instead of being typed (starting with ?, ! or /) is refused unless god passes --raw. A Codex send waits longer before its Enter for long text, so the text arrives first.
A hidden copy of every terminal's screen, kept in the app as the terminal's output arrives, so god's reading does not depend on an open tab. It never answers the program running in the terminal. If keeping it costs more than 2% of the app's time, it switches itself off and logs that it did.
Changed
God's typing counts as a person's. An automatic wake is never typed over god's text. A wake held because of god's keys counts as delivered once god's send starts a turn, and is sent again once god clears the line. A wake held because of a person's keys also counts as delivered once the agent reports a prompt submitted after god's send. That link is inferred from timing, not proven, and the log row says so (god-send). A /model switch god types sticks, as one typed by a person does.
More log rows for diagnosis: the terminal modes 5 and 30 seconds after each start, a comparison of the hidden screen copy with the visible terminal's reading, and the copy's cost every 10 minutes. They stay in the local log.jsonl.