Each update to Ownercode Pro, newest first, in plain words.
Version 0.21.0
For the people who use it
Your pages read less like a machine wrote them. Many readers now take the long dash (—) as a sign of machine-written copy. The skills that write words your customers read (pages, sales rewrites, video captions and titles) now leave it out.
For you, the owner
Your agent now builds the least that works, by its own rules. The project rules file (AGENTS.md) has a short block that runs before any new code: does it need to exist, is it already in this project, does the language or the browser already do it, does an installed package do it, can it be one line. No code "for later". Input checks, error handling that prevents data loss, security and accessibility are never cut.
Shortcuts stay visible. When the agent picks a simple version with a known limit, it tells you the limit and marks it in the code with a shortcut: comment that says when to upgrade.
A new skill finds what to delete.abstraction-police has three modes. review lists what a change could cut, one line each, with what replaces it. audit sweeps the whole project, including packages nothing uses. debt lists every shortcut: comment and names the ones with no upgrade trigger. It reports only; you pick what to cut.
The final reviewer stops code the change does not need. The critic now also blocks a new package that repeats what the project or platform already does, a second copy of an existing helper, and code with only one user that pretends to be general.
Claude now starts at medium effort, not High. Medium is the Opus 5.5 default, and its makers say to start there. The higher levels (xhigh, max) make each turn longer and use more of your allowance; use them only when a hard task shows they help. Codex stays at High.
One repo can opt out of the safety guards; the rest keep them. A repo that runs its own guards commits a file named .ownercode/own-guards whose first line says where its guards are. In that repo the Ownercode guards step aside and the session start says nothing. Every other repo on the computer stays guarded. The file must be committed and not empty, so an agent cannot switch the guards off by writing one file. The old switch for the whole computer (OWNERCODE_GUARDS=off) still works.
The page check catches a long dash that slips through. The built-page check fails a page that shows a long dash and names the page. If you like the dash, add the line long dash: allowed to docs/copy-rules.md. Chat, notes, docs and task files are not checked.
What you must do
Run the Ownercode sync skill once in each project. It brings the new AGENTS.md rules and the new .ownercode/check-guards.mjs. The new skill and the critic change arrive with the plugin update.
In Claude Code, set the effort to medium in the model picker.
If you set OWNERCODE_GUARDS=off on your computer for one repo only: commit .ownercode/own-guards in that repo, then remove the variable from your computer's environment and restart the tool.
If a built page already has a long dash, the next page check names it. Rewrite it, or allow it in docs/copy-rules.md.
If you use the Ponytail plugin, you can turn it off.
Version 0.20.0
For you, the owner
Four answers to a design review finding, not two. Blueprint's fresh-eyes review (the goldfish review) now lets you answer each finding with fix, contest, escalate or decline. A contest must point to a line the reviewer could read. "We said so in the chat" is not proof, so that answer goes into the design instead. A "a stranger could not follow this" finding cannot be contested: the text changes. Escalate is for a question only you can answer, and your answer is written into the design. Decline is for a small, correct point not worth the change. These rules come from Google's Elephant Goldfish Model, which our review is based on.
The reviews now ask about the bugs that slip through most. The design reviewer asks what happens when two people or two devices change the same record at once, and which time zone each date uses. The final code reviewer (the critic) checks every change for four bug classes first: two writers on one record, unchecked input, an error that shows as "nothing here", and a date with no time zone. A 47-bug hunt on a real app found these four classes most often.
A count that tells you if a reviewer is worth its cost. Each review finding costs you one question. A new script counts, across your task files, which review lenses led to real fixes. It prints nothing while every lens earns its place. It names a lens with no fix in its last 5 runs, and a review record it cannot read. Blueprint runs it before each review.
Fixed
Kit scripts now run when reached through a linked folder. Eight kit scripts (the safety guards, start, the architecture check, the monitor, the pages check, the look check, the push-key maker) checked that they were started directly by comparing paths. Through a Windows junction or a symlink the paths differ, so the script did nothing and exited as if all was well. They now compare real paths.
What you must do
Nothing to install. Your next blueprint design uses the new answers and the new record format by itself.
Version 0.19.0
For you, the owner
A ticked criterion is checked, not trusted. A task's acceptance criterion can now end with check: and a command in backticks. A tick on it is the agent's claim that the command passed. While that task is in progress, Claude Code runs the command behind every ticked criterion when the session ends a turn. One failure blocks the stop and shows the agent the command and its last lines, so a tick means the proof ran. Unticked criteria never run, so a half-done task can still stop to ask you something. The command goes through the kit's safety guards first, and on Windows it runs in Git Bash like the agent's own commands. Silent when no task is in progress, when the task is another session's (not named in this session, or not written since it began), when the task's latest note is a NEXT: handoff, or when no ticked criterion has a check. Claude Code only; Codex has no stop event in this plugin.
A reviewer that cannot be talked into agreeing, in the kit. The final one-pass reviewer (the critic) now ships as a plugin agent. Before it reads a change it reads your project's incident: lessons (bugs that reached users) and the ## Not a bug list in docs/gates.md (findings you ruled out, so it stops raising them). Each finding names its kind, so the builder fixes every place of that kind, not only the line found.
A tester who has never seen your code. The new first-time-user skill sends an agent that cannot read files, code or your rules to use one page the way a new customer would, and reports every place it got stuck: where, what it tried, what happened, what it expected. It never sends a form or signs in on your live site; it works on the local copy or on public pages. Confirmed points become tasks, each with a check: command.
What you must do
Nothing to install. To use the stop check, write check: and a command in backticks at the end of an acceptance criterion, and tick it only when it passed; the format is in tasks/README.md after the next sync.
If the stop check blocks a session that must stop early, the agent adds a NEXT: note to the task, as the handoff rule already says.
Version 0.18.0
For you, the owner
Sessions find the next task without reading every task file. A new script, tasks/tasks.mjs, reads only the short header at the top of each task. next shows the task to work, list shows the queue in one short table, and sql answers any other question about the queue. Your task files stay the only copy; nothing new is stored.
Two sessions can no longer take the same task number. A new task takes its number and creates its file in one step, behind a lock.
A health check for your task list.node tasks/tasks.mjs check names a misspelled status (such as in-progress), a repeated number, a missing dependency, or more than one task in progress. It prints nothing when all is well.
A warning when your rules files get too long. Each session sends CLAUDE.md and AGENTS.md again with every step it takes. When the two together pass 12 KB (or, if the starter's own copies are bigger, when they grow past those), or a project skill passes 15 KB, the session start says so and offers to move long parts into files read only when needed. It never blocks anything.
A usage meter. The new usage-meter skill shows what your Claude Code sessions read and what share of the cost each kind of reading takes. After the first run it speaks only when something grew.
What you must do
Run the Ownercode sync skill to get tasks/tasks.mjs. Until then the skill uses the copy inside the plugin.
node tasks/tasks.mjs sql needs Node 22.13 or newer (node --version). The other commands work on any Node 22.
Version 0.17.0
For you, the owner
Risky changes are found in any project, with no setup. The stop before a risky commit used to know only the starter app's exact file paths, so it missed risky files in bigger projects. It now finds them by name anywhere: database changes (migrations), login, auth, access, roles, sessions, payments, billing, invoices, checkout, refunds, payroll, the security layer (middleware) and deploy settings (wrangler). Notes and tests never count.
You can correct the list. In your project's docs/gates.md, a "Owner must see" section adds paths, and a "Not risky" section removes false matches. The agent adds a "Not risky" entry only after you say yes.
Notes now cover every app in the project. A project with extra apps in their own folders (an admin app, a customer portal, background workers) now gets notes for those too, named after the app, for example admin-customers.
Changed
Merges stop on the same risky files. When merge mode is on, the merge now stops on these files too, not only on the protected list in the merge settings. Before, a merge could skip the stop.
What you must do
If merges now stop on a file that is not risky, say "add it to Not risky" and approve it once.
Version 0.16.0
For you, the owner
Bugs your customers hit now teach your project's reviews. When the daily bug check (the monitor skill) finds a problem real users hit, the fix task now asks for a test that fails if the bug comes back. When no test can catch it, such as a confusing step, the task asks for one sentence on why. Either way, the fix adds a short lesson to your project's tasks/lessons.md file, marked as an incident. The daily bug check no longer reads that file as an open task, so a lesson can never hide a bug that comes back.
The final reviewer reads your past bugs first. Before a big piece of work is reported done, the reviewer (the critic, which looks for reasons not to merge) now gets your project's incident lessons. Each project builds its own list from its own bugs.
Fixed
The kit's own privacy tests pass again. Two skills used words the tests ban: one named a setting only the author's computer has, and one used a word from the removed receipt checks. Both are reworded.
What you must do
Nothing. New projects get the updated lessons file. In an existing project, the skills add the incident: line the first time the bug check finds a problem.
Version 0.15.0
For you, the owner
Plain-words notes on how each part of your app works. Each time work is committed, the agent now updates a short note for each part of the app it changed, in the docs/areas/ folder (for example docs/areas/contacts.md). Each note says what the part does, where it lives, how it works, which of your rules it follows and what can break. You can read how your app works now without reading code, and the next session starts from the same picture.
Risky changes wait for your yes. A change to payments, login or its settings, the security layer that runs on every request, the database structure or the deploy settings now stops before the commit. The agent tells you in plain words what changed and what could go wrong, and commits only when you say yes. Your project can set its own list in docs/gates.md, under "Owner must see".
Changed
Commit reports say which notes were updated. The ship skill's report now lists the notes it wrote and the ones it checked. A "Not explained" line names any changed file whose note was not checked. The line does not appear when every file is covered.
Merge mode uses the same risky list. The ship skill now says to copy every "Owner must see" path into the protected list of the merge settings file (.ownercode-pro/ship.json), so a merge stops on them too. When the merge script makes the commit itself, the agent still writes the notes first and commits them with the work.
What you must do
Nothing. The notes start the next time work is committed. To change which paths stop for your yes, add an "Owner must see" section to docs/gates.md.
Version 0.14.1
Fixed
Arch-guard accepts a plain JavaScript middleware. A project whose functions/_middleware.js is plain JavaScript on purpose no longer fails the architecture check. A project with no middleware still fails.
What you must do
Nothing.
Version 0.14.0
For the people who use it
Plain-English change log (new changelog skill). When work finishes, you now get a short note on what changed. You get it when a big program ends, when a task closes, and when work is committed. It uses plain words, with the things your customers see first. It no longer ends with a commit hash and a file list.
Three ways to use it. "Write" runs at each closure and adds the note to the project's CHANGELOG.md file. "Tell" answers "what changed?" or "explain the updates" in chat only and writes no file. "Release" moves the unreleased notes under a version heading and bumps the version.
Fixed layout for every note. Notes are sorted into groups, in this order: for the people who use it, for you the owner, fixed, changed, removed, behind the scenes, and what you must do. A group with nothing in it is left out. "What you must do" always appears, and says "Nothing." when there is nothing to do.
No gaps without a warning. The note ends with a "Not described" line that lists any changed file no item covers, with the reason. If every file is covered, the line does not appear.
Each change is logged once, and the log is always saved. A ship inside a task inside a program writes one note, not three: changes already in an unreleased note are skipped. The CHANGELOG.md file goes into the closing commit, so it is never left unsaved.
Changed
Program reports now carry the change log. In the orchestrator skill (the one that runs big requests as a program), the final "program done" report now holds the change log note under the progress meter. The report that a helper session sends back gets one new line, Changelog:.
Closing a task now writes the change log first. The note is pasted into the task's Notes as the closing note, and then the task is archived. The tasks skill points to this step.
Commit reports now start with the change log. The ship skill's report opens with the note, then the commit hash, the files, and each check result.
"What changed?" now has a route. The orchestrator sends "What changed?" and "explain the updates" to the change log skill in tell mode.
What you must do
Nothing.
Version 0.13.0
sales-council now judges the offer, not only the words. A new Offer review mode runs per term of a deal (floor, share, step, cap, build period, exclusivity, notice) and ends in a term table and the whole offer in one line. Three new seats: Lemonis (can the business behind the offer carry it), Musk (does each term need to exist; price from the cost of inputs), MrBeast (the title, the thumbnail and the retention curve of anything watched). Lemonis and Musk were tested with and without on a real offer and seated on the count; O'Leary was tested the same way, added nothing a sitting seat did not, and went to the candidates file with his two best plays folded into Tracy and Trump. MrBeast is provisional ("Tested: pending") until a video script comes through. A new bench row for video scripts, video ads, reels, titles and thumbnails. You: nothing. Say "review the offer" or "is this the right rate" to get the new mode.
Version 0.12.0
blueprint gets a goldfish review before it writes the task. Fresh reviewers that never saw the chat read the draft design, the code and a short checklist, and report what is wrong: is there a simpler way, can a stranger follow it, how does it fail, is any decision left to the builder, are its claims about the code true. Medium features get two reviewers, large or not-reversible ones get five. You answer each finding with fix or reject; the task keeps a "Review record" of both, so a later session does not undo a choice made on purpose. The six reviewers ship as plugin agents in a new agents/ folder. Discovery also walks six lenses (need, cost, ease of use, migration, how failures show, future debt).
blueprint now writes its task in the tasks format (NNN-KIND-slug.md). An existing .claude/blueprint-config.md with its own naming still wins.
orchestrator decides by a written rule when to design first, runs blueprint in its own session, keeps the task current until it closes it, and compares the shipped work to the design at close. It now says so when its second-opinion advisor is missing or failing.
You: nothing. Claude Code older than 2.1.271 gives the reviewers your CLAUDE.md too; update Claude Code if you want them fully fresh.
Version 0.11.13
A new row in docs/seo-rules.md: other-host-paths, for pages your build makes but another domain serves (one Worker answering two domains). The search check no longer says their canonical points elsewhere or that your sitemap leaves them out, and it does not ask for noindex, which would drop them from the other domain's search too. Every other page check still runs. You: nothing, unless your site builds pages for a second domain; then list their path in that row.
Version 0.11.12
ship can run your project's own checks before a merge, and can commit exactly the files you name. New optional key in .ownercode-pro/ship.json: "gates", a list of commands (each a list of strings, no shell), run in order at the head commit before the push and on a dry run too. Tokens {head}, {base}, {changedFilesFile} and {bodyFile} give a gate the commit, the changed files and the pull request body; a gate may edit the body (for example to add a deploy receipt), and the script uses the edited body. A failing gate stops the run. Gates run in the working folder, so with gates any file left out of the commit stops the run first; ship from a clean worktree. New flags --files a,b --message-file msg.txt: the script stages only those files and commits them (hooks run); other dirty files stay behind when the project has no gates. A named file that is already committed is skipped, so a rerun after a failed gate works. You: nothing, unless your project has a check that must pass before a merge; then list it under gates.
Version 0.11.11
A new skill, monitor, finds errors and quiet drop-offs on your live site before anyone reports them. Once a day a script reads your site's error and funnel counts and keeps only what is new or rising and has no open task. The agent finds each cause in the code, writes a task, and for a small, clear fix opens a draft pull request with a test. It never merges or deploys. It says nothing when nothing is wrong, and it tells you when the reporter itself went silent. Ask "what broke for users" or "run the monitor". Your project's facts (the export command, your task command, your site addresses) go in .ownercode-pro/monitor.json; the skill's setup mode adds a reporter that sends no personal data, and its schedule mode runs it every day. You: nothing, unless you want monitoring; then ask the agent to set it up, and say whether it may open draft pull requests.
Version 0.11.10
ship no longer hangs on a change that CI skips. Some projects skip their CI workflow on certain paths (paths-ignore, for example docs). A pull request that touched only those paths never got the named checks, so the script waited the full waitMinutes and stopped. New optional key in .ownercode-pro/ship.json: "checksSkipPaths": ["docs/"], path prefixes like protected. When every changed file is under one, the script waits for no checks, and its dry run and SHIPPED line say the checks were not run (never "checks passed"). One file outside the list: it waits as before. You: nothing, unless your workflow has paths-ignore. Then copy those paths into checksSkipPaths exactly; a wrong entry merges without CI.
Version 0.11.9
The speed, look, placeholder and search checks were run on a large live site for the first time, and most of their 29,000 failing lines were faults in the checks, not the site. Fixed: the search check reads a bare alt as an empty alt, ignores tags inside a script, counts & as one character, and does not call a link dead when it goes to a private area or a path a Worker answers. The speed check sizes only images a page, a stylesheet or a script names (not the originals the build leaves unused), finds the first image inside <main> (not the footer logo), and asks for the font preload only on pages that load the font. The look check no longer calls a quotation's left rule (blockquote) a side stripe. The placeholder check skips code samples and reads "your company name" in a sentence as plain English; a template slot ("Your Company Name", "[company name]") still fails.
Two new rows in docs/seo-rules.md: noindex-paths for public pages you keep out of search on purpose (tag lists, print copies; * stands for one path part), and server-paths for paths a Worker answers (photos served from storage). See the seo-audit skill.
A script from another server (an analytics beacon) is now its own exception item, speed-outside-js, apart from the byte budget speed-js. You: if docs/design-exceptions.md has a speed-js row that was meant for an outside script, rename it speed-outside-js.
Version 0.11.8
A new skill, microsite, for small rank-and-rent sites, in three modes: find (score niche and town pairs and plan a 40-50 page site), build (from a chosen target to a live site with a call-recording phone line, one phase at a time, with your yes before any domain is bought) and audit (how an existing site ranks against its competitors, with rank tracking). Your own facts (home towns, where new sites go, which sites you own) go in docs/microsite.md; the scores and history in docs/microsite/. You: nothing, unless you build microsites; then the agent sets up those files the first time.
Version 0.11.7
Fewer skills, same steps. next-task, new-task and close-task are one skill, tasks, with three modes: next, new and close. Start a session with /ownercode-pro:tasks next (Claude Code) or $ownercode-pro:tasks next (Codex). You: use the new name; the agent also picks it by itself when you say "next task", "add a task" or "close it".
commit is now the first step of ship: it runs your project's gates from docs/gates.md, stages only the session's files and commits locally, exactly as before. Say "commit" or "ship it". You: nothing.
grilling, domain-modeling and prototype are now modes of blueprint: grill (stress-test a plan), terms (a glossary or a decision record) and prototype (throwaway code that answers a design question). You: nothing.
research gains a deep mode for big questions: a planner splits the question, helpers research the parts in parallel (outside sources and your own project), and one review round fills real gaps. Every claim still comes from a primary source. You: nothing.
Starter files: AGENTS.md, tasks/README.md and the build conventions name the new skills. You: run the sync skill in each project when the session start asks.
Version 0.11.6
A new skill, video: plan, cut, caption, score, thumbnail and upload short marketing videos, website hero loops and YouTube Shorts, with the tools (ffmpeg and headless Chrome scripts) and worked recipes. Your brand lives in one file, docs/video-brand.js (logo, colors, fonts, site, phone); every video ends on it. Your own video rules go in docs/video.md. New: a short logo call-to-action scene and a before-and-after thumbnail. You: nothing. The first time you ask for a video, the agent asks for your logo and colors and writes the two files. The tools need ffmpeg and Chrome on your computer.
Version 0.11.5
The orchestrator has a Survey mode. Ask "what should we improve here?" and it audits the repo without touching code, shows a ranked table, and writes your picks as tasks. Then "work the task list" builds them. You: nothing.
Version 0.11.4
The design skill has a new file of places to look for ideas: design systems as token sets, ten looks to start drafts from, and effects (animated gradients, text reveals, a before-and-after slider). Each says what it needs. The agent builds the idea in your site's stack and keeps it fast.
Five more "looks AI-made" items to avoid, a logo rule (your real logo, never an invented one), and a one-line "design read" before the first drafts.
A cursor effect is now allowed as a desktop extra, when the page works the same on a phone without it.
Two more skills: sales-council (review or write anything meant to persuade) and systematic-debugging (find the root cause before a fix). You: nothing.
Version 0.11.3
The design skill now takes what works from the sites you like: layout, structure, type, color and motion. A design gallery, a template or a known brand's design file is also a fair starting point; the agent makes it yours with your colors, photos and words. Before, it could only describe those sites in words, and known brand design files were banned. Your page still uses your own photos and words. You: nothing.
Version 0.11.2
A site with a private area other than /app and /login (a /console, an /internal) can name it: one row private-paths in docs/seo-rules.md. The search audit, the speed check and the placeholder check then treat it as private. A private page must still be noindex and out of the sitemap. You: nothing, unless your site has such an area; then ask the agent to add the row.
The placeholder check now reads docs/design-exceptions.md. Words you want on a page that look like a sample (a page that says "your company name") pass with a row there: items sample-phone, sample-address, sample-name, filler-text, unfinished-marker. The starter's Worker name has no exception. You: nothing.
An exception for a route now covers that one page. Before, a row for /services/ also covered /services/roof/, and a row for / covered every page. A row with a file name (global.css) now matches the built file, which has extra letters in its name; before it matched nothing. You: if a check fails after the update on a page an old row used to cover, add a row for that page, or use all.
pre-deploy has one rule for a check your project never ran before. A failing check that is not listed under ## Accepted checks in docs/gates.md is reported under "New checks, not yet accepted by the owner", with its failing lines and your two ways out: accept the check, or record an exception. It still blocks the release. You: when the report shows that heading, tell the agent which way you want for each line.
Version 0.11.1
The handoff skill gains a named Compact Focus Arg section: the rules for what a compaction keeps and drops. A tool that compacts a session on its own can read the rules from there. You: nothing.
Version 0.11.0
One plugin. Everything that was in the free ownercode plugin is now inside ownercode-pro: the project rules and starter files, the task skills (next-task, new-task, close-task, commit, handoff), setup and sync, and the safety guards. The private marketplace lists one plugin. You: update, then remove the old ownercode plugin in your tool's plugin manager, or its safety hooks run twice. In Codex, type /hooks and trust the Ownercode hooks again; until you do, the guards are off. Then run the sync skill in each project: it updates the project's settings to name the one plugin.
Skills run as /ownercode-pro:<name> in Claude Code and $ownercode-pro:<name> in Codex. The old /ownercode:<name> form is gone. You: nothing; the agent picks skills by their short name.
decision-interview-basic is gone. The full decision-interview asks those two questions and keeps answers the basic one already wrote. You: nothing.
The free guide ships as BASICS.md, next to GUIDE.md. You: nothing.
The whole kit is under the one license in LICENSE. You: nothing.
One off switch for the safety guards, for a computer that already runs another guard: set the variable OWNERCODE_GUARDS=off in the environment your tool starts in. The session start says so when a project's own settings file sets it. You: nothing. Leave it unset unless you know you need it.
The version of a project's starter files now follows this plugin's version. A project set up from the free plugin holds 1.0.x; the session start asks for a sync once. You: run the sync skill when it says so.
Version 0.10.0
design-system now stands alone. The floor (labels, contrast, keyboard, the four states) is in its own reference file; it no longer points at the Free interface guide. You: nothing.
It opens your own design library first, when you have one: a folder with an INDEX.md that says which rule files and screen pictures to read. Ownercode ships no library. You: nothing if you have none. If you have one, set the variable DESIGN_LIBRARY_DIR to the folder, or write a "Reference library" line in your project's DESIGN.md.
Every design report now has the line "Library used: ...". If you named a library and it cannot be opened, the line says "design library not found" and the work goes on.
The impeccable plugin is the second install. When you have it, the design skill hands it four jobs: a new site or redesign with no locked brand, motion, the finish review, and the design document written from the finished build. Without it, the skill uses its own rules as before. You: install impeccable if you want those four jobs; it is a separate, free plugin.
On your request the skill runs the separate checker npx impeccable detect --json <folders> on your files. It never runs by itself.
Your locked brand and your copy rules win over the library and over the plugin.
Version 0.9.0
Eight new skills. None changes your project until you use it. You: nothing.
blueprint designs a new feature with you and writes a task file with the reasons; it never builds. Say "let's design X".
grilling asks you one hard question at a time about a plan, each with a recommended answer. Say "grill me on this".
prototype builds throwaway code to answer one design question: does this logic feel right, or what should this screen look like.
domain-modeling pins down the words of your business in CONTEXT.md and records big decisions in docs/adr/.
research sends a helper to read first-hand sources and saves the findings, with sources, as a file in your project.
credential-scoping runs before any key or token is made: which lane, which bill, the smallest scope, and how the value moves so the agent never sees it.
human-steps shapes every step you must do by hand: the exact place, values you can copy, and what you see when it worked.
learn saves a lesson from a session where the next session finds it. Say "save this lesson".
Version 0.8.0
The build skills now keep the method in the plugin and your project's own facts in small files in your project. Each skill names its file. Nothing changes until a file exists. You: nothing now. The agent writes a file with you the first time it is needed.
ship has a merge mode: push, pull request, wait for the checks you name, merge. It is off. You: nothing. To turn it on, tell the agent "turn on merge mode"; it writes .ownercode-pro/ship.json and needs the GitHub command-line tool (gh).
arch-guard reads your project's own rules from .ownercode-pro/arch-rules.json. You: the installed commit gate is now out of date; ask the agent to "run the architecture gate installer again".
seo-audit checks your own limits (title and description length, link text, enough copy) from docs/seo-rules.md. You: nothing.
pre-deploy has a new check: it fails when a public page still shows a sample phone number, example.com, filler text, or the starter's Worker name. It also runs your own checks from docs/gates.md. You: before the next deploy, give the agent your real phone number and site name if it asks.
page-template has copy and title rules, and fails a built page that shows a phrase you banned in docs/copy-rules.md. You: nothing.
worker-endpoint is one full skill; your route list lives in docs/endpoints.md. You: nothing.
New in Pro: component-builder, handoff, commit, new-task, next-task. They are fuller versions of the skills with the same names in the included Free plugin. next-task now writes a short plan in the task file before it builds, and stops when a business fact (a price, a phone number, a claim) has no source. You: nothing. Until the Free copies are removed you see each name twice; use the ownercode-pro one.
Version 0.7.0
New skill: orchestrator. It runs a big or unclear request as a program: it lists what is known and unknown, sends helpers to research the unknowns, cuts the work into chunks with a done check each, runs them, and checks every result before it reports. It asks you only for choices that cannot be undone, are public, cost money, or touch an account. Every other choice it makes is listed as "Assumed". You: nothing. To use it, say "orchestrate this" or "run this as a program".
"Work the task list" runs the ready tasks in tasks/ one after another, and stops when the list is empty or a task needs you. You: nothing.
Version 0.6.2
Before any deploy, the pre-deploy check now fails while the site address is still the sample https://example.com (in the canonical links, the sitemap and robots.txt). The go-live task sets the real address first. You: before the next deploy, tell the agent your site address: your own domain, or the workers.dev address until you have one.
A site with no database connected answers "The database is not connected" (503) instead of crashing, in the starter app and in the website added to a free CRM. You: on a site built from the starter app, ask the agent to bring in the new functions/_middleware.ts check and its test. On a free CRM with the website, see the Free 1.0.4 changelog.
Every answer now says no-transform, so Cloudflare cannot add its Web Analytics script, which the site's own security rules block (a console error, and a lower Best Practices score). You: same middleware change as above. If you want Cloudflare's analytics, ask the agent.
Session handoffs name the plugins by <pro-root> (found each session), never by a full folder path, which breaks at the next update. You: the session start names any file with an old path; let the agent fix it.
The update steps now update every installed copy, and read this changelog. Includes Free 1.0.4.
Version 0.6.1
The speed check says when sharp is missing and the site uses Astro images. You: nothing.
Version 0.6.0
Optional add-ons: instant price, visit tracking, lead alerts, phone line, customer timeline, follow-up list, review asks, visit-time requests. Each stays off unless you pick it. You: nothing. To add one, ask the agent; it names each account, cost and legal point first.
Version 0.5.1
The starter app needs Node 22.18 or later, and says so in one plain line on an older Node. You: if you see that line, install Node 24 LTS.
Ownercode Pro · A plugin for Claude Code
$399 then $99 a year for updates
A plugin for Claude Code. It runs on your own Claude plan.
Your website and CRM, built with you.
Guards on every step.
For you and every business you own.
Add startup training: $500
6 hours with us, one time, to get you building right. After that, $70 an hour.
Needs your own paid Claude plan. Claude Code comes with it.
The first year of updates is included. Stop renewing and Ownercode keeps working; only new updates stop.