Changelog
Version history, new features, and breaking changes for Pushary
Updated
Destructive Codex steps ask your phone at the keyboard under When I am out, November 21, 2026
Versions covered by this entry:
- Pushary server: from this deploy, for a Codex call the Pushary Mac app forwards.
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.230, for a Codex call the hook judges on your computer. - Pushary app: 1.0.66 or later approves such a step behind Face ID or the passcode.
Changed
- Under When I am out, a destructive Codex step asks your phone even while you are at the keyboard.
rm -rf,git reset --hard,git clean -fdx, a force push,git branch -D, dropping a table,npm publishand a patch that deletes a file used to go to Codex's own prompt while you were at the keyboard, and Codex can run some of them inside its sandbox without asking. A call nobody answers is refused, and so is a call when no device is connected. A phone older than 1.0.66 can only refuse it, so approve it from the web page or Pushary for Mac. - A deploy or release keeps today's flow.
vercel deploy --prod,npm run deploy,gh release create, and a command that only names a release, such asgh release listorgit checkout release/1.2, still go to Codex's own prompt at the keyboard and to your phone when you are away. - Risky steps need the keyboard keeps today's flow. With the workspace setting on, these steps stay with Codex's prompt at the keyboard, because the phone cannot approve them there.
Skipping permissions asks nothing by default, and Grok Build's Always-approve joins, November 21, 2026
Versions covered by this entry:
- Pushary server: Claude Code on Pushary CLI 1.9.97 or later and the Pushary Mac app follow the new default once their cached policy refreshes. No update needed. CLI 1.9.96 keeps "Stop me for risky steps" until you upgrade, because it applies the setting to Codex too. Earlier CLIs never had the setting and are unchanged.
- Pushary CLI 1.9.228: Grok Build in Always-approve follows the setting. Run
npx -y pushary@latest upgrade.
Changed
- "Never stop me" is now the default for "When your agent skips permissions". With Claude Code run with
--dangerously-skip-permissions, every step runs without a question, a merge, a force push orrm -rfincluded, unless a rule you wrote for that tool or command asks about it. Your workspace's rule for every tool (*) does not count. Until now the default was "Stop me for risky steps", which asked about those steps. A workspace or a Team member who picked "Stop me for risky steps" in Settings, Connections keeps it. To be stopped for risky steps again, pick it there. - Grok Build in Always-approve follows the same setting.
--always-approve,--yoloand--permission-mode bypassPermissionsnow ask nothing by default either. With "Stop me for risky steps" they ask only about risky steps. Before, Pushary asked about every call your policy gated, as in Grok's default ask mode. - Still applied: the kill switch, a deny rule, a rule you wrote for the tool or command that asks, a preset you picked, an approval route, a mode override, a scope you approved, and your agent's own questions.
- Unchanged: Grok's default ask mode, Claude Code's other modes, and Codex, Gemini CLI, Antigravity, Cursor, OpenCode and Hermes. Codex reports skipped permissions whatever you chose, so it keeps asking what your policy asks.
Risky git steps written around a substitution now ask, November 21, 2026
Versions covered by this entry:
- Pushary server: no update needed for steps the server judges, such as a phone call, the phone card and the editor gate.
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.227, for steps the hook judges on your computer.
Changed
- A substitution no longer hides a risky git step. A Pushary
*rule and the Autonomous preset now ask beforegit reset "$(git merge-base HEAD origin/main)" --hard,git clean -d "$(git rev-parse --show-toplevel)" -f,git push origin "$(git branch --show-current)" --forceandgit -C "$(git rev-parse --show-toplevel)" push --force, as they ask beforegit reset --hard,git clean -fandgit push --force. Before this a*rule ran all four without asking, and the Autonomous preset ran thegit resetandgit -Cones. Always allow Bash and permissions off with no Pushary rule now also ask before every risky git step written aftergit -C "$(...)", such asgit -C "$(git rev-parse --show-toplevel)" reset --hard,clean -fd,branch -D,push --force,push origin +mainandpush --prune, and beforegit push origin "$(git branch --show-current)" --delete. These still reach the phone as normal steps, where one tap approves them; that is a known gap we will fix separately. - A step inside a commit message's substitution is read. A
*rule and the Autonomous preset now ask beforegit commit -m "$(cat <<'EOF' | sh ...)", before a command written after theEOFinside the$( ), and before a$(rm -rf x)inside a heredoc whose delimiter is not quoted (cat <<EOF). Claude Code's usualgit commit -m "$(cat <<'EOF' ... EOF)"message is still not read as a command. git push --pruneis risky. With--all, a wildcard refspec orpush.default=matchingit deletes every remote branch you have no local copy of. Pushary cannot see your git push settings, so everygit push --pruneis now flagged likegit push --delete: a*rule asks before it and the phone offers Deny only.git fetch --pruneandgit remote prunestay routine.- A git alias given on the command line is read. The Autonomous preset now asks before
git -c alias.y='!$(cat notes.md)' y, as it asks before$(cat notes.md). The alias text is never matched against your rules or approval routes, so it cannot move a push out of the route you set for it. A*rule still runs both, andsh notes.mdstill runs in every mode: what a file contains cannot be read from the command.
Long Hermes commands are marked as unchecked, November 21, 2026
Versions covered by this entry:
- Hermes plugin (
hermes-plugin-pushary): 1.0.9. The Hermes catalog copy moves to 1.0.9 when its pin moves. - CLI (
pushary, also published as@pushary/agent-hooks): 1.9.226 installs plugin 1.0.9 withnpx -y pushary@latest setup --agents hermes. The Pushary Mac app installs it from its next release. - Pushary server: reads the mark from plugin 1.0.9. The risky card shows on Pushary app 1.0.66 or later.
Security
- A long Hermes command Pushary could not read in full is marked. Hermes approvals carry the command up to 4,000 characters. Until plugin 1.0.9 a longer command was cut without saying so, so a step at its end, such as
rm -rf ./buildafter a long heredoc, was not seen and the approval read as ordinary. Plugin 1.0.9 ends the cut with… [truncated], and Pushary records the approval as risky, with "Pushary could not check this step", as it does for a long command from the Pushary CLI. When Hermes offers only Allow once and Deny, Pushary app 1.0.66 or later shows it as a risky step: its card asks for Face ID, Touch ID or the passcode before Approve, and its notification offers Deny only. Older apps, Slack and calls still offer Approve there. Hermes's usual approval offers four choices, and for it the phone and the web keep those choices as they are: the only sign of the cut is the… [truncated]mark at the end of the command. - When Pushary is not Hermes's approval transport, a
terminalcall gated withPUSHARY_GATE_TOOLSsends its command, not a summary. The plugin used to send a 500-character summary of the call's arguments, so a long command was judged on about its first 200 characters. It now sends the redacted arguments up to 4,000 characters, cut with the same mark, and the summary stays the question's context. That question is a yes or no, so the risky handling above applies to a cut one. With Pushary as the approval transport, which setup selects, a gated call reaches Pushary as Hermes's own four-choice approval, carrying a label instead of the command and the 500-character summary as its context, and 1.0.9 leaves that unchanged. Every other gated tool sends the same 500-character summary as before.
One alert per refused protected file in Alert only, November 21, 2026
Versions covered by this entry:
- Pushary server: no CLI or app update needed. This covers Codex through the Pushary Mac app.
Changed
- One alert, not one per try. In Alert only, when Codex keeps trying to change a protected file such as
AGENTS.md, Pushary still refuses every try but now sends one alert per file in each session every 10 minutes, instead of one per try. A different file, another session or the same file after 10 minutes alerts again. Through the Pushary CLI's Codex hook, alerts still come once per try, within the limit of 10 notifications every 5 minutes.
More risky git spellings caught, and commit messages read as text, November 21, 2026
Versions covered by this entry:
- Pushary server: no update needed for steps the server judges, such as a phone call, the phone card and the editor gate.
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.224, for steps the hook judges on your computer.
Changed
- Reordered git flags are risky.
git reset origin/main --hard,git reset HEAD~1 --hard,git clean -d -f,git clean --force -d,git push -fu origin main,git push origin +mainandgit push --mirror originare now flagged likegit reset --hard,git clean -fandgit push --force. They ask under Always allow Bash, a*rule and the Autonomous preset, and reach your phone as risky steps. The Autonomous preset used to rungit reset origin/main --hardwithout asking. A$(git ...)before the flag, as ingit push origin "$(git branch --show-current)" --forceorgit reset "$(git merge-base HEAD origin/main)" --hard, no longer hides it: such a step is rated risky, so Always allow Bash asks and the phone offers Deny only. A*rule still runs these without asking, and so does the Autonomous preset for thegit resetone, as before.git clean -n,git push -uandgit push --dry-runstay routine. - A commit message is not a command. The quoted text of a commit or tag message (
-m,--message) and of a pull request or issue title, body or comment (gh prorgh issuecreate,editorcommentwith--title,--body,-tor-b) is no longer read for risky words, including Claude Code'sgit commit -m "$(cat <<'EOF' ... EOF)"form. Sogit commit -m "rm unused helper and enable -race in CI"andgh issue create --title "git push --delete fails"are routine steps, and an ask about one can be approved from the lock screen. This holds only when every step of the command iscd,git add,git commit,git tag, agit pushthat is not risky itself, orgh prorgh issuecreate,editorcomment, and the command has no pipe, redirect, heredoc,git -c, message or body file, template, and no$or backtick anywhere, inside the message too, unless it is in single quotes or escaped as\$. Anything else, such asnpm test && git commit -m "rm -rf the cache"orgit commit -m "rm -rf x" > run.sh, is read as before. Text printed withechoorprintfis still read as before. - Long commands are checked faster. A command made of 80,000 characters of
git pushtook 10 seconds or more to check and now takes about half a second, as long as any other command of that length.
Phone calls in German, French and Dutch, November 21, 2026
Versions covered by this entry:
- Pushary server: no CLI, plugin or app update needed.
Added
- Calls in four languages. Pushary calls in English, German, French or Dutch. Pick the language under Your number on the Calls page, with Call language. A number added on the dashboard starts in your dashboard language if it is one of the four, and English otherwise.
- The call follows you. Answer in one of the other three, or ask for one, and Pushary switches to it, voice included, until you switch again.
- The same checks in every language. Pushary reads the step back and asks again before a yes counts, in the language of the call. English calls decide as before.
Questions that closed before you answered stay on the phone, November 21, 2026
Versions covered by this entry:
- Pushary server: sends the list now. Older apps ignore it.
- Pushary app for iPhone and Android: the app's next release, once the App Store and Google Play approve it.
Changed
- A closed question stays in the Inbox for 24 hours. A question an agent asked you that closed before you answered it on the phone now shows under Closed questions asked in the last 24 hours, instead of disappearing. This covers questions from Claude Code's question tool and plan approval, Codex's question tool, OpenCode's question form and
ask_user. Tool approvals, including Codex's own approvals, and the setup test are not listed. - The label says what happened: Closed after 10 minutes, Asked in Claude Code instead when Claude Code's own prompt took the question, Cancelled by the agent, or Cancelled when the agent was stopped.
- Tap one to read it. It opens read-only with its choices and one line on where it is now, such as "Answer it in Claude Code if it's still waiting." It cannot be answered from the phone any more. Tapping the notification for a question that has closed opens the same view.
- Terminal and Updates mode still show nothing on the phone, closed questions included.
Path rules on Windows, protected files in paused sessions, and long commands, November 21, 2026
Versions covered by this entry:
- Pushary server: no CLI, plugin or app update needed. The long-command change shows on phones on Pushary app 1.0.66 or later.
Security
- A rule for one file now holds in VS Code and Cursor on Windows. The editor plugins send Pushary each file edit with its Windows path, such as
src\secret.tsorC:\Users\you\proj\src\secret.ts. Until November 21 Pushary read that path as a Linux path, so a rule that refuses one file, such asEdit(src/secret.ts), did not match it, and under an allow for edits the edit ran. A patch that reached a file outside the project, such as..\.bashrcorC:\Users\you\.gitconfig, was not asked about under an allow either. Pushary now reads the paths of a project whose folder starts with a drive letter or\\the way Windows does, so that rule refuses the edit and those files ask first. The Pushary CLI on Windows already read paths this way, and on macOS and Linux nothing changes. - A long command Pushary could not read in full is marked. A command over 4,000 characters reaches Pushary cut short, so a step at its end, such as
rm -rf ./buildafter a long heredoc, was not seen and the request read as ordinary. A command or patch that arrives cut now shows as risky, with "Pushary could not check this step". On Pushary app 1.0.66 or later its card asks for Face ID, Touch ID or the passcode before Approve, and its notification offers Deny only. A long file write, edit or plan reads as before. Through the Server API, Pushary reads the whole command you send, up to 16,000 characters, and a step it finds there is flagged by name. Older apps, Slack and calls still offer Approve on these steps, as on the steps listed as not covered yet below.
Fixed
- A protected file reaches your phone again in a paused session. Since November 20, when phone approvals for a session were paused after three unanswered requests, or your phone had left 20 requests unanswered over two weeks, an edit to a file Pushary always asks about, such as
.env,CLAUDE.mdor a CI workflow, under a rule that approves after a wait was held back with no notification. Pushary sends it again, as it did before November 20. What happens if nobody answers depends on your CLI. With Pushary CLI 1.9.219 or later, the Claude Code and Codex hooks keep waiting and never approve the file on silence. With 1.9.218 or older, the hook approves it once the rule's wait ends, as it did before November 20; between November 20 and 21, Claude Code on 1.9.211 to 1.9.218 and Codex on 1.9.204 to 1.9.218 refused it instead, and older hooks wrote the file at once with nobody asked. Runnpx -y pushary@latest upgradeso a missing answer never approves such a file.
Approve on a phone with no screen lock, November 21, 2026
Versions covered by this entry:
- Pushary app for iPhone and Android: the app's next release, once the App Store and Google Play approve it
Fixed
- Approve works on a phone with no screen lock. With Require unlock to approve on, which is the default on Android, an Android phone with a fingerprint sensor and no screen lock did nothing when you tapped Approve on an ordinary step in the app, and Approve on its notification opened the card instead. On an iPhone with Face ID and no passcode, only Approve in the app was affected; Approve on the notification already worked. A phone with no screen lock has nothing to unlock, so Approve now sends at once, as it already did on a phone without a sensor. A phone with a screen lock still asks for it first.
- A risky step still needs a screen lock. A phone with no screen lock can deny a risky step but not approve it.
- Start session and Send say why they cannot. Starting an agent and sending an instruction still need a screen lock. On such a phone, the Start session button and the send button in a session now show Device lock required where they did nothing.
Sign-up in your language, November 21, 2026
Covers the sign-up page. No CLI, app or Pushary for Mac update is needed.
Fixed
- The agent and Partner sign-up pages were in English in every language. Since November 21,
/de/sign-up,/fr/sign-up,/nl/sign-up,/tr/sign-upand/ar/sign-upshow the heading, the line under it and the panel beside the form in that language. On the agent sign-up that includes the question on the lock screen, its time, Approve and Deny. Pushary and agent names such as Claude Code stay as they are.
Risky steps from Pushary for Mac and editors leave the lock screen, November 20, 2026
Versions covered by this entry:
- Pushary server: no app or CLI update needed. It applies to phones on Pushary app 1.0.66 or later.
Changed
- Deny only on the lock screen. The steps the entry below lists as not covered yet now reach a phone on Pushary app 1.0.66 or later with Deny and Deny with note on the lock screen. That covers a risky step that Pushary for Mac or an editor plugin forwards, such as
rm -rforgit push --force, and a risky blocked action or Codex auto-reviewer rescue. On iPhone the reason shows under the title, such as "Risky · Deletes files". Tap the notification to open the card, where Approve asks for Face ID, Touch ID, your fingerprint or your phone's passcode first. A phone with no screen lock can deny these steps but not approve them; approve at your computer instead. - Older apps keep their lock-screen Approve. A phone on 1.0.65 or older, or one that has not opened the app since this change, gets these steps as before, because its card cannot approve them behind the device check. A phone names its app version each time the app opens.
- Risky steps need the keyboard still does not hold these steps.
Approve risky steps on the phone, November 20, 2026
Versions covered by this entry:
- Pushary app for iPhone and Android: 1.0.66, once the App Store and Google Play approve it
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.215
Changed
- Approve a risky step from the phone. In Pushary app 1.0.66, a risky step such as
rm -rf,git push --forceor a deploy opens as a card with a red Risky badge naming the first reason, the other reasons and the command, with Show all for a long one. Approve asks for Face ID, Touch ID, your fingerprint or your phone's passcode, and sends only after it succeeds. A step that does not match what you asked shows what you asked above the step, and Approve opens a confirm before the device check. Handle on your machine hands the step back to the agent's own prompt. A risky step's card offers no session grant, standing rule or Never stop choice. - A phone with no screen lock cannot approve a risky step. The card says to add a passcode or approve at your computer, and sends nothing. A prompt that fails or is locked out after too many tries says so. Deny still works.
- A risky approve that does not go through is not retried for you. If it never reached Pushary, the card says Not sent. Tap Approve again., and the next tap asks for Face ID again. If Pushary may have recorded it, the card says so and the list refreshes. If Pushary refuses it, for example because your team turned on Risky steps need the keyboard after the request arrived, an alert says why, and a step your team keeps at the computer then shows Deny and Handle on your machine only.
- Risky steps need the keyboard, on the phone. With the setting on, a risky step from the Pushary CLI's hooks, MCP or the Server API shows the risk with Deny and Handle on your machine, and says "Your team approves risky steps at the computer."
- The lock screen says why a step is risky. On iPhone, the notification for a risky step from the Pushary CLI's hooks, MCP or the Server API, which has no Approve action, names the reason under its title, such as "Risky · Deletes files". The reason shows without an app update. Approving the step from the phone needs Pushary app 1.0.66, where the card asks for Face ID or the passcode first.
- Not covered yet. Risky steps that Pushary for Mac or an editor plugin forwards, the Codex auto-reviewer's rescue and blocked actions keep their lock-screen Approve, get no reason under the title, and are not held by Risky steps need the keyboard. In Pushary app 1.0.66, their card still asks for Face ID or the passcode before it approves.
- The Pushary skill 0.13.6, in CLI 1.9.215 and the Claude Code, Codex, Cursor, VS Code and Grok Build plugins, tells agents the phone can approve a risky step behind the device check.
Older versions of the app show a risky step as they did before.
Risky approvals, November 20, 2026
Covers the dashboard, /decide, /notify, the live page, the browser inbox and Slack. No CLI, app or Pushary for Mac update is needed.
Security
- Slack, a one-tap link and the Partner relay could approve a step the phone could not. A risky step, such as
rm -rf,git push --forceor a step that does not match what you asked, shows no Approve on the Pushary app or the lock screen. Until November 20, Approve on its Slack message, the yes link in a notification and a Partner's server relaying an answer could still approve it. Since November 20, the server refuses an approve on such a step unless the place it came from showed the warning and you confirmed it: the confirm step on the web, or Pushary for Mac on the computer the agent runs on. A refused approve is not recorded, so you can still deny it or approve it elsewhere. - Slack shows the risk instead of Approve. A risky step's Slack message says why it is risky and shows the full command, with Deny and Review and approve, which opens the request's decision page, where you confirm it. An Approve on a message posted before the change is refused, and the message links to the same page. A yes link in a notification for such a step opens that page too.
Changed
- A confirm step before a risky approve on the web. Since November 20, the dashboard,
/decide,/notify, the live page and the browser inbox show a risky step with a red badge, its reasons and the full command. Approve opens a short confirm that restates the step, and for a step that does not match your request, what you asked beside it. Handle on your machine hands any gated request back to the agent's own prompt. - Risky steps need the keyboard, a team setting. A workspace owner or admin can turn it on in Agent settings, on every agent plan. It is off by default. With it on, a risky step from an agent connected through the CLI's hooks, MCP or the Server API can be approved only at the agent's own prompt or in Pushary for Mac on the computer the agent runs on. The phone, the web and Slack can still deny it: the web shows Deny and Handle on your machine with no Approve, Slack shows Deny and Open, and a call says you'll need to approve it at your computer. A request already waiting when you turn it on still shows Approve, and that approve is refused with a message saying why. An agent with no prompt of its own, and a Server API question, cannot get a risky step approved while it is on. Steps Pushary for Mac forwards are covered in a later release. Permission policies has the details.
- The audit trail says how a risky step was approved. Opening a row in Activity shows, for example, "Approved on the web after a second confirm. Risky: Deletes files." The CSV and JSON exports gain
riskConfirmationandriskReasonsat the end. See How a risky step was approved.
Phone calls, November 18, 2026
Versions covered by this entry:
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.142 to 1.9.184 - Pushary app for iPhone and Android: 1.0.61
Pushary can call your phone when an agent waits on a request you did not answer. The push still comes first. If the request is still waiting after Call after, Pushary calls you, reads the request and takes your answer. Calls shipped as a beta on November 12 and left beta on November 15. The Phone calls guide covers setup, the agents that can be called, the countries Pushary can call and every limit.
Security
- A call could approve a risky step. Until November 15, a request that reached Pushary through Pushary for Mac, or as an action Codex's auto-reviewer or the CLI had blocked, carried no risk check for calls. A call about
rm -rf,git push --forceorsudocould approve it after the read-back. Since November 15, a risky step asked that way can be denied on a call but needs the Pushary app to approve, as a step asked by the CLI's hooks already did.
New
- A call when the push did not reach you. Pick Call after: 1, 2 or 3 minutes. The default is 2. Call sooner when an agent waits less is on by default, so an agent that stops waiting sooner is called as early as it needs. On the call, Pushary says which agent is asking, in which project, what it wants to run and how long it has waited. You can approve it, say no, or say what the agent should do instead. Your activity record marks the answer by phone call, and Pushary app 1.0.61 shows it in your history.
- A yes is read back first. Before a yes counts, Pushary reads the step back and asks again. Some requests can only be approved in the Pushary app. On those, the call can still say no or pass on an instruction.
- One call, up to three requests. If more than one request waits for you, one call asks up to three of them.
- A missed call rings once more ten minutes later, if the request is still waiting then.
- Limits you set. Calls stay inside your calling hours, 08:00 to 22:00 in your time zone by default, with a Weekdays only option. The default is at most 5 calls in 24 hours. You can set 1 to 10, or no limit. Calls are in English.
- Call minutes. Agent includes 10 call minutes a month, Agent Pro 20 and Team 60, shared by the workspace. The 3-day trial includes 5. You can buy more as a one-time purchase. When the minutes run out, Pushary stops calling and push notifications keep working.
- The Calls page. Since November 17, Calls is its own page in the dashboard sidebar. It holds your number, the test call, your rules, your minutes and Call history, with a transcript of each call kept for 30 days. Links to the old Phone calls tab in Settings open it.
pushary calls. With CLI 1.9.161,npx -y pushary@latest calls onasks for your number and first name, places the test call, waits for you to confirm the number, asks you to agree to the calls and turns them on.pushary callsshows whether they are on, andpushary calls offstops them. CLI reference- A notice when a key changes your number. When an API key sets a new number or turns calls on, Pushary sends a push to your paired phone and an email, both showing the number in full. Keys can make five such changes for you in 24 hours, counted across all your keys. After that, a key's change is refused, and you set your number or turn calls on in the dashboard. Since CLI 1.9.184,
pushary calls onsays so when it is refused for that reason. - "Call me if you get stuck." Since CLI 1.9.144, the instructions and the skill Pushary installs tell an agent what to do when you ask it to call you: ask through
ask_user. The reply says whether a call is planned, and the agent is told not to promise one when none is. - Team. A routed approval calls the person the route names. From CLI 1.9.142, a request from an agent with a prompt of its own, such as Claude Code, waits long enough for that call, even when the person whose agent asked has calls off. With a When away rule, the default, Antigravity CLI and Codex when it has no prompt of its own wait only as long as the rule of the person whose agent asked, 60 seconds by default, so the call to the teammate is skipped when that wait ends first. With an Always rule the wait is raised for them too, up to the time the agent gives a hook.
- Antigravity CLI. Its approvals can ring since November 16.
Calls are set up in the dashboard or from the terminal. They are not offered for OpenClaw yet.
OpenCode questions, Codex in Always and more agents, November 18, 2026
Versions covered by this entry, November 5 to November 18:
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.101 to 1.9.184 - Pushary for Mac: 0.1.128 to 0.1.137
- Pushary app for iPhone and Android: 1.0.54 to 1.0.62. 1.0.62 is on Android now, and on iPhone once Apple approves the update. Until then the App Store has 1.0.61.
- OpenClaw plugin (
@pushary/openclaw): 0.1.7 - Hermes plugin (
hermes-plugin-pushary): 1.0.2
Every CLI release except 1.9.153 has its own notes in the CLI changelog, and every Pushary for Mac release has notes on its releases page. Update the CLI with npx -y pushary@latest upgrade.
Security
- A folder rule covered a path that left the folder. With
Write(/Users/you/**)set to approve, a write to/Users/you/../outside.txtran without a question, because the rule was checked against the path as written. Since CLI 1.9.139 (November 10), rules see the path with.and..worked out, and a deny on a folder catches a path that reaches it through... Symbolic links are not followed. The server reads paths the same way for approval routes and for the gate Pushary for Mac and the editor plugins use. - CLI 1.9.140 let some commands past a deny rule. For about 16 hours on November 11, with
Bash(git push:*)set to deny beside a broad allow, a push was allowed when it sat inside somebash -cscripts, behindfind -exec, or after a leading input redirection. CLI 1.9.141 fixed it the same day. - Shell forms that hid a command from your rules. Before CLI 1.9.119 (November 8), a quote or backslash inside a shell comment could hide the lines after it from deny rules. A script fed to a shell by process substitution, such as
bash <(curl ...), ran under a broad allow. Since CLI 1.9.111, a deny, ask or route rule on a command also catches it behind a runner such asuv run,npm exec,bundle execordocker exec, soBash(git push:*)asks aboutuv run git push --force. - Another MCP server could pass as Pushary's own. Any tool whose name started with
mcp__pushary__counted as one of Pushary's own tools, so a different server namedpushary__fsskipped the kill switch and standing deny rules. With CLI 1.9.150 and 1.9.151, Claude Code also allowed that server's tools without a prompt. Since CLI 1.9.152 (November 15), only the server named exactlypushary, the claude.ai Pushary connector and the Pushary plugin count. - Silence could approve a call outside an agreed scope. Before CLI 1.9.118 (November 8), an Always allow rule on Claude Code withdrew a scope question at once and ran the call, and a standing auto-approve on Gemini CLI or OpenCode ran a call outside the scope. A scope question now waits for a real answer. On Gemini CLI, OpenCode and Cursor, silence refuses the call.
- Antigravity CLI ran workflow scripts and file writes without a question. Antigravity CLI 1.5.9 added
run_workflow, which runs a Python script the agent writes, and the Pushary hook did not cover it. Since CLI 1.9.155 and Pushary for Mac 0.1.135, the script is judged aspython3 -c '<script>'under yourBashrules. Antigravity CLI builds that sendwrite_fileoredit_fileare gated since CLI 1.9.153 and Pushary for Mac 0.1.134. Runnpx -y pushary@latest setup --agents antigravity_cliagain, or repair the Antigravity connection in Pushary for Mac, then start a new AGY session. - The OpenClaw plugin sent some
.envvalues and missed some writes. Before plugin 0.1.4, a quoted.envvalue that ran over several lines had its later lines sent with the approval. Plugin 0.1.5 closed further.envleaks in patches and odd paths. It shipped only inside CLI 1.9.165 and was not published to npm, where the plugin went from 0.1.4 to 0.1.6. Plugins 0.1.5 to 0.1.7 make a write ask when it reaches OpenClaw's own config through a symbolic link, a hard link, a path that is not plain, or on macOS a path under/System/Volumes/Data. CLI 1.9.169 carries plugin 0.1.7. Update withnpx -y pushary@latest upgrade, oropenclaw plugins update pusharyfor an install from ClawHub or npm. - One phone could read every member's session history. Before November 7, on a workspace with more than one member, session history was sealed to the one phone that registered the workspace's history key, and the server gave any signed-in member the records of any session. Now a session's history is shown only to the person signed in on the computer that ran it.
- The kill switch and deny rules cover more of Claude Code. Since CLI 1.9.120 and Pushary for Mac 0.1.130, the kill switch and standing deny rules apply to Claude Code's MCP tools, web fetches and subagent tasks in every permission mode, even when Claude Code shows no prompt. Nothing new is asked on those calls.
- Setup no longer prints your API key. Before CLI 1.9.148, setup printed the full key when your shell profile already set
PUSHARY_API_KEYin a line Pushary did not write. Since CLI 1.9.154, setup, npm and OpenClaw failure messages redact credentials too. - The Hermes history helper got keys it did not need. Since Hermes plugin 1.0.1 (CLI 1.9.103, Pushary for Mac 0.1.129), the helper that syncs session history receives only what it reads. Hermes' other provider keys no longer reach it.
New
- OpenCode's questions on your phone. Since CLI 1.9.161, a question OpenCode asks with its own question tool shows in OpenCode and on your phone and the Mac notch at the same time. The first answer wins. Since CLI 1.9.163, a question with five or six options shows on the phone as a plain list, a call with two to four questions is asked one question at a time, and a question the phone cannot show sends a notification telling you to answer it in OpenCode. With Pushary for Mac 0.1.136, OpenCode asks through Pushary's
ask_userinstead. Supported agents - Start OpenCode from your phone. With CLI 1.9.178 and Pushary app 1.0.62, which is on Android now and on iPhone once Apple approves the update, Fleet, then New session, offers OpenCode on a computer that runs
pushary daemonand has OpenCode 2. The session opens in OpenCode's own background service, so it shows in the OpenCode desktop app too. A phone-started OpenCode session cannot be resumed from the phone yet. On a computer where OpenCode is the only agent, runnpx -y pushary@latest setupagain to turn this on. - More of OpenCode 2 reaches your phone. Since CLI 1.9.161 and Pushary for Mac 0.1.136, web searches, skills and subagents OpenCode asks about go to your phone, a deny reaches the agent with its reason so it carries on, and a turn that ends on an error sends a push. A subagent's prompts are no longer reported as yours.
- OpenClaw.
npx -y pushary@latest setup --agents openclawinstalls a Pushary plugin that asks you before everyexeccall and before the tools an agent could use to switch the plugin off. It needs CLI 1.9.147 and OpenClaw 2026.11.3 or later. The plugin is also published as@pushary/openclawon ClawHub and npm. OpenClaw setup - A Grok Build plugin.
grok plugin install Pushary/grok-build-pluginlets Grok Build ask you questions and send a push when work finishes, signed in through your browser and with no hooks on your computer. Approving tool calls from your phone still needsnpx -y pushary@latest setup --agents grok. Grok Build - Codex permission requests. Since CLI 1.9.140, when Codex asks for a permission profile, such as network access or folders it may read or write, the request goes to your phone. You can allow it for the turn, allow it for the session, or deny it.
- Hermes Desktop. With Hermes plugin 1.0.0 (CLI 1.9.102, Pushary for Mac 0.1.128), a Pushary chip in Hermes Desktop's status bar shows the approvals waiting on your phone, and you can answer them there. Only one answer counts.
- An audit trail you can check yourself. On Agent Pro and above, Verifiable export in the ⋯ (More) menu of Activity downloads a signed hash chain of your questions and decisions. It shows once your site's first entry has been recorded. With CLI 1.9.104,
npx -y pushary@latest audit verify <file>checks every hash and signature with no account and no network. Check it yourself pushary sessions --jsonprints your agent sessions and the questions waiting on them as one JSON object. CLI 1.9.170. CLI reference- Yearly billing. Agent and Agent Pro can be paid yearly: $99 a year for Agent and $199 a year for Agent Pro, which is two months free. Monthly stays preselected at checkout.
- Agents on your iPhone lock screen. Since Pushary app 1.0.57, a Live Activity on the lock screen and in the Dynamic Island shows which agent needs you, for how long, and how many are running. Turn it on in Settings, under Notifications.
- Team: route a set of risky steps in one click. Since November 8, an owner or admin can add a pack of approval routes in Policies: Risky steps, Money and customer actions or Installs and agent settings. Overview offers Send them to me for the risky steps. Team guide
- Activity is one filterable record. Since November 8, Activity has Agent runs, Questions and What changed, with filters and paging.
Changed
- Claude Code: a plan approved from your phone starts the work. Before CLI 1.9.175, a plan you approved on the phone or the Mac notch still waited for a second yes at the terminal. Pushary now asks about the plan when Claude Code shows Would you like to proceed? and answers it there. Approve, start coding starts coding in the mode the session had before planning, and Keep planning keeps Claude planning. While Pushary for Mac is running, this works without the CLI update, and since 0.1.137 a plan you answer at the keyboard leaves the notch and your phone. On CLI 1.9.175 to 1.9.181, a wait for your phone that ended without an answer, or a Stop from your phone, declined the plan. Since CLI 1.9.182 the plan dialog stays up in Claude Code for you to answer there. Claude Code
- Claude Code: Claude's questions show in both places. Since CLI 1.9.176, a question Claude asks you shows in Claude Code and on your phone at the same time, and the first answer counts. Before, only the phone could answer it. A question asked through Pushary for Mac is still held for the phone or the notch. Since CLI 1.9.174, a multiple-choice answer with several picks reaches Claude the way Claude Code writes it.
- Codex in Always: a burst waits for a free place. Codex can send many calls at once, and a session can have 8 open questions that push to your phone. Before CLI 1.9.162, every call past the eighth was refused at once. Now a call that arrives when the session already has its eight open questions waits for a free place. How long it can wait depends on your rule's wait time. If no place frees in that time, the call is refused, never approved. With CLI 1.9.179, a call past the limit whose rule would not approve it unanswered is stored without a push instead, listed in the app, and waited on like any other question. Pricing and limits
- Codex in Always: one tap clears a burst. Since November 17, once three calls of the same tool wait in a session at once, their cards offer the session grant, such as Allow Bash for this session, in the Pushary app and Pushary for Mac. With CLI 1.9.162, the calls still waiting read your rules again, run, and take their questions off your phone. The grant covers
Bashcommands that are not risky and MCP tools that only read. It never covers an edit. - Codex in Always: a late answer counts. Before CLI 1.9.179, a call nobody answered within the hook's time, about 170 seconds, was refused and its question withdrawn. Now the question stays on your phone and Codex is told to run the same call again. When it does, in the same turn, it joins that question instead of asking again, and your answer applies once. A retry that asks for escalated permissions, or, after you press Esc, one made less than 180 seconds after the hook began waiting on that question, gets a card of its own. This depends on the model repeating the call. Codex
- Codex: pressing Esc no longer blocks a session's questions. Codex stops a hook when you press Esc, so its question could not be withdrawn and kept its place for ten minutes. With eight of them, every gated call in that session was refused. Since November 17, such a question stops counting once its hook can no longer wait on it.
- Fewer rings for routine approvals nobody answers. A question whose rule approves on timeout rings every time. Since November 17, once ten of them in a row have gone unanswered, a routine command or edit rings at most once in ten minutes and arrives silently otherwise. The question itself is unchanged, and any answer ends the quiet stretch.
- When away is offered, never switched for you. Since November 17, after you answer a question in the dashboard while in Always, Overview offers Switch to When away. With CLI 1.9.179, Pushary app 1.0.62 offers the same during a Codex burst while you are at your computer, and Not now hides it for a week. 1.0.62 is on Android now, and on iPhone once Apple approves the update.
- Your API key moved out of your shell profile. Since CLI 1.9.171, setup and
pushary loginsave the key in~/.pushary/env, readable only by you, and add one line to your profile that loads it. A key export an earlier version wrote there is replaced by that line the next time you run setup orpushary login. A profile kept in a dotfiles repository no longer receives a copy of the key. Get your API key - Setup asks which agents to set up. Since CLI 1.9.174, setup at a terminal asks Which agents do you use? after you pair the phone app too, with the agents it found already ticked. Since CLI 1.9.147, setup and
pushary pairlead to the Pushary app, and push notifications in a browser need--connect weborpushary pair --browser. - Claude Code's own worktrees carry the repository's name. Since CLI 1.9.174, a session in one of Claude Code's own worktrees, under
.claude/worktrees, is named after its repository, not the worktree folder. CLI 1.9.174 to 1.9.182 gave that name to every git worktree, so two worktrees of one repository showed the same name. Since CLI 1.9.183, a worktree you made yourself, or one a tool such as Conductor made, is named after its folder again, so parallel worktrees can be told apart on your phone. Sessions whose hooks Pushary for Mac runs still use the folder name. - Grok Build's read-only tools stay off your phone. Since CLI 1.9.167 and Pushary for Mac 0.1.137, Pushary never asks about
grep,list_dir,glob,read_file,todo_writeorsearch_tool, and Grok'sread_fileworks again. Write a rule for one of them with Claude Code's name, such asReadorGrep. - Qwen Code runs the command on a yes. Before CLI 1.9.136, a yes on your phone still left the command waiting at Qwen's own prompt.
- A hook keeps waiting through a network blip. Since CLI 1.9.143, one failed poll no longer withdraws the question a hook is waiting on. The hook keeps polling until its own deadline, and a wait that saw a failed poll never ends in a timeout approval.
- Team routes reach further. In Claude Code's auto mode, a shell command, file edit or MCP tool call that a route sends to someone else now waits for that person (CLI 1.9.113 and 1.9.116). A route on pushes to
mainmatches a plaingit pushon that branch (CLI 1.9.116), reads a push by where it lands, such asgit push origin HEAD:refs/heads/main(CLI 1.9.119), and applies to the Codex app permission cards Pushary forwards (CLI 1.9.109). A rule that names someone who left the Team goes to its owners and admins. - Agent Pro is for one person. New Agent Pro subscriptions from November 17 cover one member. Workspaces on Agent Pro before then keep five. Plans and limits
- The pricing page shows the price checkout starts from. Since November 17, visitors outside the US, Britain and the euro area see the euro price. For Agent and Agent Pro, checkout converts it to their currency. Before, they saw dollars.
- You stay signed in. Pushary app 1.0.60 and Pushary for Mac 0.1.133 renew your sign-in while you use them, so they no longer sign you out every seven days.
- Pushary app. 1.0.62 leads each inbox card with its project and shows the agent as a badge, and its approval cards show a command once. It is on Android now, and on iPhone once Apple approves the update. 1.0.59 opens a Codex permission request with its details showing. 1.0.55 puts your choices first in Questions & approvals.
- Pushary for Mac. A deny with a note from the notch reaches the agent with your note (0.1.131). Home shows a Repair button for an agent whose connection is missing or damaged, and a card says when an answer could not be sent (0.1.133). Claude Code is offered for a start from your phone when it is installed with volta, fnm, asdf, mise, pnpm or nix (0.1.132).
- The background service on Linux and Windows. On a Linux machine you reach over SSH, starting sessions from your phone keeps working after you log out (CLI 1.9.134), and sessions started from your phone keep running when the daemon restarts (CLI 1.9.125). On Windows the daemon runs without a console window (CLI 1.9.138).
- Hermes in a container. Hermes skips its own dangerous-command approvals when its terminal backend is a container, and on
dockeronly when no host folder is mounted, so no shell command reached your phone. Since CLI 1.9.145,pushary doctorsays so and setup offers to gate the terminal tool. Hermes
Fixed
- A failed install blocked every later setup and update. An install lock left behind made setup,
pushary upgradeand the daily update fail with "An earlier installation did not finish safely" until you deleted~/.pushary/auto-update.lock. CLI 1.9.101 and 1.9.148 fixed some causes, 1.9.154 brought the stuck lock back, and 1.9.161 stopped an npm release-age refusal from leaving one. Since CLI 1.9.164, on macOS and Linux, setup and upgrade, and automatic updates on macOS, clear a lock that is more than an hour old once the process that started the install has exited and no npm install of Pushary is running. A computer that CLI 1.9.154 to 1.9.161 left stuck is not fixed by its daily update, which still runs the older release. It clears the first time CLI 1.9.164 or later runs setup or upgrade there, for examplenpx -y pushary@latest setup. On Windows you still remove it yourself. Troubleshooting - OpenCode on Windows left answered questions on your phone. Enough of them used up the open question limits, and after that every question on the workspace was refused. CLI 1.9.168 withdraws them. Run
npx -y pushary@latest upgradeto rewrite the plugin. - OpenCode 1 lost a phone answer when the request that hands it to OpenCode failed. Since CLI 1.9.180 and Pushary for Mac 0.1.137, your phone tells you to answer it in OpenCode. On a computer that is already connected, rewrite the plugin first: run
npx -y pushary@latest upgrade, or repair the OpenCode connection in Pushary for Mac and start a new OpenCode session. pushary cleanremoved other tools' lines from your shell profile. On CLI 1.9.171 to 1.9.180,pushary cleanremoved every profile line shaped like the one that loads your Pushary key file,if [ -f FILE ]; then . FILE; fi. The Google Cloud SDK adds itself to a profile with lines of that shape. CLI 1.9.181 removes only the line setup wrote. If you use the Google Cloud SDK and ranpushary cleanon one of those versions, run itsinstall.shagain to put its lines back.- Claude Code asked again for commands you had allowed. Setup, upgrade and the session start repair deleted every allow rule of yours that contained the word "pushary", such as
Bash(pushary doctor). Fixed in CLI 1.9.130. - Setup said "You're set" when Codex could not gate a call. Since CLI 1.9.112, setup ends with Codex approvals are not on yet or Codex is not running every Pushary hook yet when that is the case.
- Setup's test question was cancelled early. A tap between 48 and 55 seconds was reported as "nothing came back". Fixed in CLI 1.9.106.
- The background service restarted forever after
npm uninstall -g pusharyon macOS and Linux, writing an error every 10 seconds. Fixed in CLI 1.9.137. The daemon also keeps running after your computer wakes from sleep (CLI 1.9.123) and starts again after a lock file was cut off by a full disk or a power cut (CLI 1.9.133). - Windows. The Cursor gate runs on Windows (CLI 1.9.124), a message from your phone reaches a phone-started Gemini session as you wrote it (CLI 1.9.128), and your phone shows Claude's reply as it streams (CLI 1.9.132).
- Cursor and VS Code showed one computer as two. Their gates now use the computer's saved id. CLI 1.9.120.
- Answers from a locked iPhone. Pushary app 1.0.57 makes Approve and Deny from the lock screen reliable, including when the app was closed. Since 1.0.55, an answer that could not be sent is kept and tried again for up to five minutes while the app is open.
- Starting an agent from your phone started it twice. A retry after a dropped connection, or a second tap on Start, could start a second agent. Fixed in Pushary app 1.0.57, 1.0.59 and 1.0.62, CLI 1.9.140 and Pushary for Mac 0.1.132. 1.0.62 is on Android now, and on iPhone once Apple approves the update.
- A notification for an older session said it was gone. Pushary app 1.0.61 opens a session that has been quiet for more than a day.
- Pushary for Mac warned that Codex was not connected while it was. Fixed in 0.1.132. Since 0.1.130 the app no longer checks for sessions to wake when there are none.
Corrections
- CLI 1.9.115, 1.9.146, 1.9.159, 1.9.160, 1.9.172 and 1.9.173 have entries in the CLI changelog but were never published to npm. Their changes first shipped in 1.9.116, 1.9.147, 1.9.161 and 1.9.174. Each entry now says so, as do the entries for 1.9.95 and 1.9.98.
- CLI 1.9.153 was published on November 15 without an entry in the CLI changelog. It gates the
write_fileandedit_filetools of current Antigravity CLI builds. - The troubleshooting page said CLI 1.9.159 checks with npm before installing. That check first shipped in 1.9.161.
- The SDK versions below listed CLI 1.9.100. They now list 1.9.184.
Pushary for Teams, November 4, 2026
Versions covered by this entry:
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.100 - Pushary for Mac: 0.1.127
- Pushary app for iPhone and Android: 1.0.52 (1.0.53 rolling out)
Team is for a group of people who each run AI agents, such as Claude Code, Codex or Cursor. Each agent asks the person who runs it. A rule can send a risky step to a lead instead. Everyone sees the same record of whose agent asked and who answered in Pushary. Team shipped on October 29 and changed through November 4. The Team guide covers setup, roles, seats and limits, Pushary for teams has the overview, and How Pushary for Teams decides who an agent asks explains the rules.
New
- Price. $59 a month for 5 people, plus $12 a month for each extra person. Team is sold in US dollars, euros and pounds at the same numbers. New workspaces get 3 days free. Team subscriptions started at an earlier price keep it. Plans and limits
- Each question goes to the person whose agent asked. Once a second person joins, an approval goes to that person's phone and dashboard. Until then, questions also go to push notifications in the browser, and to Slack if you connected it. Task updates an agent sends still reach every member's phone.
- Send a risky step to a lead. Team includes approval routing. An owner or admin adds a rule in Policies with a tool pattern such as
Bash(git push:*), one repository or all of them, and a role or one person. A routed approval reaches that person in the phone app and the dashboard. If they have not answered when the wait ends, the step goes back to the agent, where the person whose agent asked can answer at its own prompt, and that answer is not recorded. The phone app can deny a force push,rm -ror a deploy, but not approve one. Approve those in the dashboard or at the agent's prompt. - One screen to start. Starting a Team trial is one screen: check the team name, then check out.
- Overview is the Team home. Its People block is the setup list, with one row for each person: connect a computer, pair a phone, invite the team.
- Invitations and a team link. Invite a teammate by email from People. A new account can accept an invitation, and someone who signs up or signs in from it joins on arrival. Signed in with another address? The invitation page says which address it was sent to. If the owner's email is at a company domain, an owner or admin can turn on a team link for anyone with a verified email at that domain. It is not offered for free email domains.
- Seats that grow. 5 seats are included. When someone joins and every seat is taken, Pushary adds a seat at $12 a month, prorated on the next invoice, and emails the owner and admins. When someone leaves, their seat is dropped, never below 5. An invitation holds no seat and expires after 7 days. A Team on the older 10-person plan, or paid through the App Store or Google Play, keeps a fixed number of seats.
- Team activity. Activity has Sessions, Audit trail and What changed. Sessions shows one row for each session, with whose agent it was, the commands, file edits and commits the agent reported, and its questions. Worth a look sits on top. Choose this sprint or the last 7 or 30 days, and everyone, yourself or one person. Copy for retro, Export CSV and Export JSON are in the ⋯ menu. Exports name people by user id. Nobody is scored or ranked. Team keeps 365 days of history. See what happened
- What the record does not show. It does not show who said yes at the agent's own prompt. In Team activity, only answers in the Pushary phone app, Mac app or dashboard carry a name. Answers from the web inbox or a link in a browser notification carry none, and the Audit trail shows a Slack answer under the Slack member's name. Steps your rules approve on their own have no answer. A command shows only its first two or three words, such as
git push. Each person's record starts when their agent first connects. What the record does not show - The sprint email. The owner and admins get the sprint record by email on the sprint's last day. New Teams get it instead of the weekly digest. It is not sent during the trial.
- Leads hear when a teammate joins or first answers. Owners and admins get one push when a teammate joins, unless an email already told them, and when a member first answers a question from their own agent.
- A Team welcome email replaces the Agent welcome after a Team checkout and links to inviting people.
Changed
- Only owners and admins change team rules. In Team and Enterprise workspaces, policies, approval routing rules, presets and team defaults need an owner or admin. Members see team rules read only and still answer their own questions. Viewers see Policies and Overview read only.
- A member's Always allow covers only their own agents. It saves a personal rule. Two members can each keep one for the same tool, and a team rule for the same tool wins. An owner's or admin's Always allow is a team rule and covers everyone's agents. From Pushary app 1.0.52, members do not see Always allow on a teammate's request.
- Each person has their own delivery settings. Approval mode, temporary mode, task updates and When your agent skips permissions cover only that person's agents. The push window and the
*timeout stay team rules. Owners and admins can set a team default for everyone who has not picked their own. New workspaces start on Always. - Claude.ai and ChatGPT connections start on your Team. A person's first connector connection is made on their Team's site instead of their newest site. A Team whose subscription has lapsed is skipped.
- Pushary for Mac. 0.1.124 lets you choose the Team workspace when you connect your Mac, and opens Team activity. 0.1.126 shows an invitation with Accept instead of the plans when you have no plan, and each person pairs their own phone.
- Pushary app. 1.0.52 says when a Team setting covers only your agents and marks the team default.
Fixed
- Slack could not be connected on Team, Partner and Enterprise. Since July 13, the channel install asked Slack for a set of permissions Slack refuses, so every connect on those plans stopped at Slack's consent page. Fixed on October 25.
- A Team moved to Agent or Agent Pro was billed for its seats. A plan change kept the seat count, so five or more seats were billed at the new plan's price. Since October 31 a plan change bills one unit.
Permissions off, session history and more agents, November 4, 2026
Versions covered by this entry, October 16 to November 4:
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.61 to 1.9.100 - Pushary for Mac: 0.1.104 to 0.1.127
- Pushary app for iPhone and Android: up to 1.0.52 (1.0.53 rolling out)
- Server SDK (
@pushary/server): 2.5.1 - Python SDK (
pushary): 2.4.1 - Hermes plugin (
hermes-plugin-pushary): 0.9.2
Every CLI release has its own notes in the CLI changelog, and every Pushary for Mac release has notes on its releases page. Server SDK 2.5.1 and Python SDK 2.4.1 change only the package description and README.
Security
- Pushary for Mac could grant a standing permission from a key pressed in another app. From October 24 to October 26, Pushary for Mac 0.1.118 to 0.1.120 let Control-Command and a number pick an option on a question from any app. A Codex or Hermes tool approval offers "Allow this conversation" or "Always allow", so a key pressed while you typed elsewhere could grant one. Since 0.1.121, a tool approval takes an option key only after you click the notch. The approve and deny keys, and number keys for your agent's own questions, work as before.
- Some Codex calls ran without a question on October 31. CLI 1.9.94 and 1.9.96, and the server's gate, which Pushary for Mac and the editor plugins use, applied the new permissions-off checks to every agent that reports
bypassPermissions.codex execreports that mode whatever its approval policy is, so a routine Codex call that used to reach your phone ran without a question. Risky steps still asked. CLI 1.9.97 and the server fixed it the same day. Update withnpx -y pushary@latest upgrade. - Some setup keys had full access. Before October 18, the key that Resume setup in onboarding and Email me the command created was a full-access key instead of an agent key. An agent key cannot switch phone approvals off. Both now create agent keys.
- The emailed setup command carried an API key. Until October 24, Email me the command put a new API key in the email. It now sends the setup link and the plain
npx -y pushary@latest setupcommand, and you connect with the Pushary app's QR code. - A departing member's keys kept working. Before October 19, removing a member, or a member leaving, turned off their agent keys but not the other keys they had created on the workspace's agent sites. Now those are turned off too, except keys bound to one of your customers. Leaving now asks you to confirm.
- A phone's history key could be replaced without notice. Since October 25, the server refuses to register a different history encryption key over the one a phone already holds, so a second phone or a reinstall can no longer silently take its place. CLI 1.9.90 also stops Gemini CLI, Cursor and Hermes history from trusting a new phone key on first use when the computer already trusts a phone. To move history to a new phone, the owner chooses Use this phone for history in the app's Settings, then verifies it on each computer with
pushary transcripts trust --fingerprint <fingerprint-from-phone-settings>. - Hermes exported history for any key. Before plugin 0.9.1, the Hermes plugin compared the key it was given with itself, so it exported session history for any Pushary key. Since plugin 0.9.1 (CLI 1.9.84, October 24) it exports history only for the account the computer is signed in to.
- A Team that moved to Agent kept posting to its Slack channel. Before October 26, anyone in that channel could answer approvals on a single-seat plan. Now a channel connection on a plan that sends direct messages stops posting new approvals, and Settings shows it as paused with a reconnect.
- Decision webhooks connect to the address Pushary checked. Since October 23, a decision callback connects only to the public address Pushary validated for its URL.
New
- Cline.
npx -y pushary@latest setup --agents clineadds Pushary to Cline's MCP settings, with a global rule that tells Cline to ask and notify through Pushary. Cline chooses when to ask. Its own approval prompts stay in VS Code. Supported agents - Session history from Gemini CLI, Cursor and Hermes. Their conversations sync to your phone, including edits and rewinds, with CLI 1.9.77 or Pushary for Mac 0.1.114 and an updated phone app. Cursor exports have no timestamps and leave out some tool details.
- Live replies from Claude Code sessions opened on your Mac. Your phone shows live previews of replies with Pushary for Mac 0.1.112, Claude Code 2.1.308 or later, and Pushary app 1.0.45.
- Pause phone notifications from your Mac. Pushary for Mac 0.1.122 pauses phone notifications until a time you choose, or until you resume them. The Mac keeps receiving questions while they are paused.
- Sign in to Pushary for Mac from your browser. In 0.1.127, choose Continue in browser, then Continue on pushary.com, and the browser opens Pushary on this Mac to finish signing in. If you are already signed in on pushary.com, there is no second password or email code.
- A custom answer from the notch. Pushary for Mac 0.1.122 adds an Other field.
- Windows installer.
pushary.com/install.ps1checks for Node.js, offers to install Node.js LTS with winget when there is none, then runs setup. It never replaces a Node.js you already have. pushary transcripts onandpushary transcripts offturn session history on or off for one computer. CLI reference- Slack test and welcome. Settings has a Slack test button, the app sends a welcome the first time someone opens it, and Pushary joins the channel you pick.
- Claude from its directory listing. Setup, Pushary for Mac and the Pushary app connect Claude through Pushary's listing in Claude's connector directory. Pasting the connector URL still works. Claude.ai, Claude Desktop, and Cowork
- A starter prompt for chat connectors. After the setup test, people who connect only Claude.ai, Cowork or ChatGPT get a prompt that has the agent ask one real question before it starts work. Phone first run for a connector now ends on a real answered question.
- The
pushary-human-in-the-loopskill is for people building approval steps into their own AI agents. It covers who answers, question types, waits, what happens when nobody answers, and code for MCP, REST, the Partner decisions API and each framework adapter. It ships in the skill repository and the Claude Code and Codex plugins.
Changed
- With permissions off, Claude Code asks only about risky steps. With
claude --dangerously-skip-permissions, routine commands and edits run without a question. Steps that are hard to take back wait for your phone. Examples aregh pr merge,git push --force, a tag push,npm publish,vercel --prod,rm -rf, a secret written to a host, a push to a URL or host address,git remote add, and an edit to.env, a CI workflow, a shell profile or an agent's own settings. Agit pushto a remote the repository already has runs. A rule you wrote, a preset, an approval route, the kill switch and a scope you approved still apply. Other agents, and every other Claude Code mode, ask about what they asked about before. Pushary for Mac does this through the server. The CLI needs 1.9.97, or 1.9.100 for a push to an address. Claude Code - "Never stop me" when permissions are off. Settings has When your agent skips permissions, with Stop me for risky steps (the default) or Never stop me, which lets every step run. A card that stops you only because a step is risky offers Never stop me while permissions are off in Pushary for Mac 0.1.125 and Pushary app 1.0.51. The kill switch and approval routes still apply, and so does a rule you wrote for the commands it names.
- No push when a turn ends cleanly. Since November 1, a clean turn end is recorded without the "Your agent finished its turn" push. A turn that ends on an error still pushes "run failed", and a session started or driven from your phone still gets "finished". CLI 1.9.99 brings that push back for turns you drive from your phone through
pushary claudewhile Pushary for Mac is running. - Answer Claude Code at the terminal or on your phone. Since CLI 1.9.66, in an interactive Claude Code session in default or accept-edits mode, when only Pushary's built-in rules ask about a call, Claude Code shows its own prompt at once. Pushary sends the same question to your phone, which keeps it for up to 10 minutes, and whichever answer comes first wins. A rule you wrote, a preset or a Team approval route works as before. Read the story
- Session history is on by default. A computer that never chose now syncs each session's conversation to your phone, redacted and encrypted on the computer. An owner or admin can turn it off for the workspace under Session history in Agent settings. A session already running when Pushary first sees it syncs only its last 20 MB, inside your plan's history window. This is in CLI 1.9.76 and Pushary for Mac 0.1.113. Task titles and privacy
- A blocked action's question stays open while the agent waits. When Claude Code's auto mode or Codex's auto-reviewer blocks a call and Pushary asks whether to allow it once, the question now stays open for as long as the agent is held. Before October 25, a rule with a one-minute wait closed it after a minute.
- A closed question says so. The phone, the Mac and the dashboard refuse an answer to a question that has closed, and the decision page says "Question expired". An expired question is never an approval. The agent is told not to take the step and to ask again in its current client.
- Codex Full access stops asking about reads in connected apps. Since CLI 1.9.71 and the October 21 server, a call to an app you connected in ChatGPT or Codex that only reads, such as searching Gmail or listing GitHub issues, runs without asking in Full access. Calls that change something, reads that name a secret, any database or BI app, and any call a rule of yours covers still ask.
- OpenCode asks before shell commands, edits and web fetches. Setup,
pushary upgradeand Pushary for Mac 0.1.123 setbash,editandwebfetchtoaskinopencode.jsonwhere no rule of yours names them, so those approvals reach Pushary. - Hermes keeps working after
hermes update. Setup and Pushary for Mac 0.1.107 install the plugin in~/.hermes/plugins/pushary. Runhermes gateway restartonce after updating. A failed tool call no longer pushes "Tool error in Hermes". A failed turn pushes "Hermes run failed" with the reason. - Checkout takes a card or Link. Trial checkouts no longer accept bank debits.
- Billing emails. You get an email when the first charge after a trial goes through, with the amount, when a cancellation is scheduled, and when a plan ends. A failed payment gets reminders about 3 and 7 days later while the invoice is still unpaid. Unsubscribing stops optional emails only. The weekly digest goes to paying workspaces only.
- Pushary for Mac. Home opens from the Dock (0.1.112), and a simpler Home keeps your agents and recent activity together (0.1.122). The notch shows a subagent's state with a count on the agent's mark (0.1.105) and follows what Claude Code says it is doing (0.1.106).
Fixed
- Old CLIs locked a session's approvals. Before October 18, approvals from CLIs before 0.83 never expired, so after three of them every later ask in that session was refused. Until October 27, approvals the server held back from CLIs 0.26 to 0.95.2 also filled the session. Update with
npx -y pushary@latest upgrade. - Phone rule saves. From November 3 to November 4, the Pushary app could not create or edit a rule, or Undo an Always allow that replaced one. On a Team, the app on 1.0.52 or older can edit a team rule but not create one.
- Setup failed while agents were stopped. Since CLI 1.9.91,
pushary setupno longer reports your phone unreachable because Stop all agents is on. - Windows Claude Code hooks failed after an upgrade.
pushary upgrade, an automatic update or a Claude Code update rewrote the hooks with a path Git Bash cannot read, so every Pushary hook failed silently. CLI 1.9.82 fixes it. If Claude Code on Windows stopped reaching your phone, runnpx.cmd -y pushary@latest setuponce. - Cursor asked twice. With Pushary's Cursor gate installed, each shell command and file change was asked about by two hooks. Since CLI 1.9.62 it is asked once.
- "Stop hook blocking error" in Claude Code. A phone instruction that arrives as a turn ends now reaches Claude as context on Claude Code 2.1.163 and later, in CLI 1.9.73 and Pushary for Mac 0.1.110.
- Git worktrees. In a repository with no git remote, "Always allow ... in this repository" now covers its linked worktrees, such as the ones Claude Code creates for subagents. CLI 1.9.92.
pushary login --force, then setup. Setup in the same terminal no longer saves the old key back over the new workspace. CLI 1.9.99.- The phone kept its encryption key. Pushary app 1.0.49 keeps its key when Pushary is reinstalled or secure storage is briefly unavailable, so saved conversations stay readable.
- The phone inbox. Pushary app 1.0.50 loads waiting requests faster, says when it is still checking or cannot reach Pushary instead of showing All clear, and opens a tapped request that is not part of a session on its own.
- Pushary for Mac. A question card stops offering answers when its time is up, and the app no longer uses a full CPU core after your Mac wakes (0.1.119). Signing out and back in no longer leaves Codex showing "Reconnect MCP server pushary" (0.1.104, CLI 1.9.61).
- fx after an idle stretch.
pushary upgradeswitches a hand-written HTTP entry for Pushary in~/.fx/mcp.jsonto the local bridge, so the first call after an idle stretch arrives. CLI 1.9.68 and 1.9.69.
Corrections
- CLI 1.9.74, 1.9.95 and 1.9.98 have entries in the CLI changelog but were never published to npm. Their changes first shipped in 1.9.75, 1.9.96 and 1.9.99.
- The SDK versions below listed CLI 1.9.18, Server SDK 2.2.1 and Python SDK 2.2.0. They now list the current versions.
Codex questions and connection updates, October 15, 2026
CLI 1.9.60 is the current published version as of this update. Releases 1.9.57 through 1.9.60 add supported typed Codex questions, site-access and transfer questions through Codex app cards, and fixes to setup, upgrades and temporary mode expiry.
Read the October product update for the user-facing changes and their limits. The CLI changelog has the complete version history.
Antigravity CLI, Codex and corrected notes, October 2026
Versions covered by the October 1 to 5 entry:
- CLI (
pushary, also published as@pushary/agent-hooks): 1.9.23 - Pushary for Mac: 0.1.66
- Server SDK (
@pushary/server): 2.2.1 - Python SDK (
pushary): 2.2.0
This page skipped the releases between July and October. Every CLI release has its own notes in the CLI changelog, and every Pushary for Mac release has notes on its releases page. This entry covers October 1 to 5.
Security
- A read-only check approved some commands that hide code. From October 1 to October 3, CLI 1.9.1 to 1.9.5 and the server's gate (which Pushary for Mac and the editor plugins use) treated a few shell forms as read-only when they are not: shell arithmetic (
$[ ]), subscripted parameters andprintf -vcan make the shell run text it was given as data. A command built that way could run without asking, with no rule of yours involved. An escaped;,&or|could also hide a destructive word from the check that keeps a broad Bash auto-approve away from destructive commands. CLI 1.9.6 and the server fixed both on October 3, andhermes-plugin-pushary0.5.9 shipped the same day. Update withnpx -y pushary@latest upgrade. Automatic updates install it at the next session start. - A
printfthat hides code was approved as a read. Through CLI 1.9.22, the server's gate passedprintfwith a numeric conversion such as%dor%xas read-only. zsh, the default shell on a Mac, evaluates the argument of a numeric conversion as arithmetic, so an argument like'PWD[$(command)]'ran the command without asking. CLI 1.9.23 and the server fixed it on October 5. Aprintfthat only prints text, or only plain numbers, still runs without asking. - Codex with approvals off ran a call you handed to your machine. With Codex's approval policy set to never, answering "Handle on your machine" on your phone let the call run, a destructive one included, because Codex has no prompt of its own there. Since CLI 1.9.23 and the October 5 server, that answer refuses the call. With Codex's own prompts on, the call still goes to that prompt.
- Antigravity CLI ran calls once a subscription ended. Pushary is agy's only gate, and when the subscription had ended a call Pushary would ask about ran with nothing asking. Since CLI 1.9.23 it is refused with a message that says to renew, or to disconnect agy so its own setting comes back.
New
- Antigravity CLI (
agy).pushary setupand Pushary for Mac connect it. Its shell commands, file writes and edits, and MCP calls wait for your answer on your phone or your Mac. Pushary is agy's only gate: when you have not chosen an agy mode, setup setstoolPermissiontoalways-proceed, and disconnecting puts it back. A call nobody answers, or one Pushary cannot decide, is blocked. Antigravity CLI - Phone start in Pushary for Mac. The Mac app can start agent sessions from your phone without the CLI. It is off by default. Turn it on in the app's Settings.
- fx in Pushary for Mac connects natively over MCP, with no Node needed.
- The CLI's
pair,sign-outanddisconnect fx.pairandsign-outare the new names forconnectandlogout, which still work. CLI reference
Changed
- New workspaces start on When I'm out. Existing workspaces keep their mode. Under When I'm out your phone is asked when you are away from the computer, and at the keyboard the agent's own prompt asks you. Since June these docs called When I'm out the default, while workspaces created on the web started on Every time. Now both agree.
- From October 12, new workspaces start on Always again. A new workspace's phone is asked for every approval, whether or not you are at the computer. The agent waits for the phone for up to 60 seconds and then asks in its own prompt. Workspaces created between October 4 and October 12 stay on When away, and every existing workspace keeps the mode it has. You can change it under Policies.
- Agents with no prompt of their own ask your phone every time. When the Pushary CLI or the Cursor plugin runs the gate, these wait for your answer for your policy's timeout, under When I'm out and also while you are at the keyboard: Cursor file edits, sessions the CLI starts from your phone, Antigravity CLI, a Codex edit outside your agreed scope, and Codex with its approval policy set to never. Before, they were often refused with nobody asked. For Codex with approval never, a call nobody answers is now refused. Before, it ran once the wait ended. When nobody answers
- Codex asks less, and asks about scope. In Every time, Pushary asks before a Codex call only when it leaves the scope you agreed, matches a rule or preset you set, destroys, rewrites history, reaches the network, or deletes a file. A Codex edit outside your agreed scope is now asked or refused. Before, Codex applied it without asking.
- More read-only commands run without asking, for every agent.
rg,sort,uniq,sleep,pgrep,git branch --show-current, and Codex'swait_threadsandsend_message_to_thread. Options that run another program or write a file still ask. A deny rule now blocks read-only commands too. This is in CLI 1.9.12 and, through the server, in Pushary for Mac and the editor plugins from October 4. - An interrupted approval stops the call. When an approval is cancelled, expires, cannot be recorded or withdrawn, or Pushary for Mac quits or signs out while an agent waits, the call is stopped and the question is withdrawn. In Claude Code the stop also ends the agent's turn. Before, the server denied only that call, and Pushary for Mac sent it back to the agent's own prompt. This is on the server and in Pushary for Mac 0.1.66.
- The daily budget counts Pushary for Mac sessions. A Mac that runs agents only through the app can now reach the budget and stop its agents. The app sends token counts, never prices, and the server prices them. A session the CLI and the app both report is charged once.
- Pushary for Mac opens at login after a new install, and
brew uninstall --zaprevokes the Mac's device key on pushary.com before deleting it. pushary clean --everythingrevokes the keys the CLI wrote before removing them.
Fixed
- Cookie consent on pushary.com. PostHog captured nothing before consent, but it still stored its identity cookie and browser storage before you answered, and kept a cookie after Reject All. Since October 4 nothing is stored until you accept. Visitors in the EEA, the UK, Switzerland, Canada and Australia now have a "Cookie settings" link to change their answer, and the banner appears on your first click, key press or scroll instead of after a timer.
Corrections
- The CLI changelog entries for 1.9.8, 1.9.12, 1.9.13, 1.9.15, 1.9.17 and 1.9.18 were incomplete, and 1.9.17 said something the code does not do. They were corrected on October 4, and each corrected entry says what changed.
- Pushary for Mac 0.1.66 was released with no notes. They were added to its releases page on October 4. From the next release, the update window in the app shows the notes too.
- This page listed Server SDK 1.4.0 and 1.0.1 while 2.0.0 had shipped on September 5. The SDK versions and breaking changes below are now current.
Claude Agent SDK adapter, September 2026
@pushary/claude-agent-sdk 0.1.0 is a canUseTool for the Claude Agent SDK. Before a
tool that needs permission runs, it asks your user on their phone and returns allow or
deny. A denial, an expiry or no answer all deny, and cancelling the run stops the wait.
See the Claude Agent SDK page.
Decision layer: policy answers first, one gate for every caller, August 2026
This page did not list these changes when they shipped. They were added on November 4, 2026. Together they are the part of the decision layer that decides before anyone is asked: a rule answers what you already delegated, and a person is asked only for the rest. The manifesto explains why.
New
pushary.authorize()lets policy answer before a person is asked.POST /api/v1/server/authorizereturnsallow,denyorrequires_human, andauthorize()callsdecisions.ask()only forrequires_human. Rules made withPOST /api/v1/server/authorization-rulesmatchtoolName, optionallytoolTarget, and scalarparameters. An unmatched action asks a person.evaluateAuthorization()returns the verdict without opening a decision. These need the Partner plan. Let policy answer- One authorization, one execution.
protect()authorizes, asks a person when policy defers, and runs your action. Its permit binds the authorization to the action, and a second claim of the same authorization is refused. A custom executor callsPOST /api/v1/server/authorizations/consume, thenPOST /api/v1/server/authorizations/{permitId}/receiptwith the outcome. Run it once - The Partner decisions page shows the outcome your server reported for an approved action, and exports add two columns for it.
- One gate for every caller.
POST /api/agent/gateruns the same engine as the CLI's hook, so Pushary for Mac and the Cursor and VS Code gates get the same verdict over the same state. The editor gates no longer carry their own copy, which had drifted: the VS Code copy matched rules on the tool name alone.
Enforced approval gates, August 2026
Until now every adapter shipped one shape: a tool the model chooses to call. A model in a hurry can decline to call it, which makes it a suggestion. Every framework we support had already shipped a real gate the runtime evaluates before a tool runs, and no adapter of ours used it. This wave uses all of them, so "ask a human" and "this cannot happen without a yes" become two things you pick between. See the adapters guide.
Current SDK versions:
- Server SDK (
@pushary/server): 1.4.0 - Python SDK (
pushary): 1.4.0
New
@pushary/ai-sdkpusharyApproval/pusharyToolApproval. Plugs into the AI SDK'stoolApproval, which runs before a tool executes. Gate one tool by name, a named set, or the whole run. Needsai@7; the ask tool still works fromai@5.@pushary/openai-agentsandpushary-openai-agents:resolvePusharyInterruptions/resolve_pushary_interruptions. The OpenAI Agents SDK'sneedsApprovalonly decides that a human is needed and then stops the run withresult.interruptions; nothing asks anyone. This asks about each interruption in order, approves or rejects it on the run state, and hands a denial back to the model as the rejection message.@pushary/mastraresolvePusharyApprovals. The same split in Mastra:requireApprovalsuspends the run, this asks and then resumes or declines. With norunspassed it drains the agent's own suspended runs, so a background worker can clear a whole thread.@pushary/server/adaptersandpushary.adapters. The kernel every official adapter is built on, now public in both languages. Writing an adapter for an in-house harness is a thin binding overaskExternalUser,createDurableDecisionandcreateGaterather than a reimplementation of idempotency, fail-closed semantics and callback verification.
Changed
- Every adapter now shares one kernel instead of carrying its own copy of the same ~150 lines. Public APIs are unchanged, and a bug fixed in the kernel is fixed in all of them at once.
@pushary/everequireseve@0.31or newer, and its README'spusharyApprovalexample was wrong: Eve names a tool by its file path, sodefineTooltakes nonamefield and does requiredescription. The old snippet did not compile.- The three PyPI adapters now run in CI. They never had before.
Framework adapters, July 2026
New wrapper packages so each framework gets a ready-made human-in-the-loop tool, not a
snippet to paste. Every one is a thin, fail-closed wrapper over the same two-call
contract (enroll + decisions.ask), so there is no new durability to trust. See the
adapters guide.
New
@pushary/langgraph/pushary-langgraph. A blockingask_humanfor a LangGraph node, pluspushary_interruptthat wraps LangGraph's nativeinterrupt()for a durable, webhook-resumed wait. The idempotency key is derived fromexternal_id + node + question, so a node that re-runs on resume never pages the human twice.@pushary/mastra.createPusharyAskToolfor a blocking approval, andpusharyApprovalStepfor a durable Mastra workflow step that suspends and resumes on a signed webhook.@pushary/durable. Orchestrator-agnostic building blocks for Inngest, Temporal, and Vercel Workflow:createApprovalopens the decision,resolveApprovalverifies and parses the callback in one call. No framework dependency of its own.@pushary/openai-agents/pushary-openai-agents. A function tool for the OpenAI Agents SDK that bindsexternalIdin code, so a prompt-injected model cannot ask the wrong person.pushary-crewai.make_ask_human_tool(external_id)returns a CrewAIBaseToolthat replaces the consolehuman_input=Trueprompt with a phone approval.- Hermes partner tools.
hermes-plugin-pushary0.4.0 addspushary_enrollandpushary_ask_end_user, so a Hermes agent in gateway mode can ask its own end-users durably over the REST decisions ledger.
Partner platform, July 2026
Current SDK versions:
- Server SDK (
@pushary/server): 1.3.2 - Python SDK (
pushary): 1.3.2
This wave hardens the embed decisions path for partners running human-in-the-loop for their own end-users.
New
- Reachability signal. A decision addressed to an end-user now returns
reachable,reachableChannels, anddeviceCountat create time, so an approve-gate can tell "no human was reachable" apart from "the human declined." PassrequireReachable: trueto refuse with a 409 instead of opening a decision that reaches nobody. Embed - Callbacks echo your context. The signed webhook body is now
{ correlationId, answer, value, answeredAt, context }.answerstays canonical (valueis an alias), andcontextechoes whatever you passed at create time, so a stateless handler can carry its own run id through the callback without keeping a map. NewparseDecisionCallbackhelper in both SDKs. - Per-end-user (bound) keys. Mint a key bound to one end-user for a session with
keys.issue({ externalId }), hand it to the agent, and revoke it withkeys.revoke(keyPrefix). A bound key can only create, read and cancel decisions for that exact user, over both the REST API and MCP, so a prompt-injected agent acting for one user can never reach another. It cannot enroll devices or answer a decision. Enrollment requires your full-access key on your trusted backend. The answer comes from the person, or from a full-access key on your server. Since October 2026,decisions.answer()with a bound key returns 403 with the codebound_key_cannot_answer.
Remote Control, July 2026
Current agent-side version:
- Agent hooks (
@pushary/agent-hooks): 0.46.0
This wave closes the distance between "my agent stopped and I am not at my desk" and "I handled it from my phone." Your phone now shows what each agent actually did, answers its real questions, warns you when one dies, and can drive an agent that has gone completely idle — from the same pushary claude you already run, no flag.
New
- Reach an idle agent from your phone — by default. Run
pushary claudeand you get the exact native Claude Code terminal. The moment you send an instruction from your phone, Pushary takes the wheel and drives the agent, even if it was sitting fully idle; press Ctrl-] to take the terminal back. It is the same session throughout, so nothing is lost across the hand-off. No flag to remember: the reach is on whenever you launch throughpushary claude. It sets itself up on first run and uses your existing Claude login, never a key of ours.--remotestarts headless for a kick-off-and-walk-away run. With no Pushary key, or for a one-shot-pprint,pushary claudestays a transparent passthrough that behaves exactly likeclaude. - Real questions on your phone. When your agent asks a multiple-choice question, the options are pushed to your phone and your tap answers it in place, instead of the question waiting at a terminal picker.
- Live plan mirror. As the agent updates its to-do list, your phone shows the checklist ticking off, so you can watch progress without reading the terminal. Control Panel
- Real completion summaries. The finish notification now says what the agent actually finished, taken from its last message, instead of a generic "done."
- Rate-limit and crash alerts. If a session dies on a rate limit, an overload, or an auth or billing error, your phone gets told instead of the run going silent.
- More tools reach your phone. Tools that per-tool matching could never gate (web fetch, web search, sub-agent Task calls, and MCP tools) now route their approval to your phone as well. Policies
- Session presence. A session appears the moment it starts and drops to offline the moment it truly closes, rather than only looking idle between turns. Control Panel
Changed
- Remote mode needs no setup step. The first preview asked you to install the Agent SDK by hand and run
claude setup-tokenif you signed in with a subscription. Nowpushary claude --remotefetches what it needs on first run and uses your existing Claude login, whether that is a subscription, an API key, Bedrock, or Vertex, so a plain install is all it takes. It reuses theclaudeyou already have rather than shipping its own copy, so the download stays small. - In Claude Code's auto, plan, and don't-ask modes, Pushary now steps out of the way instead of pushing a second approval on top of Claude Code's own. Auto mode no longer buzzes your phone for calls it was going to approve anyway.
pushary upgradenow rewrites your Claude Code settings, so newly added hooks take effect on upgrade instead of staying inert until a fresh setup.pushary doctornow checks every installed hook and tells you the one command to fix a missing one.
Fixed
- Failed tool calls are visible again. The post-tool hook was reading the wrong response field, so a tool that errored could look like it succeeded. Receipts and status now reflect real failures.
Intelligence Layer, June 2026
Current agent-side version:
- Agent hooks (
@pushary/agent-hooks): 0.14.0
Pushary started reading your decision history and proposing rules from it. It now shows what every agent is working on, records what they actually did, and brings the same depth to Codex that Claude Code already had.
New
- Policy Autopilot. Pushary mines your approval history and suggests the rules you keep approving by hand. Approve
git push14 times in a repo, get a one-tap rule that auto-approves it. Your phone buzzes less every week. Policies - Fleet Board. Fire several prompts and walk away. One board shows every agent as a card with its task title and live state (working, waiting on you, done, errored), answerable inline. Control Panel
- Agent Receipts. Every session ends with a receipt of what changed: files edited, commands run, commits created. Paths and command heads only, never your code. Searchable and exportable. (Agent Pro / Team.)
- Weekly digest. One email a week: how many times your agents stopped to ask, how many your policies cleared automatically, and the rules worth adding next.
- Codex native hooks. Codex now gets the full control loop through its native hooks: per-tool policy, phone approvals, the kill switch, task titles, and receipts. Setup writes
~/.codex/hooks.jsonon Codex 0.122 and above, with a notify fallback for older versions. Codex setup - Arg-aware policies. Rules now match command shape, not just tool name:
Bash(npm test:*),Bash(git status), or a bareBash. "Always denyrm -rf" is finally something you can express. Policies - Mobile-aware onboarding. Sign up on a phone and the flow connects your phone first, then emails you the one-line setup command for when you are back at your computer. No more dead-ending on a terminal step.
Changed
- A rule set to deny with a zero-second timeout now denies instantly with no push, instead of pushing and timing out into a deny. If you had such a rule, the push will stop. That is the rule doing what it says.
- The policies panel and the subscription page were redesigned: plain-language rules with controls behind a tap, and a calm billing page in place of the upgrade wall.
Fixed
- Accepted policy suggestions and saved rules take effect on the agent immediately, instead of waiting up to five minutes for the local cache to expire.
- The append-only audit trail now also pins the tool and command-target columns, so the immutable record covers receipts.
Awareness Layer — May 2026
Current agent-side version:
- Agent hooks (
@pushary/agent-hooks): 0.9.0
The multi-agent control layer for AI coding agents. Run agents in parallel, see which one is blocked, and stay in control from your phone.
New
- Per-agent identity. Every session is keyed by
sessionId+ a hashedmachineId, so parallel windows of the same agent no longer collapse into one row. Control Panel - Multi-agent control panel &
list_sessions. A read-only MCP tool and dashboard view of the live fleet, with a Blocked Queue that floats the agent waiting on you to the top. Control Panel - Kill switch. Halt a runaway session from the dashboard or the lock screen — denies every tool call, even auto-approved ones, until you release it. (Agent Pro / Team.) Kill Switch
- Permission policy editor. Set per-tool approval mode, timeout, and timeout action in the dashboard. Policies
- Teach-on-Tap. "Always allow this tool" on the decision page mints a durable auto-approve policy. Policies
- Append-only audit trail + export. Database-enforced immutable record of every question, approval, and notification, with CSV/JSON export. (Agent Pro / Team.) Audit Trail
- Delivery receipts, fallback & telemetry. Web-push acknowledgment tracking, latency telemetry, and an email fallback for undelivered agent approvals. Receiving Notifications
- Agent pricing tiers. Free / Agent / Agent Pro / Team, every paid plan a card-first 7-day trial. Pricing
Fixed
- Parallel-agent decision loss. Ending one session no longer cancels another live session's pending approval — cleanup is now scoped to the finishing session.
- Per-session rate limits. Approval questions are bounded per session (3) and per site (25) so a fan-out of agents is never throttled by a single one.
- MCP scaling. Replaced a full Redis keyspace scan with a per-site pending-question index.
Initial release, January 2026
Versions at release:
- Client SDK (
@pushary/sdk): 1.0.5 - Server SDK (
@pushary/server): 1.0.1 - Platform: 1.0.0
Initial public release of Pushary push notification platform.
Features
Client SDK (@pushary/sdk)
- Browser push notification support (Chrome, Firefox, Safari, Edge, Opera)
- Service worker integration
- Auto-prompt and manual permission flow
- Subscriber identification with external IDs
- Tag-based segmentation
- Click and impression tracking
- iOS Safari PWA support
- Retry queue for offline reliability
- Device detection utilities
- TypeScript support with full type definitions
Server SDK (@pushary/server)
- Complete REST API client for Node.js
- Subscriber management (list, get, update, delete, count)
- Campaign management (create, update, send, pause, resume, stats)
- Template management for reusable notifications
- Direct notification sending (by subscriber ID, external ID, or tags)
- Flow automation (trigger-based notification sequences)
- TypeScript support with full type definitions
Platform Features
- Multi-site support with isolated VAPID keys
- Team collaboration with role-based access
- A/B testing for campaigns (Starter plan and above)
- Flow automation with triggers and steps
- Analytics dashboard with delivery and engagement metrics
- Rate limiting per plan tier
- Plan-based feature gating
- Subscriber segmentation with tags and filters
- Campaign scheduling
- Data retention based on plan
Authentication
- Two-tier API key system:
- Site Key (
pk_xxx) for client-side use - Full API Key (
pk_xxx.sk_xxx) for server-side use
- Site Key (
- Domain validation for client requests
- SHA-256 hashed API key storage
- Rate limiting per workspace and plan
Integrations
- Shopify integration (via extension)
- Webhook support for events
- REST API for custom integrations
API Endpoints
Client SDK Endpoints:
POST /api/v1/subscribe- Subscribe to notificationsPOST /api/v1/unsubscribe- Unsubscribe from notificationsPOST /api/v1/track- Track eventsPOST /api/v1/identify- Set external ID and tagsPOST /api/v1/subscription-change- Update subscription endpointGET /api/v1/site/:siteSlug- Get site configuration
Server API Endpoints:
POST /api/v1/server/send- Send notifications directlyGET /api/v1/server/subscribers- List subscribersGET /api/v1/server/subscribers/:id- Get subscriberPATCH /api/v1/server/subscribers/:id- Update subscriberDELETE /api/v1/server/subscribers/:id- Delete subscriberGET /api/v1/server/subscribers/count- Get subscriber countGET /api/v1/server/campaigns- List campaignsPOST /api/v1/server/campaigns- Create campaignGET /api/v1/server/campaigns/:id- Get campaignPATCH /api/v1/server/campaigns/:id- Update campaignDELETE /api/v1/server/campaigns/:id- Delete campaignPOST /api/v1/server/campaigns/:id/send- Send campaignPOST /api/v1/server/campaigns/:id/pause- Pause campaignPOST /api/v1/server/campaigns/:id/resume- Resume campaignGET /api/v1/server/campaigns/:id/stats- Get campaign statisticsGET /api/v1/server/templates- List templatesPOST /api/v1/server/templates- Create templateGET /api/v1/server/templates/:id- Get templatePATCH /api/v1/server/templates/:id- Update templateDELETE /api/v1/server/templates/:id- Delete templateGET /api/v1/server/flows- List flowsPOST /api/v1/server/flows- Create flowGET /api/v1/server/flows/:id- Get flowPATCH /api/v1/server/flows/:id- Update flowDELETE /api/v1/server/flows/:id- Delete flowPOST /api/v1/server/flows/:id/activate- Activate flowPOST /api/v1/server/flows/:id/pause- Pause flow
Pricing Plans
- Free: 100 subscribers, 500 notifications/month
- Starter: 5K subscribers, 25K notifications/month - €29/month
- Growth: 25K subscribers, 100K notifications/month - €99/month
- Enterprise: Unlimited, custom pricing
Known Limitations
- Free plan does not have API access (client SDK only)
- iOS Safari requires PWA installation (not available in browser tab)
- Service worker must be at root path (
/pushary-sw.js) - HTTPS required (except localhost for development)
- Rate limits vary by plan (100-1000 requests/minute)
- Data retention: 7-90 days depending on plan
Migration Guides
For New Users
No migration needed. Follow the Getting Started guide to begin.
Server SDK and Python SDK 2.0
Upgrade the framework adapter together with the SDK. Durable helpers now require an explicit idempotencyKey tied to the run and the step. The 0.3 LangGraph, Mastra and OpenAI Agents adapters support it. Earlier adapters must stay on the 1.x SDK until they are upgraded. Independent blocking calls no longer share a decision because their question text matches.
Version 2.1 changed how human gate keys are built. Finish operations already in progress on the SDK version that started them, or migrate their stored bindings yourself. Do not replay an operation that already ran through the new key scheme.
Deprecation Policy
Policy
- Breaking changes will be announced at least 90 days in advance
- Deprecated features remain functional during deprecation period
- Migration guides provided for all breaking changes
- Enterprise customers receive dedicated migration support
Current Deprecations
No deprecations at this time.
Breaking Changes
Server SDK and Python SDK 2.0.0, September 5, 2026
Durable helpers require an explicit idempotencyKey, independent blocking calls no longer reuse a decision by question text, and @pushary/mastra 0.3 requires an operationId for durable inputs. See Server SDK and Python SDK 2.0.
Platform 1.0.0 / SDK 1.0.5 / Server SDK 1.0.1
Initial release, no breaking changes.
SDK Versions
CLI (pushary)
See the published CLI version and the CLI changelog. The same CLI is published as @pushary/agent-hooks.
Client SDK (@pushary/sdk)
Current: 1.1.1
Compatibility:
- Works with Pushary API v1
- Supports all modern browsers (see Browser Support)
- Service worker compatible with all supported browsers
Server SDK (@pushary/server)
Current: 2.5.1
Compatibility:
- Node.js 18.0.0 or higher
- TypeScript 5.0 or higher (optional)
- Works with Pushary API v1
Python SDK (pushary)
Current: 2.4.1
Compatibility:
- Python 3.9 or higher
Roadmap from January 2026
This is the push notifications roadmap as it was published in January 2026. It is kept here as published and is not a current plan.
Q1 2026
- Advanced segmentation with computed properties
- Webhook event stream
- Campaign A/B testing enhancements
- Custom notification actions
- Rich notification templates
Q2 2026
- Mobile SDKs (React Native, Flutter)
- Advanced analytics dashboard
- Predictive send time optimization
- Multi-language template support
- Workflow builder UI
Q3 2026
- AI-powered notification optimization
- Advanced automation rules
- Custom integrations marketplace
- Geo-targeting
- Time zone optimization
Future
- SMS and email channels
- In-app messaging
- Push notification previews
- Advanced permission prompts
- Message scheduling AI
Want to influence the roadmap? Contact us at business@pushary.com
Security Updates
Security fixes are listed on this page under the release that ships them, with what was affected and for how long. Report a vulnerability to business@pushary.com.
API Versioning
Pushary uses URL-based API versioning:
- Current:
/api/v1/ - Future versions will be announced with migration guides
Version Support:
- Latest version: Full support
- Previous version: Security updates only (12 months)
- Older versions: No support
Getting Updates
Stay informed about new features and changes:
Where each part of Pushary publishes its release notes:
- CLI: the CLI changelog, also inside the package on npm.
- Pushary for Mac: the releases page, and the update window in the app.
- Pushary app for iPhone and Android: the release notes in the App Store and Google Play.
- Framework adapters and SDKs: the
CHANGELOG.mdin each package on npm and PyPI. - Server changes: this page. A change on the server can change what your agents see without a new CLI or app release.
Breaking changes are listed under Breaking Changes with a migration guide.
Feedback
Have feedback or feature requests?
- Email: business@pushary.com
- Docs: the feedback buttons at the bottom of each docs page
- Enterprise customers: Contact your account manager
We review all feedback and use it to shape our roadmap.