Missions wait out a provider limit instead of failing
A job that runs out of every provider account saves its work, waits for the reset and carries on, up to six times. Missions keep the PRD, and the Operation room says what it waits for and until when.
A Virtual Fleet now has a level its missions inherit. Own branch merges each finished PRD into a mission branch a person lands on main; Works alone also lets the fleet's committee approve PRDs and answer questions, sending a person what it must.
Jobs sent with an API key run as that key's own GitHub App
When an API key holds its own GitHub App, every job it sends gets its GitHub access from that App, and its bot signs the work. A key with no App still runs as the Omni App. A send is refused when the key's App is not installed on a repository the job would reach.
A mission's pull request is ready only when five checks pass
A mission's pull request goes ready for review only when five checks pass on its latest commit: CI green, full tests and lint run, every acceptance criterion proven, no test loosened, the change in its area. The PRD page shows them in a Ready panel.
Typing status now shows where this terminal's work stands in plain words: the step it is at, whether you need to do anything, what comes next, and links to its page and pull request. The overview of every PRD moves to the dashboard command.
A Reviewer reads every mission's work before it goes ready
In a mission, the Reviewer hero checks each finished PRD's pull request and leaves one review: a pass marks it ready, and up to five findings send it back for one fix. After two rounds, a person decides. The Operation room shows each review.
Every hero shows all its work and builds across repos
Each Virtual Hero lists every skill of its role, marked runs alone or needs a person, and a new Designer hero joins the fleet. Runners now build a PRD across several repositories, so a mission whose PRD lives in a plan repository runs with no terminal open.
The before/after page has a third tab that shows each screen a change touches at its real size, today beside after, each with a short reference to quote. An agent that can show you the page first lets you ask for changes before it is committed.
Any member creates a Virtual Fleet of agent heroes from the Fleet section, then gives it a mission: a goal and the PRDs or roadmaps it takes. Once started, the fleet builds the approved PRDs on its runners and stops for approvals, answers and merges.
When a runner is lost, its unpushed work is saved and the job is queued again on its own, up to three attempts. The Runners page shows each job as Lost, resuming, and lists every loss if the third one ends the chain.
A pull request that names another one it must merge after stays a draft until that one is merged and, when its repository says how, until it is served. The loop says what it waits for and checks again on its own.
Personas take six new trades with their own portraits, or any trade typed under Other, and the stance excited now reads enthusiastic. Members and agents add or edit a persona from the terminal, and every persona change is kept in the product's knowledge history.
A workspace owner sets a product's key once, on its Settings tab, starting from a key suggested by the product's name and confirmed before it is saved. Until a key is set, the MCP server card says the key comes first.
A bug fix's proof must name its pull request's last commit, or omni bug refuses it. After a review fix is pushed, the agent runs the failing test again and adds a new Re-proved line, never changing the old one, or puts the pull request back to draft.
Points a bug fix or a review puts off for later are kept in the repository, one file each, and counted in the status. A fix or a feature does not ship until each point it named is kept. Nothing opens an issue or notifies anyone.
Agents find your product's rules in their own words
A product's MCP server now searches its knowledge by meaning, so an agent's plain question reaches the right business law, decision or rule even when it uses other words. Exact ids still come first, and the word search answers whenever meaning search cannot.
A workspace owner can turn on an MCP server per product. Any agent then uses one of the workspace's API keys to search and read the product's business laws, architecture decisions, rules, Legal entries and terms, across all its repositories, as of the last merge.
Merge gates that stay on, whatever a vendor's quota
The outbox, inbox, approval and budget checks now run on Omni's own job queue, so every pull request into a default branch gets them again. A check that cannot run shows as delayed on the pull request, then ends as a failure rather than staying in progress.
When a merge changes a repository's knowledge, the product makes its knowledge pack through its own job list. A job survives a crash, a restart or a missed GitHub event, and a jam opens an issue by itself. Specs can be judged against business laws again.
A repository can give each pull request the loop opens the reference of the ticket its issue came from, in the title and as a link in the body. Nothing changes until a repository turns it on, and an issue without a reference gets the fallback title it sets.
A product can now keep its business rules, each with its reason, in one place that every repository and every agent reads. A spec on a PRD's page, or in a phase-0 pull request, is judged against them, and a spec that breaks one cannot be approved until it changes.
Ask an engineer to approve the plan before the build
A PM can ask one named engineer to approve a PRD's plan before its product approval. That person reviews the coding side with an agent and gives the approval on the PRD's page, and the build waits for both. Approval email is now on by default for everyone.
Each product now keeps one list of its domains in Knowledge. Glossary terms can name domains from it, and folders in the product's repositories that are not on the list are shown and counted, with one action to add them. Nothing is blocked by a folder off the list.
A bug fix's pull request opens as a draft on the failing test, then shows its proof from the best level it reaches: a preview filmed by the repository's own hook, the checks of its runs, or the machine that made the fix. Without one, it stays a draft and says why.
Runner work signs as the Omni bot, never as a person
Every job a runner runs now pushes, commits and opens pull requests as the Omni GitHub App's bot, whoever sent it and whatever account the runner's machine uses. A job that cannot run as the bot fails before it starts, and a runner on an older kit takes no job until updated.
A runner can hold several Claude, Codex and OpenCode accounts in order. When one nears or reaches its usage limit, the job carries on in the same work on the next account, and the Runners page shows each part and when every account is out.
Approve a lesson, and its rule is proposed for you
Approving a rewind finding about a missing rule now proposes that rule in the repository's own knowledge, as a pull request a member merges. The card says where it went: the proposed rule, the rule that already says it, or one question to answer.
A workspace owner picks, on Settings › Playbook, which of the workspace's repositories holds its tech playbook. Until one is chosen, the plan repository stays the default, and one click keeps it.
Your fleet's pull requests land on its GitHub board
A workspace owner can link a fleet to a GitHub organization project on Settings › Fleets. Each pull request the loop opens into the default branch for a member, or for a runner job one of them sent, is added to that board, and the fleet's card says how it went.
A product's Knowledge page now lists each of its repositories with the playbook pages and decision records agents follow there. Each one shows whether it comes from the repository, its plan repository, the Omni settings or the kit default, and who owns it.
A bug fix now lists the other problems it saw and left alone, with the repositories each one touches. An app that sent the job with an API key reads that list when the job ends and can offer to launch a fix on each problem, as a new issue.
One playbook for how agents work, in every repository
A workspace's settings repository now holds a tech playbook every repository reads where its own pages are silent, kept offline by sync. Owners see it on Settings › Playbook, and a change there opens a pull request a person merges.
The kit removes the worktrees its agents create once their work is merged and on the remote, and keeps any that hold uncommitted, unpushed, claimed or in-use work. When a session starts, one line says how many can be removed and how.
An Anti-loop button in the top bar opens a space for the tech team's harness, architecture and retro work, in its own colours with the ANTI LOOP logo. Its menu holds Anti-dashboard and Anti-work, neither open yet. Omni loop takes you back.
On the Runners page, a queued job that no runner can ever take reads Cannot run in red, with the reason: no runner holds its repository, the runner or fleet it went to is gone, or none can run it as its key's GitHub App. It reads Queued again once a runner can.
The PRD list loads its rows sooner, a click on a PRD shows its page's outline at once, and the page fills in as each part arrives. Every row, filter and tab shows the same content as before.
For seven days after a merge, each PRD's page says whether it broke main, and the Engineering page counts how many did. When one breaks, a rewind proposes what to fix, each fix approved or rejected in one click.
An app that sends a bug fix with an API key now sees what the job did, reads its lines, answers or dismisses its questions, cancels it and sends it again, without opening the Runners page. It can also be told each time the job changes.
See which level each of your agents' rules comes from
Each playbook section and decision record now names its level: the repository, its plan repository, the workspace's Omni settings, or the kit default. The repository's own wins, and a product rule that contradicts a higher level fails the knowledge check.
Your app's bug fixes arrive under its own GitHub name
An owner can attach a GitHub App to an API key in Settings › API keys. Every bug fix that key sends then runs on GitHub as that app, never as a person's account, and a key without one works as before.
A PRD's page shows how long it took from idea to merged and its longest step, with a Timeline tab listing each step, its duration and who it waited on. The Engineering board shows the median time of each step.
A new Waiting page lists every PRD and pull request that waits on a review, a merge, an approval, an answer or a law's owner, with whose turn it is and for how many working days. The bell now counts the reviews and merges that wait on you.
Pitch titles can highlight one word, every scene can carry a short line under its title, and each step can show its own clip. A new scene counts up figures with their source, the closing scene names the team, and stills show each scene once it has settled.
A runner job that stops halfway no longer reads Done
A runner job whose build ends before its hand-off reads Stopped mid-way in red, on the Runners page and in the terminal, with how many slices it merged and the claims it released. One Resume button sends the same request again.
Before changing a file, an agent now sees the rules and design decisions of that file's area first, then those that apply everywhere, and no longer the laws of other areas. Layers and repository roles use one shared list of words.
When a retro keeps a lesson, a runner writes it into the playbook, or says why it stays out, on the retro's own pull request. Merging that pull request applies each rule, quoted word for word. Guard tests are optional per repository.
A workspace owner can allow a repository's own skills, one by one, in Settings › Repositories. An allowed skill can then be sent to a runner like the loop's own skills, and runs only while it is still allowed and present on the repository's default branch.
The code base names no internal repository, real person's email or secret, and a fresh install starts with no workspace but the demo one. Every push and pull request is checked, and a secrets scan runs on each pull request.
OMNI DUET now gives 1 point for each obstacle passed in a chapter and 2 in endless mode, instead of points per wave; a hit still gives back the points since the chapter began. DUET's board starts over, and earlier bests are kept aside.
Every agent session now starts with the product's words, its legal entries and its personas. The skills read the full glossary when they write, and an avoided word in a spec or a mockup is flagged with the word to use instead.
Before/after pages that show the technical change too
Every new before/after page has two tabs: one for the journey and the screens, one for the flow between services and, per service, the database and API changes shown today beside after. Diagrams are drawn into the page itself, so it reads with no network.
Runner jobs pick up again once your answers are in
A runner job that stopped on questions starts its fix by itself on the same runner once every question has an answer, and reads Answered · fixing. A switch per repository, on by default, turns this off and leaves a Fix with answers button instead.
Statements, never lines, legal entries and glossary terms can be added and edited from the terminal, one at a time, with the same rights as on the Knowledge tab. In a Claude session, each item is shown and saved only after a yes.
Send a build to a runner and follow it from your terminal
A PRD can now be sent to a runner from the terminal and followed there, one milestone at a time, until it ends with a clear next step. The Runners page names the PRD each job builds, and the status band shows how many runner slots are free.
When a bug does not show in tests, a bug fix now checks the built app with a check each repository commits. Without that check it stops and keeps what reproduces the bug, so nothing ships unproven. Its triage appears at once, with the release the bug first showed in.
When a bug fix on a runner stops for a decision, its job reads Need human on the Runners page and the sidebar counts it. Anyone in the workspace answers its questions in a drawer, or dismisses them, and an answer starts the fix again with it.
Everything agents know about a product, in one place
A product's Knowledge tab now holds its Constituents, Personas, Legal, Glossary, Pitch and Settings. Each list sorts by any column and shows who made each item and when, and every item has its own page. Settings no longer carries these sections.
Workspace owners can archive a product: it leaves every member's list, the terminal and agents are told it is archived instead of reading its rules, and nothing is erased. Owners find it under Archived and restore it with its repositories and history.
A PRD's page now counts the answered questions and marks the pull request as a draft, then offers Fix with answers once every question that needs you has a reply. A runner job that stopped on questions reads Waiting on answers, Held or Ready, with a link to the PRD's page.
A concept now reads shipped once every area's PRD has shipped, or when a workspace member marks it done. Its Areas tab shows where each PRD stands, and areas a closed concept never built read not built.
See a product's knowledge whole, across its repositories
Each repository sends its knowledge to the Omni page, by hand or after every merge. A product's page, the terminal and connected agents then show its rules from every repository joined to its Legal and Glossary entries, with who sent each repository's knowledge and when.
The Constellation tab returns on a product's home as a night sky where every repository has its own place and no label covers another. Lines join the repositories that its PRDs' plans target together, with a card per line listing those PRDs.
A workspace owner can remove a repository from Settings, after a confirmation, and add it back later with its history. A removed repository no longer appears in product tabs, pickers, the product home or the ideas board.
When a runner job stops while building, its runner releases the slices it held and says so on the Runners page and on each pull request. The next run reuses that work when it still matches the plan, or rebuilds it, and a runner job never waits on a question.
The knowledge graph command takes one rule, or the files you are about to change, and lists only what touches them: the principle or Legal entry it serves, the test that proves it, the screens and decision records beside it. It works offline and never prints a whole register.
When the player's kart skids or drifts, its own skid marks now show under its rear wheels, over its shadow, on the wide and the tall screen. The rivals' marks are unchanged.
Every bug fix arrives with its before-and-after proof
A bug-fix pull request now carries an animated before-and-after image, drawn with no extra tools, and is refused without one. Runner machines bring their Omni plugin up to date before each job. A bug found already fixed names the change that fixed it only once that is proven.
A product's home has a new Constellation tab: its repositories as stars, the dependencies its owners declared as lines, and each PRD's reach. A star shows its most urgent state, a chip filters the map to one PRD, and a docked star is tethered to its phase-0 pull request.
A law change or an open question that waits on its owner is passed on after two working days to the stand-in the settings name, then to product leadership. Asking never approves anything: the check stays red until a person answers.
The knowledge base takes a fourth kind of entry, the domain contract: a plain line on how an entity is shaped or moves through its states, pointing to the file that enforces it. The knowledge check refuses a contract whose file is gone, and rewording one asks a person first.
A person answers every change to the agents' rulebook
A pull request that changes the briefing, or a page it points to, stays red until a person answers for it. The people who review the laws are now asked to review the briefing too.
One place for what the law requires and your words
Each product page has a Knowledge tab holding legal requirements and a glossary, kept by named owners while other members propose changes. Every repository of the product reads them, a rule can cite a legal requirement, and the word check flags words the glossary avoids.
Every open question in one Outbox, repository PRDs included
The Outbox now lists repository PRDs' open questions beside product PRDs', in one list filtered by product. A repository question shows when it was last read and links to its PRD page, where you answer it as before.
Roadmaps, bug records and knowledge live on your product
Roadmaps and bug records spanning your product's repositories are written and approved on the Omni page, and a bug record closes when its last fix merges. Questions shows every product's open questions in one Outbox, and plan repositories keep working as before.
The game room gains OMNI PONG, open from level 5: hum high to move your paddle up and low to move it down against an AI. The microphone is required and its sound never leaves the browser; the score is the points scored before the third miss.
See each runner machine's health and token use, live
The Runners page shows each runner's CPU, memory, disk and Claude tokens, with the last hour's trend and each job's tokens. A runner under pressure stops taking new jobs until it recovers, and a queued job says why it waits.
Plans name things the way your repository already does
Before a brainstorm or a plan names a table, a column or a schema field, it reads the pages where each repository writes its conventions down. Every spec now lists the names it chose, each with the page it came from.
API keys let your other tools read and start the loop
Workspace owners make API keys on Settings, each optionally held to some repositories. A key lists the workspace's PRDs and bug fixes page by page and starts a bug fix on a runner. Every route is described in the API reference in the docs.
FROST EIGHT's ice now slides, and two of its bends freeze over. SPACE hops into a drift that pays a blue or orange mini-turbo, GAS at the right beat of the countdown gives a rocket start, and a kart tucked behind another is pulled past it. ITEM moves to SHIFT.
Every session opens by saying whether this checkout is ready to plan on: main pulled within a day, Omni on its latest release, the knowledge and the Omni page fresh. One command runs the safe fixes, and the status line shows the state at a glance.
Every bug fix's pull request shows the bug before and after the fix in a short animation, also kept on the bug's page; when none can be made, it opens as a draft that says why. Services can queue a bug fix on the workspace's runners with a member's sign-in.
Merged pull requests pay their author and reviewers
A pull request merged into a tracked repository's default branch pays its author 30 points and each reviewer 10 for a feature or fix, or 10 and 5 for a retro or knowledge change. Pull requests merged into a feature branch no longer earn points, past seasons included.
A plan now advises a tier for each slice, and a team can map each tier to a model and effort for each coding agent. The wave then starts each slice on that model and says on its pull request which one built it; with no mapping, nothing changes.
The decision list stays short as decisions pile up
Agents reading the decision list now see the latest 20 decisions and the next free number, with one command to show every decision. The list stays the same size however many decisions a team records.
When end-to-end tests are written without anyone to ask, the first list of critical flows is drafted and each candidate test comes with a question offering standard or critical. Replying with the letter sets the test's level, without editing its tags by hand.
When a job ends, the runner deletes the job's local branch unless it was pushed. Once at start and every hour, it removes the folders, branches and old logs that finished jobs left behind, and keeps logs for 14 days unless told otherwise.
FROST EIGHT, a second kart circuit across a frozen lake
OMNI KART gains FROST EIGHT, a figure eight on a snowy mountain whose two diagonals cross on a frozen lake where karts slide. A CIRCUIT step picks the circuit, the cabinet shows each circuit's crew times in turn, and COMET RING's times start over.
A repository can ask every plan to declare, slice by slice, which of its contract kinds a slice changes, or none. Plan check refuses a slice that declares nothing or names a kind the repository does not list, and plan repositories read each target's own kinds.
Each product can list its critical flows, and every end-to-end test is now marked critical or standard by the flow it covers. The status report shows each test's level and flow, and refuses a critical test that covers no known flow.
Every screen of the website now speaks of your coding agent rather than Claude, from the question page to the play dock. The home page says it works with Claude Code, Codex, Cursor, Gemini CLI and any coding agent, and shows how each one runs the loop.
Each job and each runner on the Runners page has a Terminal button that opens a live, read-only view of the job's log, one line per step. A job keeps its last 20,000 lines while the page lists it.
Runner jobs sign in to Claude on a headless server
When the Claude sign-in token from setup-token is exported before the runner starts, every job it runs gets that token and signs in to Claude with it. Without the token, jobs keep using Claude's stored sign-in, as before.
Install the Omni Loop for Codex, Cursor and Gemini CLI
A repository now lists the coding agents its team uses, and the install puts every Omni skill where Codex, Cursor and Gemini CLI find them, kept in step with each update. Help shows each person how to run a skill in their own agent.
The skills say each step in plain words any coding agent can follow, with the Claude Code way shown in brackets. A new section of the rules every session reads first tells Codex, Cursor or Gemini CLI how to ask, wait and work in parallel.
COMET RING drives in curves, with a vaulted tunnel
OMNI KART's COMET RING turns through a curve at every corner, from wide bends to tight hairpins. Its tunnel sits in a bend under a vaulted ceiling, and the walkway shortcut is one tile wide, with a gap only top speed clears. Best times are kept.
Send work from the app to a server and watch it run
A server running omni runner shows on the new Runners page, and an owner can give it to a fleet. Send to terminal, beside Build it and the other copy buttons, runs that line there. Each job's log, status and pull request come back to the page.
Codex, Cursor or any coding agent, on a laptop or in a cloud session, can now start an Omni session, ask its questions on your ask page and show it is working, with no Claude hooks. Claude Code works as before.
The game loads its map first, details when you open them
The game now loads only what its map needs. A planet's dossier and log load when the planet is opened, and the star chart when it is opened, each kept for the rest of the visit.
Every commit and push now runs the same checks in git itself, whichever coding agent or person makes it, and they refuse when a tool is missing. Draft pull requests and pull requests into a feature branch are checked by CI on every push.
Builds across repositories install before they test
A repository the loop builds in can name its install command in its own config. Builds across several repositories run it before that repository's tests, so the tests run locally on real dependencies. Setting up a repository proposes the command from its lockfile.
COMET RING gains a station tunnel, two jumps, boost pads and two shortcuts, with fleet stands, new props and a busier sky beside the track. A kart now tips over only after half a second with both wheels of one side over the void.
Locked laws, the few rules only their owner can change
A repository can now lock up to twenty laws, listed in a plain-words thesis. A pull request that changes one stays red until its code owner approves it. Agents read only the laws for the files they touch, and invade drafts a thesis everywhere.
The kit reads AGENTS.md before CLAUDE.md, links a dossier opened under Codex, Cursor or any agent to its checkout without setup, and points flow hooks to .agents/hooks/. Nothing changes for Claude Code.
The arcade and the GitHub App now say when they break: one JSON log line per webhook, Inngest run and sync, credentials masked; a health endpoint on each; and an hourly check that opens an issue while an app is down. The guide lists the Vercel alerts to turn on.
Context budgets, a size limit on what every agent reads
A repository can now cap what each agent reads before it starts. Invade sets the limits at today's sizes, each pull request gets a budgets check saying in plain words what grew and how to trim it, and a limit only goes up when a person agrees.
OMNI KART gains an OPTIONS screen where players set their keys, the laps, the rivals' difficulty and their kart's colour and design, kept in their browser. The arrow keys and space drive by default, and every finished race shows a podium before the results.
Secret scanning and push protection, switched on and kept on
This public repository now refuses any push holding a real credential: secret scanning and push protection are on. Test fakes are built at runtime and a guard refuses literal ones. A daily workflow opens an issue if either setting goes off.
OMNI KART's circuit climbs, dives and tips you over
COMET RING now rises and falls, slowing karts uphill and speeding them downhill, and a kart with both wheels of one side over the edge tips into the void. Item boxes stand on every straight, a minimap shows every kart, and best times start over.
Signing a terminal in and signing up now open on a page of their own, with no menus: the Omni Loop crest over a starfield and one card. The terminal's code is shown in the arcade's pixel letters, and sign-up says what matters in a few short lines.
The loop reads its state from the Omni page, not from gh
The loop's commands now read pull requests, reviews, comments and issues through your Omni sign-in instead of the gh tool, so any coding agent with the kit can drive the loop. A read that cannot be made prints one line and the loop waits.
Every coding agent, Claude Code, Codex, Cursor or Copilot, now reads a root AGENTS.md first: the repository map, the rules, every command with its time, and the pnpm 9 it needs. Claude reads it through a one-line CLAUDE.md.
The e2e validation no longer stops on a test made only of exact steps, because only a test that calls the agent must have a recording. The guide now starts from the framework's own login, and the skill writes that login setup when a repository has none.
In a cloud session, over SSH or in a container, the omni command now prints a link and a short code instead of waiting on this computer's browser. Approve it in any browser and the terminal signs in, or makes a personal token. On a laptop, the link shows as a fallback.
See on every pull request that its PRD is the one approved
A PRD approved on its page now gets the approved check on its feature pull request and its target pull requests, failing when an approved file changed. A single command writes a changed file back, and approving or voiding reruns the check everywhere.
A product's PRDs live on the server, with no plan repository
A PRD across your product's repositories is now written, approved and answered on the Omni page, with no plan repository or plan PR. Each target pull request shows whether it builds what was approved, and members answer the agents' questions on the product's Questions tab.
Each merge now deploys the arcade and the GitHub App at most once, and only the one it changes; a merge of documents deploys nothing. Previews build for ready pull requests once their pushes settle, not on every push. A failed release still deploys.
Claude Code on the web or the phone can now reach the Omni page. Make a personal token from the terminal or your profile, paste it once into your cloud environment, and every session signs in as you. Tokens expire and stop when revoked.
The game room gains OMNI DUET, open from level 4: turn two orbs around the crest to dodge falling blocks through four chapters, then an endless mode worth double. A hit restarts the chapter at its starting score, and the score is saved once when the game ends.
Shared screens, and only what you touch can hold you
Your app shell, notifications and sign-in become screens of their own, drawn and checked like any page. A bug you see in a part you did not change is reported apart and never holds your work. Every PRD's before/after now shows what it does, then how it looks.
OMNI KART's ready screen offers SINGLE, today's race against the CPU karts, or MULTI, where up to four players of the same workspace race COMET RING live, each from their own browser. BLOBs and ORBs hit the other players, and a MULTI time counts toward the cabinet's best time.
Every screen gets a mockup; locks guard only its look
Your agents draw a mockup for each screen in your library, and one page shows them all. A locked screen asks you first only when its look changes, not the code behind it. Word checks read only what is on a screen, and test-only changes are no longer screen work.
Lock the screens you are happy with and your agents build them as written, never redrawn; changing one asks you first. Every mockup has its words read before you see it, and each change names the screens it touches. Your agents remember your product's design from run to run.
Screens that look like your product, checked for you
Switch on design craft and your agents learn your product's own look. Each change to a screen is checked on mobile and desktop against the mockup you approved, and the pull request says what was fixed. Ask /omni:pixel-perfect to polish any screen. It never blocks a delivery.
Products get a home, and a service can serve several
Each product has its own page listing what waits on whom, its PRDs, ideas, roadmap, fixes, questions, repositories and approvers. A repository can belong to several products, and a PRD takes one product or none, chosen when it is written.
COMET RING hangs over a starfield with no walls or verge beside the road. A kart that drives off the edge falls for a second and comes back at rest on the racing line, its lap kept, and rivals follow the same rule. An ORB turns back at the road's edge.
OMNI KART records each player's best race time instead of a point score. The results screen shows the time of the three laps, a new best means a faster race, and the cabinet's crew top five ranks times, fastest first. Earlier kart scores are cleared.
OMNI KART races through space, with music and sound
COMET RING becomes a neon circuit in space, with pylons, asteroids, a moving sky, shadows and sparks. The race plays its own music, an engine that follows your speed and a sound for every item, hit and box, all muted by M. The track and the driving stay the same.
Race your team in OMNI KART, the arcade's third game
The game room's third cabinet opens OMNI KART at level 3: three laps of COMET RING against five karts driven by your workspace's fleets, with BOOST, BLOB and ORB items. Your place and time make a score saved to the arcade.
A decision becomes a rule only when it is worth one, and each new rule gets a test that fails when the rule is broken. When a change touches a rule or its test, a person answers before it ships, on fix pull requests too.
The ask pages and the bell read and write through the app's server instead of the database. Once signed out, they stop checking for news and the ask pages ask you to sign in. Screenshots on an answer go one at a time, up to 4 MB each.
The e2e skill is followed by /omni:yolo when the spec asks
A spec can now ask for the e2e validation with one line in its front matter. When it does, and e2e is turned on, the loop runs it after the pull request is ready and never blocks on its result. Brainstorm asks the question. Off by default.
The e2e beta now works in a plan repository: tests and recordings stay there, each record gets a name unique to the run, the run signs in from a saved session, and a new guard refuses credentials before a commit. Still off by default.
Approvers are asked on their phone, and the build starts
Each product lists who approves its PRDs. They are asked by phone alert and email, the approve screen opens with the problem and the solution, and the build starts once one approves. A later change to an approved file voids the approval and asks again.
A healed e2e step now shows what the screen looked like before and after. Its changed recording waits aside until a person confirms it; rejecting it keeps the old one and the test red. The beta is still off by default.
Approve a PRD on its page, with no phase-0 pull request
A workspace owner can switch a repository so its new PRDs are approved on their PRD page. The page records who approved and when, and building stops if an approved file changes. Repositories not switched work as before.
The e2e beta skill now sets up its own e2e project, checks your package manager matches your CI's, and can film against a demo address when the preview has no data. It is still off by default, so a repository that has not turned it on sees no change.
Work in the menu now lists Concepts. Open one to see its brief, its areas linked to their PRDs, the clickable vision tour, every board and the debate, marked in review or in the inbox.
Keep a PRD's acceptance criteria as e2e tests (beta)
A new beta skill turns the acceptance criteria of a ready PRD into e2e tests that stay in the repository, with their recordings, in a sub-PR to review. It is off by default, so a repository that does not turn it on sees no change.
Members of a workspace vote on the ideas of their own board while it is still private, and take a vote back the same way. Those votes stay counted once the board goes public. People outside the workspace still cannot see or vote on a private board.
A repository can publish an ideas board that anyone can read, in Now, Next and Later lanes. People who sign in with GitHub vote once per idea. Members add and arrange ideas on the page or from the terminal, and turn the board on in Settings.
Know what a roadmap needs before the loop gets stuck
A roadmap lists what must be ready first, such as Docker or access to a private package. The agent checks each item and fixes the safe ones. The rest wait on a Prerequisites tab with the command to copy, what it does and who can do it. Only the PRDs they block wait.
See what your Claude session is building, one click away
Claude Code now shows what the session works on, whether a PRD, a bug fix, a visual fix or a roadmap, with its stage, the slices being built and links to its page and pull requests. The band hides when there is nothing to show.
A roadmap's page lists the work only a person can do, sorted as business, development, dev ops or delivery ops, each with what to do and where to answer it. Settled work stays listed, so a finished roadmap shows the human work it took.
When you drive several PRDs, the loop now keeps up to three steps running at once, each from a different PRD, and starts one only when it shares no files or open blocker with those running. The Loop page shows what runs where, and why any step waits.
A phased plan becomes a roadmap of PRDs in one sitting, with one review. The loop then builds every PRD that is not blocked, across repositories, and says which pull request each waiting one needs merged. A Roadmaps page shows each milestone as a Gantt.
Answering the agent's questions now earns points: 2 points and XP per answer on a PRD, on the Omni page or in the Claude terminal, counted back to when the game began. Terminal answers are no longer lost, and points always reach a real GitHub account.
After a feature is merged, each rule and invariant written back to the knowledge base names the test from that feature that proves it, once a person confirms it. The status screen shows how many rules are proven by a test.
The People table ranks by points and sorts by any column
Every board's People table now lists people by points, highest first, with a Rank column that shares places on ties. Click any column header to sort by it, and click again to reverse. The sort stays in the page's address, so a sorted table can be shared.
Drive the loop across your PRDs, one step at a time
One command now plans every PRD you own together and builds them step by step, in order, until each one waits on a person. A new Loop page shows every running loop, its plan, and each step linked to its PRD.
List the files your repository builds in the config, and the loop stops treating them as conflicts: no fake territory breaches, no extra waves, no merge clash on a bundle. Each wave rebuilds them once, and a wrong list turns omni check config red.
When a feature built across several repositories merges, its retro now reads every one of them and groups what it found by repository. Each repository's merged pull request gets a comment with its own findings and the retro's link.
/omni:mega-pr-care keeps every pull request of a multi-repository PRD green, conflict-free and review-handled, and waits when a PR is red only because another repository's is not merged. /omni:mega-bug-fix fixes a bug across repositories, back-end first, one linked PR each.
A pitch is now an animated launch video in your product's own colours and fonts: the feature's title builds on screen, the camera zooms on what matters, with free music and an intro and outro you set in Settings.
A shipped PRD now has a Pitch button: pick customers or the team, and the loop makes a short launch video, a square cut and a GIF in your product's look, kept on the PRD's page.
A repository can now set its own rules for how Omni Loop plans, tests, opens and merges its work, part of the code by part of the code, with no code of its own. A plan that breaks a rule is refused before anything is built.
PRD and fix pages now show a saved copy of what GitHub said, with the time it was read, and refresh only when something changed. Both apps share one GitHub allowance, so background work gives way to people, and if GitHub asks for a pause, everything waits instead of failing.
Each night, Omni Loop now changes its own core code on purpose, thousands of small changes, and checks that its tests catch them. Each part of the core has a score that may never go down, and a fix or a new test can check just the code it touched in a few minutes.
The rules for which part of Omni Loop may use which are now checked on every change. Code that reaches across a boundary, or browser code that reaches server-only code, is stopped before it builds, with the path that led there.
Every part of Omni Loop now checks its settings when it starts. A setting given only half, or in the wrong shape, stops the start with the setting's name, never its value; an optional feature left unset stays off.
PRD, pull request, issue and comment numbers, slice ids and decision ids now each have their own type, checked where they are read. An agent that passes one where another belongs is stopped before the change builds, and a check keeps every new identifier typed.
Agents building this repository now check a change in seconds: one command typechecks, lints and tests only what moved, and audits it against its base. Typechecking runs on the native TypeScript compiler, and pull requests run their tests and lint in parallel shards.
Every row, answer and saved step the loop reads from outside is now checked where it arrives, so bad data fails at once and names the field instead of slipping through. Nothing changes for people using the loop: production's data was checked before shipping.
The last strict compiler checks are on: a missing value and an empty one are no longer confused, and an unused parameter or a forgotten return is an error. Nothing changes for people using the loop.
Every pull request is now linted for the mistakes types cannot see, such as a promise nobody waits for or a value printed as "undefined". The code holds zero findings, with no rule switched off. Nothing changes for people using the loop.
When you are signed in, the home page shows your hero, or your GitHub photo, beside Continue your game, one click back to the game. Visitors who are not signed in see the page as before.
Omni Loop now has its own domain, omni-loop.xyz. Search engines and link previews keep it, and the site has a robots file and a sitemap. Signed pull requests and new installs link to it. The old address keeps working, and the guide says how to switch an installed repository.
The home page shows the loop, your customers and the game
The home page draws the loop as the docs do, shows how your business, products and personas steer the agents, and invites your team to build a fleet and climb the leaderboard, with real counts as proof. The sections about the loop's internals are gone.
Every panel on the Home, Fleet, Workspace and Engineering boards now sits in a card like the tiles. Cards side by side end on one line, and tile figures line up even when a label wraps. On a phone the cards stack, and tables scroll inside their card.
Fewer unchecked shortcuts, and a limit that only goes down
The code now holds 222 places that tell the compiler to stop checking, down from 552, and each one says why. A check counts them per part of the repository and fails if a count rises, so the number can only go down. Nothing changes for people using the loop.
Pick your app when you sign up: the board or the Arcade
Signing up from the home page now opens a select screen: the Omni app for the board, or the Arcade for the game. The pick decides where you land, and Remember my choice skips the screen next time, with a change link under the button to pick again.
The kit, the game and the GitHub App are now strict TypeScript, type-checked on every pull request, and data read from outside is checked against a schema that names the wrong field. Nothing changes for people using the loop: same commands, same output.
Join a repository that already runs the loop in one step
Setting up the loop on a repository that already has it now only installs the plugin and signs you in. It opens no install pull request, and it tells you whether the knowledge base still needs filling or only a refresh.
Product constituents: the rules every agent reads first
Owners give each product a Statement and a Never list on Settings › Business, with a history of who changed what. Every Claude session reads them first, even offline, and the inbox check flags a spec that breaks one, judged by Jev once switched on.
Settings › Business makes a read-only link for Cursor, Claude Code or any MCP client, so your editor's agent reads who you sell to and cites claim ids. A question no claim answers lands on the page to answer once. Jev can set junk questions aside.
Phase-0 PRs now check the spec against your business
Add Never lines on Settings › Business. The inbox check on phase-0 PRs gains a canon gate: a spec that breaks a Never line or misses your customer turns red, quoting the spec, with Rewrite and Change the claim buttons.
Let Jev make three decisions, one switch at a time
Settings › Jev lets an owner add a TypeSafe AI key and hand Jev three decisions: question category, outbox risk and bug risk. Each is Off, Shadow or On, with its own threshold and 30-day record. When Jev can't answer, today's answer counts.
Think-big and brainstorm speak through your product's personas: the one a design misses objects once, with receipts. Each PRD page has a User voice tab showing every persona's score by stage, and Rework with this feedback.
A second game in the game room: SUPER OMNI WORLD, an original SNES-style platformer that opens at LV 2. Run and jump through three stages with your own hero, stomp entropy and reach the flag. The corner Game Boy lets you pick it too.
Run /omni:prove on a ready PRD and Claude films each acceptance criterion on the preview. The PR gets a ✓/✗ comment with a moving preview, and the PRD page a Proof tab with every clip. It's off by default, and it never blocks a PR.
Give your customers faces: personas on the Business page
Settings › Business has Personas: a cast per product, each with a stance, a pixel-art portrait picked from a trade's variations, who they are and how they use the product. Agents read them with omni business show.
Claude keeps your feature PR green and answers its reviews
Run /omni:pr-care on a PRD and Claude watches its feature PR until it merges: it fixes red CI and merge conflicts, fixes review comments worth fixing, and pushes back on nitpicks with a reason. The PRD page shows a health chip and a PR care tab with every thread's verdict.
Omni drafts your business from your repos, with receipts
Draft from my repos in Settings › Business reads your repositories and pasted web pages, proposes who you sell to, each fact quoted, and That's us confirms them. A weekly recheck brings what changed to the bell.
While Claude works on a PRD or a fix, its page and your /ask tab offer Entropy Invaders on a small Game Boy in the corner, under the arcade's usual rules. The game pauses when Claude asks you something, and says when Claude is done.
Settings has a Business tab where members tap who they sell to, where, and against which suggested rivals. Each repository can belong to a product. Think-big reads the confirmed facts and shows which ones it cited.
Readable long questions, and Claude's message beside them
Long questions now show a bold lead and fold the rest, with steps as a list. When Claude asks you to approve something it just wrote, that message shows above the question, so you no longer go back to the terminal.
A PRD delivered through Omni Loop now earns points in every repository tracked in Settings → Repositories, multi-repository PRDs included. Points and XP start again from 0 the day this ships; games you already unlocked stay unlocked.
The menu is shorter: Dashboard, Work, and one Settings entry at the foot that opens Fleets and Repositories as tabs. Shared with me and History are tabs on the Questions pages. On a computer, « folds the menu to a rail of sprites, and your browser remembers it.
New /omni:think-big: a studio of agents draws concepts for an idea too big for one PRD, over rounds of boards and prototypes you react to. You crown one, and /omni:brainstorm --concept turns each of its areas into a PRD.
Engineering board — main-branch merges and loop health
The Engineering board now counts only merges into main, master or develop, so sub-PRs and promotions no longer inflate it. A new Loop health panel lists what the loop has stuck right now, each linked, and how often merged sub-PRs needed a fix first.
Every person and fleet name in the app is now a link. A fleet opens its board, and a person opens their new profile: their numbers and their PRDs, fixes, pull requests and reviews. "My profile" in the avatar menu opens your own.
Bug Fixes and Visual Updates open at once with a placeholder list, then show their rows without waiting on GitHub. Each fix's state and labels are refreshed every 15 minutes, and again whenever someone opens that fix.
See who is who, and their fleet, everywhere in the app
Wherever the app names a person, their face shows first: their arcade hero, else their GitHub photo, else their initial. Wherever it names a fleet, its mascot shows in the fleet's colour. Home's people table is now headed Your fleet.
The PRD list and the dashboards show their frame at once and fill each block in as its data arrives. The sidebar moves between pages without a full reload, and the PRD list no longer waits on GitHub for its open-question counts.
Phase-0 pull requests get an **inbox** check from the Omni Loop app: green when the PRD's spec, plan, before/after, signed commits and open issue are all in place, red naming what is missing. Here, they no longer run the tests or build previews.
On the Engineering board, the top people lists show each person's face, their count in its own column and a bar against the leader. Each repository in the table opens its own page with the same stats for that repository alone.
Bug fixes and visual updates have their own pages under Work
Work now lists Bug Fixes and Visual Updates beside PRDs. Each fix has a page with its record, every round of variations a look was picked from, and a Timeline of who asked, picked, approved and merged it, and the release it went out in.
When a question on the ask page needs more than words, Other now takes up to five screenshots: paste, drop or add them. They stay private to your workspace, and Claude in your terminal opens each one and sees what you see.
Real stages for every PRD, and counts of where they are
Every PRD shows seven stages, synced from the repository every 15 minutes, so a PRD in shipped reads as shipped and no page says the stage is unknown. The PRD list counts and filters by stage, and the dashboard shows how many are open, building or shipped now.
Choose the repositories you track and see how PRs move
Workspace owners add repositories from Settings, and each one is tracked by default. A new Engineering board shows pull requests opened, merged and open, time to merge per repository, and who opens, merges and reviews most. It also shows the share Omni Loop signed.
Tests and a dead-code check on every ready pull request
A ready pull request into main now runs the tests and a Fallow check for any unused or duplicated code it adds. Runs wait three minutes, so a burst of pushes costs one. Claude runs the same check before every commit and push.
A guide page for working across several repositories
The guide has a new page, Several repositories, that shows in three drawings how a feature spread over several repositories is planned in one place, built with one pull request per repository, and reviewed once, and in which order its pull requests are merged.
The app's menu now opens on Dashboard: Home, Fleet and Workspace. Each lists every member, even at 0 points, with PRs merged, PRDs drafted, in progress and shipped per day, repositories involved and questions answered, over 7 days, 30 days or the season. Fleets moved to Settings.
Build one feature across several repositories, one review
A feature planned across repositories can now be built from the repository that holds its plan: each repository gets its own pull request, every decision taken along the way is gathered in one place for a single review, and the plan says in which order to merge them.
The docs now list every Omni Loop skill by what you want to do, with a page each: what it does, when to use it, what to type, who runs it and its related skills. The docs search finds them, and /omni:help shows the same example in the terminal.
An alert when a spec, plan or before/after lands on your PRD
When a spec, a plan or a before/after page lands on a PRD you opened, the bell lists it under New documents until you open that PRD. With alerts on, you get one per PRD once its pushes go quiet for 30 seconds.
A new command, /omni:bug-fix, takes a bug from one line or an existing issue to one pull request: it triages the bug on its issue, shows a test failing before the fix, fixes it, adds a guard, and records what it found.
A fleet's owner now picks its mascot from eleven, not six: Atom Eve, a shark, a turtle, Allen the Alien and a robot join the library, and each plays its own short sound in the arcade.
One plan across repositories, reviewed in one place
A feature that spans the back-end and the front-end can now be designed once, in the repository that holds plans: each step of its plan names the repository its code lands in, and one review covers the whole feature before any code is written.
A new command, /omni:visual-fix, takes a small visual change from one line to one pull request: it shows four or five options, applies the one you pick, and records a before-and-after page.
On a PRD's page, the Questions tab has more breathing room. Each round sits in more space, an unfolded answer reads as separate, taller lines instead of one dense block, and the "answered" count is set a little further from the list. On a phone the list looks as it did.
The knowledge map now has a Repository menu: beside Omni Loop's own, it lists every repository of your workspaces set up with Omni Loop and draws the one you pick on the same map. A link to one of its entries keeps the repository, so it can be shared.
A plan repository that knows its target repositories
A repository that holds plans can now list the repositories its features land in, with a role for each, and see which ones run Omni Loop and whose knowledge is current. One command sets it up in a single pull request.
Answer the questions the agents leave in the terminal when a build ends, or on a PRD's Outbox tab beside its spec, where Send posts your answer on the pull request as you. The PRD list shows how many questions each PRD has open.
Every app page now shows, without a reload, the questions and outbox items waiting for you: a count in the browser tab, counts in the menu, and a bell in the top bar listing each one. Desktop alerts and a chime can be switched on from the bell.
Lighter app pages, and clear questions left to answer
Every app page uses the full desktop width, and the top bar stands out in every theme. On a PRD's Questions tab, answered rounds fold to one line, rounds awaiting an answer stay open, and the tab counts questions: "7/10 · 3 to answer".
Retros and knowledge updates only when they teach something
Omni Loop now opens a retro or knowledge pull request only when it adds a new lesson or rule; otherwise it leaves one short comment on the merged pull request. The knowledge base no longer records how screens look.
The Omni page is open to every workspace, not only Vertuoza
No more @vertuoza.com address needed. Install the Omni App on your GitHub account or org, and its members can sign in from the terminal, push dossiers and use ask mode. Sign-in says which workspace your repository goes to, and a refusal says why.
A clearer PRD page, with Approve spec always in reach
A PRD's page opens on one header box, which stays at the top on larger screens: the title and its buttons, its stage, repository and links, then the tabs. Questions use the full width, Approve spec is readable, and outlines and later stages are easier to see in every theme.
Every app page now has a sidebar: Home, PRDs, Questions, Knowledge, Fleets, Docs and Release notes. The top bar shows the page's title, the theme and Game mode, and your avatar, whose menu signs you out. On a phone, ☰ opens the sidebar as a drawer.
On a computer, the arcade no longer switches to fullscreen when you first click or press a key. A small ⛶ button in the bottom-right corner of the page, or the F key, takes it fullscreen when you want; Esc or F brings the window back. Phones still go fullscreen on the first tap.
A PRD's page now shows where it stands, from idea to retro, with one button for what to do next: approve the spec, answer the outbox, review and merge, or read the retro. "PRD #n" links to its issue, and new Outbox and Retro tabs show the decisions and the retro.
Install the Omni Loop with one npm line and omni init
With Node, git, gh and Claude Code on your laptop, one npm line puts omni on your PATH in any shell, and omni init opens the install pull request, installs the Claude Code plugin and signs you in. The Install guide goes from twelve steps to five.
Fleets are now optional: pick one or play solo, and a workspace with no fleets goes straight to the game. Each workspace's owner creates, edits and retires its own fleets on the new Fleets page, with a name, colour, motto and mascot.
Checks, brainstorms and shipping no longer crash in a repository whose list of files passes 1 MB. File names with spaces or accents are now read correctly too. Run omni update to get the fix.
A PRDs link at the top of every page and a My PRDs card on your dashboard open the PRD list, which starts on the PRDs you opened, drafts included, with one click to the whole workspace. Claude now gives a PRD's page link whenever it answers about that PRD.
The PRD page opens on its questions and refreshes itself
A PRD's page now opens on its questions and shows each answered one as its answer alone. Quick questions are answered with one click on the list, a bigger one brings you back to the list once answered, and new questions appear with no reload.
The HOME poster fills a desktop screen, its lines share one left edge, and its top line is cyan and easy to read. The planet turns, the crest sits under it, and Omni-man flies past every 12 seconds, except when your device asks for reduced motion.
Anyone can now sign up from the home page with GitHub: install Omni Loop on your organization or your own account, and get a workspace named after it. Members of an organization already on Omni Loop join it when they sign in. GitHub is now the only way to sign in.
omni on your PATH, and badges on every guide code block
The Install guide now puts omni on your PATH with one block you paste once per laptop, so omni works from any folder of a repository that has the kit. Every code block in the guide says where it goes: Terminal, Coding agent (or both), a file, or a GitHub comment.
Know which kit you run, and update it in one command
Every merge to main now publishes a numbered release of the Omni Loop kit. omni version says which one a repository runs and whether a newer one exists; omni update opens one pull request that brings it to the latest, and updates your Claude skills.
Claude Code shows two lines at the bottom of the terminal: the model, how full the context is and the 5-hour usage, then the PRD the session works on, its stage and, while it is built, its wave and slices. Setting up the loop turns it on for everyone.
A Getting Started guide, and Release notes in the top bar
A public guide at /docs shows how to install Omni Loop, invade a repository and ship a first PRD, with fixes for common errors. The home page opens it from Getting Started. Every app page now shares one top bar, with Release notes and Docs.
The app's home page now shows your hero, your fleet, your season's points and rankings, a chart of the pull requests you got into main this week, and counts of your answered questions, settled decisions, new PRDs and the questions waiting for you.
Ask the loop how it works and where your work stands
Help explains the loop's stages, its principles and every command, in the terminal and inside Claude. Status shows how many PRDs have shipped, wait in the inbox, are being built or wait for review, a bar of delivered work, and the PRDs you worked on.
The homepage now opens with what the loop gives you: agents ship, you steer. It speaks to engineering heads, developers and product managers, shows what you see while agents work and how to get in and out, and explains every loop term it uses.
A build now ends by showing where its files are, which stage of the loop it has reached, and the next steps for how it ended: review and merge a finished build, answer the questions it left open, or clear what holds it up. Reworking a build after those answers ends the same way.
The app opens in Omni, with Light and Dark one tap away
The app's pages open in Omni, the homepage's own colours, and the theme switch offers Omni, Light and Dark. A saved choice of Light or Dark is kept, and anyone who had chosen System now sees Omni.
A brainstorm now ends by showing where its PRD's files are, which stage of the loop it has reached, and the next steps in order: review the proposal, merge it, then start the build. Planning a PRD on its own ends with its next steps too.
Every PRD Omni Loop ships now comes with a release note, reviewed with its code and published on a public release notes page, grouped by week, with a version number that never changes. The PRDs shipped so far form the initial release, 0.0.1.
The first public release of Omni Loop gathers everything shipped from 24 to 27 September 2026: the kit, its plugin for Claude Code, skills for any coding agent, the omni-loop GitHub App, ask mode and its question history, a knowledge base that fills itself, PRD dossiers, a design system, and the arcade that turns delivery into a game.
Install the delivery loop in any repository
The Omni Loop kit packages the brainstorm, build, review and ship loop as one command-line tool and one configuration folder, so any repository can run it without a hand-made copy of its own.
PRD 3
Brainstorm, build and ship with three commands
The omni plugin for Claude Code gives the loop its commands: /omni:brainstorm turns an idea into a reviewed PRD, /omni:yolo builds it in parallel slices, and /omni:yolo-fix reworks what a reviewer disagreed with.
PRD 7
Every open question, visible on the pull request
The omni-loop GitHub App adds an outbox check to every pull request. It stays red while a decision the agents took still waits for a person, so reviewers see open questions before they merge.
PRD 28
Set up Omni Loop with one line
omni init installs the loop in a repository with a single command: it writes the configuration, adds the command-line tool, creates the labels the loop uses and says what is left to connect.
PRD 39
A playbook that tells agents how your repo works
Each repository gets a playbook of short forms (how it tests, what a merge publishes, what must never break), filled from what the repository already documents. Agents read it at every step of the loop.
PRD 45
Decision reviews with a sense of humour
Every question the agents leave for a person on a pull request now opens with a short intro and ends with a punchline, so a long list of decisions reads like a conversation rather than a form.
PRD 50
Your repository's rules, found and written down
/omni:invade explores a repository and proposes its knowledge base: principles, business rules and invariants, pointing at the pages that already state them or drafted from what the code enforces.
PRD 68
Answer Claude on a page, not in a terminal
Ask mode shows the questions Claude asks during a session on a clean web page, with options and previews side by side. It is switched on per checkout, and the terminal takes over whenever the page cannot answer.
PRD 71
Every delivery comes with its own retrospective
When a PRD merges, the Omni Loop app counts what went wrong during its delivery and opens a retro pull request with findings, evidence and proposed lessons. It runs again 14 days later.
PRD 72
Every decision lands in the knowledge base
When a PRD merges, the decisions taken while building it are sorted into the knowledge base as decision records, rules or invariants, with who answered and who merged. One pull request carries them all.
PRD 82
The arcade, in your pocket
On a phone, the Omni Loop arcade appears as a handheld console with readable text and real buttons; on a computer, the screen alone fills the window. The game itself does not change.
PRD 94
Omni-man signs the loop's work
Every commit, pull request and issue the loop makes carries Omni-man's signature, and a new command, omni credits, counts his contributions across the organisation.
PRD 99
One loop, many teams
Everything the game holds now belongs to a workspace, with its own members, fleets, repositories and colours. Vertuoza is the first workspace, and others can join without seeing each other's data.
PRD 100
One look for everything Omni Loop
A single design library holds the Omni Loop logo, colours, fonts and pixel sprites. The arcade and every reading page draw from it, and a catalogue page shows each piece.
PRD 141
Every terminal gets its own tab
Ask mode keeps one session per Claude Code terminal. The page shows each terminal as a tab, so questions from sessions running side by side no longer overwrite each other.
PRD 142
No answer is ever lost again
Every question Claude asks is kept for good, with its repository, branch, PRD, cost, category and who answered. The whole workspace can browse the history, and a live question can be shared with a teammate.
PRD 144
See your knowledge base as a galaxy
Each domain of a repository's knowledge base appears as a star system of its principles, rules and invariants, in the arcade and on a plain reading page.
PRD 149
Shipping now levels you up
Every point earned by delivering also counts as experience that never resets. Levels unlock games in a new game room, starting with Entropy Invaders at the very first point.
Each PRD's spec, plan and before/after page are kept on the Omni page with every version, next to the questions that shaped them, for the whole workspace to read.
PRD 216
Jump between work and play in one tap
A Game mode button on every app page and an App mode switch in the arcade move you between the reading pages and the game, with a confirmation before each switch. The app gets its own home page.