Release notes

Added

• A World view: the machine's work as territory. Every project with an agent working in it becomes a plot of land with a figure standing on it, swinging while the agent runs tools and going still when it stops; the rest of the library sits around it as unlit terrain, and a dormant project rises into a plot the moment something starts work there. GSO-1 could only ever see the Claude sessions it had started itself, so six agents working in your repositories were invisible to it; it now finds all of them from what Claude Code already writes down, and shows what each has its hands on.

Fixed

• On an iPhone home screen, GSO-1 was a screenshot of itself. The icon was offered to iOS as an SVG, which iOS does not accept for a home-screen icon, so "Add to Home Screen" fell back to a picture of the page. It is a PNG now, drawn full-bleed with square corners because iOS applies its own rounded mask, and the web manifest carries PNGs alongside the SVG for everywhere else.

Added

• Start a new project without leaving GSO-1. The Library has a "New folder" button: name it, choose which of your folders it goes in when there is more than one, and it is created and opened in VSCode in one step, as a git repository on main unless you say otherwise. Creating somewhere GSO-1 does not watch was never the point, so it only offers the folders you have already told it about.

Added

• The Ops Room can be taught new things without writing code. A skill is a folder with a SKILL.md: a line saying what it is for, the binaries it needs, and instructions written for the model. Three ship with it, for GitHub, Docker and dependencies, and your own go in data/skills/ where an update will not touch them. A skill whose binaries are missing is never offered as usable, so the agent can say it would need gh installed rather than running a command that is not there.
• The Ops Room shows its plan while it works. A job of several steps used to be a blank panel and the word "thinking" for a minute, which is indistinguishable from a hang. The agent now writes down what it intends to do and marks each step off as it goes, on the desktop and the phone. One checklist, revised in place, not a stream of announcements.
• The Ops Room remembers the conversation. It answered every question as a stranger: "which repos are dirty?" then "commit the first one" were two unrelated requests. It now carries the exchange forward, on the desktop and the phone, which share one conversation so you can ask at the desk and follow up from the sofa. It survives closing the window, and Clear genuinely forgets rather than only blanking the screen. Old turns are dropped once the transcript outgrows what a local model can hold, newest kept.
• The phone can see what is open in VSCode. It could already open a repo in the editor on your Mac but had no way to show which were already open, which the desktop has had all along. Repos now filters between what needs you and what is open in the editor, marks the ones that are both, and lists folders open in VSCode that are not among your apps, which a tap brings to the front.
• The phone gets the rest of GSO-1. A More tab leads to the Local LLM, where you can see the server, load a model, download one and run the whole library's summaries with progress; to which model answers in the Ops Room; and to the scheduled jobs, which can be switched on and off or run on the spot. The Planner can now create boards and tasks, rename, link, move and delete them, rather than showing what the desktop made.
• The phone can finish a piece of work too. A repo on mobile now shows what it is, from the same summary the desktop writes, with the measured facts beside it; its branch, which taps to switch, including branches that only exist on the remote; the files that would be committed; and Commit, Commit & push, Push and Pull. Its logs and run configuration are there as well, so "that lives in the desktop app" is no longer the answer.

Changed

• GSO-1 stopped being a background process that costs you battery. With the dashboard open it spawned 28 helper processes every 100 seconds; it now spawns 2. Most of that was top, run from scratch every 2.5 seconds to produce one reading: 673 ms of the 1774 ms is start-up, paid over and over, 42% of wall-clock spent measuring the machine for a number a human glances at. One top is now kept open and read frame by frame, at 1.4%.
• Asking whether the model server is up cost a quarter-second every time. Finding out which process holds the port ran lsof, which walks every open file descriptor on the machine, 267 ms a call, on every status check from the desktop, the phone and the Ops Room, several times a minute. It is now asked the cheap way first, a connection attempt on the port, so a machine with no model running, which is most machines most of the time, never pays for lsof at all; when something is listening the answer is cached until it stops.
• Closing the dashboard now actually stops the sampling. Opening it once started a loop that ran for the life of the process, whether or not anybody was looking. Nothing samples unless something is asking, and a window you cannot see stops asking, so a backgrounded or hidden GSO-1 no longer holds a top open on the other end. Coming back to the window refreshes it at once rather than showing you what was true when you left.
• The scheduler sleeps until there is something to do. It woke every 30 seconds around the clock, 2,880 times a day, to notice that two daily reports were still hours away. It now waits until the next job is actually due, 97 wakeups a day, and edits to the schedule wake it immediately rather than waiting out the interval. Jobs still run at the minute they are set for.

Fixed

• A long Ops Room conversation would have started failing every turn. A tool call and its result are two separate messages, and trimming the transcript to fit the model's context could cut between them, leaving a result answering nothing. Providers reject that outright, so the conversation would have worked perfectly until it grew past the budget and then broken, for a reason nothing in the error would connect to trimming. Orphaned results are now dropped with the call they belonged to.
• Release verification was checking a different set of tools than the agent runs. The list was written out by hand in two places, so a tool added to one was invisible to the other: verification reported sixteen tools while the agent had seventeen, and could not have noticed the seventeenth breaking. Both now read the same list.
• ./ops in a terminal refused to run whenever a different model was loaded, comparing what the server was serving against a name hardcoded when the file was written. The local model's name is now discovered from the running server unless you deliberately pin one, and the mismatch warning is kept for the case it was written for: a model swapped underneath somebody who had chosen it.
• Six controls on the phone's newer screens did nothing. Stop the server, the summary run and its cancel, the download cancel, and both Claude controls all fell through a guard written for the repository screen, which returns early when no repo is open. Starting a model from the phone and then being unable to stop it is how this was found.
• Asking the Ops Room a question with no model loaded was a dead end. It answered "llama-server is not reachable, start it from the Local LLM view", which is an instruction to go elsewhere, do four things and come back, for a model the app already knew how to start. It now offers to start the one you used last, waits for it, and then asks the question you were holding. When the real obstacle is something else, no llama.cpp, no model on disk, or Claude selected without an account, it says which rather than naming a port.
• The Ops Room header showed a green light for Claude even with no account connected. Choosing Claude is not the same as being able to reach it: the status said "running" and named a model while every question was refused for a missing credential. It now says so, and reading whether a credential exists no longer spawns the sidecar, which took the Settings screen from about forty-five seconds to half a second.
• Six buttons did nothing in the desktop app. New board, Rename board and Link app in the Planner, and New file and Publish in the Site tab, all asked for a line of text with window.prompt, which Electron does not implement and has said it never will. It returns undefined, so each of these read that as "cancelled" and returned without a word. They worked when the same page was opened in a browser, which is what made it look like the feature was simply missing. GSO-1 now asks in its own dialog, which works in both.

Fixed

• The Windows build served an error page instead of the app. Every text file was read with the platform's default encoding, which is UTF-8 on macOS and Linux but cp1252 on Windows, so index.html failed to decode and every request returned 500. Everything the app reads, your READMEs, manifests, settings and logs, is now read as UTF-8 explicitly, which also stops non-English filenames and content being mangled on Windows.

Added

• The Ops Room can run on Claude. Its model is now a choice between the local llama.cpp server and Anthropic's models, made in Settings and stored with the rest of your configuration. Everything else is unchanged on purpose: the same tools, the same sandbox, the same build-verify-promote sequence. The guarantee that matters, that the agent cannot touch the supervisor which rolls back its mistakes, comes from the path policy rather than from which model is answering, so it holds either way.
• Sign in from the app. Connecting Claude runs the OAuth flow from Settings, opens the authorisation page for you and stores the token in GSO-1's data folder, owner-readable only. Tokens refresh under a lock, so two Ops Room runs waking at once cannot race each other's rotation.
• GSO-1 can tell you what a repo is. Half a typical library detects as unknown, with no type, no language and often no README: the folder name is everything you get. Open a repo and GSO-1 reads it, asks your local model what it is for, and shows a headline, what it does, its stack, where to start and anything worth knowing before you open it. Summaries follow a fixed schema, constrained at sampling time, so they cannot drift into whatever shape the model felt like.
• Facts are measured, not generated. Tests, CI, lockfile, licence, contributors, file count and last-commit age are counted by GSO-1 and shown beneath the model's text. They are correct even when no model has ever run, and the card renders them either way rather than waiting on one.
• A model recommended from your own memory. GSO-1 works out which model suits the machine it is on, prefers one already downloaded when the size is comparable, warns when the machine is too busy to load it right now, and estimates a full pass from speed measured locally rather than a number carried over from a developer's laptop.
• Summarise every repo in one run, serially, resumably and cancellably, most-recently-touched first so the repos you are working on are described in the first minute. Summaries cache per commit, so it is paid once per repo per change, and a commit that touches nothing the model was shown re-stamps rather than regenerates.
• Finish a piece of work without leaving GSO-1. A repo's drawer now shows the branch it is on, exactly which files are uncommitted, and how many commits are unpushed or unpulled, with Commit, Commit & push, Push and Pull beside them. Git's own output is printed verbatim when something refuses, because that text is the reason.
• The branches of a repo, in the drawer. Local and remote-only branches, newest first, with the current one pinned to the top. Picking a remote-only one creates the local tracking branch for it.
• Unpushed and unpulled counts in the Library list, next to the branch, so a repo that needs attention says so without being opened.

Fixed

• System load disagreed with Activity Monitor, and the app was the one that was wrong. Two separate causes. Memory was divided by 1,000,000,000 while macOS labels binary units as "GB", so the same 34,359,738,368 bytes read as "34.4 GB" here and "32.00 GB" there. And "used" came from top, which counts the file cache as used, where Activity Monitor lists cached files separately because the kernel reclaims them the instant anything needs the space. GSO-1 now reports app, wired, compressed, cached and available the way Apple does, and shows swap, which is the honest sign a machine is over-committed.
• The model recommendation warned about the wrong number. It read free memory, which on a healthy Mac is near zero by design, and told you 0.2 GB was left when 8.8 GB was genuinely available. It reads available memory now.
• An update could arrive half-applied: new markup wearing the previous release's stylesheet. Static assets were served with an etag but no Cache-Control, so browsers cached them heuristically and reused a stylesheet for hours without revalidating. Asset URLs now carry the file's own modification time, so an unchanged file still caches hard while a changed one is a URL that cannot be served from cache.
• Settings showed a permanently disabled Save button on tabs that have nothing to save, which reads like something is broken. Only the tabs made of environment fields show it; Project folders and Ops Room apply immediately.
• Committing from GSO-1 failed on any machine that signs commits. The app is normally started by a LaunchAgent, whose PATH is four directories long, so git could not find gpg and every commit died with "cannot run gpg" while the identical commit worked in a terminal. Git now runs with a login-shell PATH, and with GIT_TERMINAL_PROMPT=0 so a credential prompt fails readably instead of hanging on a terminal that is not there.
• The first changed file in every repo had its name truncated by one character. git status --porcelain marks an unstaged change with a leading space, and the output was being stripped before parsing, so M .DS_Store became M .DS_Store and the path read as DS_Store. This also fed the Site tab's list of changed files.

Added

• GSO-1 can get you a local model from nothing. With llama.cpp missing, the Local LLM tab now shows the install command for your package manager and, on a machine with Homebrew, offers to run it and streams the output. With no .gguf files, it offers a short list of tool-capable models sized against your actual RAM, resolved against Hugging Face so the sizes are real, and downloads one into the folder it already scans. Downloads resume after an interruption. Both are explicit button presses behind a confirmation; nothing installs or downloads because you opened a tab.
• The Local LLM tab remembers how each model was started. Context window, reasoning format, sampling, cache type: every flag is saved per model and restored when you pick it, so loading a model is one click instead of fifteen fields. Start again repeats the last launch verbatim.
• A "Where things live" section on the Local LLM tab for the llama-server binary and the folders scanned for .gguf files. Changes apply to the next scan with no restart.
• Settings names the file that pinned a value. A setting locked by an exported variable now reads "set in ~/run_manager.sh" rather than "set by your environment", because the old wording told you to unset something without telling you where it was set.

Fixed

• The Ops Room header claimed a model was connected when nothing was running. It read GLM-4.7 · local from a hardcoded string, and both status dots were painted green unconditionally. It now reports the model that is actually loaded, or says so when there is none.
• The Library search box was partly covered by the filter chips. The chip row bled upward 22px into a 16px gap, and painted over the bottom of the input. Search, chips and the column header are now one flush sticky stack, so the search box also stays put instead of scrolling away in a long list.
• The Site tab said "not found at ." when nothing was configured. Unset and set-but-missing are now different messages, and both link to the setting instead of asking you to restart.
• Model folders you never chose were listed as if you had. The built-in guesses are labelled as guesses, and folders that do not exist say so.

Changed

• The Ops Room agent follows the running server. OPSROOM_LLAMA_URL and OPSROOM_MODEL default to the endpoint and model GSO-1 has loaded, instead of being two fields to keep in sync with the Local LLM tab by hand. Export either one to point the agent somewhere else.

Removed

• The Settings → Local LLM tab. Everything in it moved to the Local LLM tab, which now owns the model end to end.
• The Chat tab. It duplicated the Ops Room, which is docked on every view, talks to llama-server directly rather than through the Anthropic shim, and already knows your project roots. Existing chats.json is left on disk.
• The Archived nav item, a placeholder for a concept that does not exist.

Changed

• First run always asks which folder holds your projects. It previously adopted ~/Projects without asking whenever that folder happened to exist, which meant a new install could open on a list the user never chose and had no obvious way to change. The picker now opens on a suggested folder, so the common case is still one click, but it is a suggestion rather than a decision. Only an environment override or a saved choice counts as an answer.

GSO-1 can now be set up entirely from the application. Nothing in this release requires a terminal.

Added

• Settings, with five tabs: project folders, Telegram, Local LLM, phone access, and the rest. Values are saved to settings.json and injected into the environment at startup, which is the only way a GUI can configure things that are read from os.environ at import time. A real environment variable still wins and is shown as read-only rather than accepting an edit that would never take effect.
• Add and remove project folders from the app. The Library could previously only ever show the folders chosen at first run; adding another meant setting MANAGER_PROJECTS_DIRS and restarting, which a packaged install has no obvious way to do.
• A Telegram test button. Configuring a bot is four steps across two apps and the usual failure is silent: a good token with the wrong chat id, and nothing ever arrives.
• Restart from the app, since saved settings are read at startup. The restart re-execs in place and keeps the pid, so a supervisor or LaunchAgent watching the process does not see it die.

Security

• settings.json is written 0600. It holds the bot token, the phone access token and any API keys set in Settings, and the mode is applied before the rename so there is no instant where a complete file of secrets is world-readable.
• Secrets are write-only from the browser: the server reports whether one is set, never what it is.

Documentation

• The macOS install instructions described a route Apple removed in macOS 15. Right-click and Open no longer works; the dialog offers only Done and Move to Bin. Every instruction now describes Privacy & Security → Open Anyway.
• A page for the Ops Room: all ten agent tools, the sandbox model, and how to connect a local model.
• The git and GitHub answer, which is that there is nothing to connect: GSO-1 uses the git already on your machine and stores no credentials of its own.

Both of these were found by downloading the v0.1.0 installer and running it, which is the only test that would have caught either.

Fixed

• macOS builds reported themselves as damaged. electron-builder copies the frozen backend in after Electron's own code signature is applied, which invalidates it, and with no signing identity nothing put it back. An invalid signature is worse than no signature: macOS refuses the app outright rather than offering the usual unidentified-developer prompt, and right-clicking to Open cannot get past it. Builds are now ad-hoc signed after packing, and the hook verifies its own work and fails the build rather than shipping a signature that does not validate.
• Application state was written into Electron's browser profile. The packaged app pointed MANAGER_DATA_DIR at Electron's userData directory, so the registry, event log and job state sat interleaved with Chromium's cookies, caches and lock files. State now lives in a data subdirectory, where clearing a cache cannot take the registry with it.

Changed

• The install instructions describe what actually happens on macOS: right-click and Open rather than double-click, and what "damaged" means if it appears.

The first public release: a dashboard that finds every app in your projects folder and runs it, on your machine, not just on the one it was written on.

Packaging & distribution

• Installers for macOS (.dmg, Apple Silicon and Intel), Windows (NSIS), and Linux (AppImage, .deb). The Python backend is frozen with PyInstaller and bundled, so no Python is required to run GSO-1.
• First-run screen: pick the folder holding your projects, browse or type a path, add several roots. No config file, no terminal.
• State is stored per platform (Application Support, %APPDATA%, XDG_DATA_HOME) when packaged, and beside the code in a source checkout.
• Apache-2.0 license with NOTICE, plus CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, and an .env.example documenting every configuration variable.

Security

• Remote-auth public paths are matched on a path boundary rather than as a bare string prefix. Previously a route named /models or /metrics would have been reachable from the network without a token; no such route existed, but the gate was one route name away from opening.

Fixed, portability

• The dashboard abbreviated paths against a hardcoded home directory; the server now reports the real one via /api/apps.
• The greeting, sidebar and avatars were hardcoded to one name; they now come from the OS account record.
• The Jekyll CMS defaulted to one specific personal repo; MANAGER_SITE_DIR now has no default and the CMS stays off until it names a real checkout.
• Ops Room's sandbox root, the only directory the agent may write to, defaulted to a fixed ~/Projects/… path; it is now resolved from the sidecar's own install location.
• Ops Room's readable roots named a specific work folder; they now follow MANAGER_PROJECTS_DIRS.

Registry & launcher

• Scans every top-level directory across one or more project roots.
• Start / stop any app in its own process group, with logs streaming to data/logs/<app>.log.
• Health tracking, probes the app's port and reports healthy, starting, crashed, or stopped.
• Port-conflict detection before launch.
• Auto-detected run and setup commands for explicit run.sh / Makefile, Node, Django, Streamlit, FastAPI, Flask, plain Python, Rust, Go, and static sites. Per-app overrides live in data/registry.json and are editable from the UI.
• Git panel per app: branch, dirty file count, ahead/behind, last commit, and remote, plus git pull --ff-only from the dashboard.
• Descriptions pulled from the README's first paragraph or package.json.
• Full UI rebuild against the redesign handoff: design tokens, light and dark themes, a sidebar shell, the Library view with a status strip, a Details dock, and per-view chrome.
• Ops Room as a dock column with its own overview endpoint.
• Works offline, keyboard-accessible, and honest about failure states.
• Ops Room: a pi-based agent sidecar with an HTTP bridge.
• Claude bridge with an MCP permission server, read-only tools run automatically, writes and commands require explicit Allow/Deny.
• Anthropic→llama proxy (/v1/messages) so agent sessions can run entirely on a local llama-server.
• Recurring scheduled jobs.
• Telegram approvals and an activity monitor with kanban boards.
• Supervisor with versioned release snapshots, a health gate, and rollback.
• Electron desktop shell that owns the window and child lifecycle, and leaves no orphaned processes on quit.
• iPhone companion at /m, gated behind a shared token for remote access.
• launchd integration for start-on-login and crash restart.