chore(deps): update dependency can1357/oh-my-pi to v18.22.1 #266
Loading…
Reference in a new issue
No description provided.
Delete branch "renovate/can1357-oh-my-pi-18.x"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
This PR contains the following updates:
18.22.0→18.22.1Release Notes
can1357/oh-my-pi
v18.22.1Compare Source
@oh-my-pi/pi-agent-core
Fixed
filterProviderReplayMessages, the defaultconvertToLlm) now sends a tool call stopped by a provider error as "not executed" without the provider's error text, so a refusal explanation is never fed back to the model (#19419 by @flosrn)@oh-my-pi/pi-ai
Added
buildAnthropicSearchHeaderstakesextraBetas, request-specific betas sent beside the web search beta (#18125 by @will-bogusz)omitDemotedReasoningtoStreamOptions: an Anthropic request leaves out other models' reasoning that it would otherwise replay as text, and a refused request reports how many such blocks it carried indemotedReasoningBlocks(#10407) (#17610 by @will-bogusz)metadata.credentialIdto usage reports probed from stored credential rows (API-key and OAuth), and a read-onlyAuthStorage.sessions.stickylookup of the credential a session is currently served from (#19456) (#19492 by @thekitaev).riskAcknowledgementtoOAuthPrompt, set on the login-warning acceptance prompt so hosts that refuse input before the auth URL can still show it (#19523)Fixed
prompt_tokens/completion_tokensinstead ofinput_tokens/output_tokens(#19053 by @Xytronix)thread_unsupported_request), so every such request first paid a rejected round trip before resending without the thread (#19473 by @alphastorm)yieldwith a typed output schema) being published as empty objects, which left subagents unable to submit results (#19502)@oh-my-pi/browser-relay
Fixed
@oh-my-pi/pi-catalog
Added
openai/gpt-6.2-astra($8/$40, $0.40 cached) andopenai/gpt-6.2-sol($2/$10, $0.10 cached) are now bundled with a 1.05M window, a higher price tier above 272K input, and Ultrafast at 3x through 2027-02-28. On Codex,gpt-6.2-astraandgpt-6.2-solshow API-equivalent costs, bill Fast at 2x and Ultrafast at 6x, offer a 922K window with extended context on, and use only Plus-or-higher accounts.bifrost), a self-hosted OpenAI-compatible AI gateway: its/modelsroster is discovered live per virtual key fromBIFROST_BASE_URL(defaulthttp://localhost:8080/v1), chat models run over Chat Completions, known decision models becomeopenai-decisionsjudges on the gateway's/decisionsroute, and/login bifroststores the virtual key (BIFROST_API_KEY) (#19053 by @Xytronix)imagerole's Factory defaults (#19191 by @will-bogusz)Fixed
duo-chat-opus-*,duo-chat-sonnet-*,duo-chat-haiku-*) not being treated as Claude when quoting earlier reasoning: compaction summaries and interrupted-turn notes now leave it out for them, and areasoning_extractionrefusal is retried without it, as for Claude elsewhere (#17610 by @will-bogusz)@oh-my-pi/pi-coding-agent
Breaking Changes
commithelper now throws aTypeErrorfor option names it does not know (e.g.commit.draft({ cwd }), which used to draft the session's own checkout) and for a non-object options argument, naming the expected options; scripts that passed extra keys must drop them. Committing a draft of an unchanged repository now says so instead of suggestingtake()(#19440)Added
providers.webKeylesssetting (default on): turn it off andweb_searchand URL reading use only services you configured, skipping keyless Parallel, Exa, Firecrawl, Keenable, and Perplexity, the Jina reader, and scraped engines even when yourwebchain lists them;omp q --model web/<engine>still runs one on request (#13694, #16336)live.endpointandlive.apiKeysettings: live voice (/live, RPClive_start/live_connect, extensionstartLive) can run on any OpenAI Realtime WebRTC-compatible server instead of Codex, with no ChatGPT sign-in and in offline mode;live.voiceaccepts any voice nameel.setValue(75)now writes the number as"75"instead of refusing a non-string;NaN,Infinity, booleans and objects are still refused, as are a whole number past 2^53 and a number that prints with an exponent, which are passed as text in digits instead (#19405 by @will-bogusz)<select>menu, colour or date picker, or a print, save, sign-in, passkey or permission sheet or bubble a call raised over its tab is named with itscomputer.window()id while the browser's tab list shows that tab in front of its window, by the next call when it shows just after the call returned, and on failed calls too; connected CDP endpoints, remote relays and page-backed popup windows are excluded (#18925 by @will-bogusz)select()accepts{label},{value}, and zero-based{index}choices alongside strings, including mixed choices for multi-selects, on Chromium, Tern, and cmux tabs (#19518 by @will-bogusz)Changed
computertool's takeover notice now says what it did with your keyboard ("handed the keyboard back to Notes's window", "left as the user chose") instead of "focus returned", which it could not verify (#19208 by @will-bogusz)tab.observe()andtab.ariaSnapshot()on Chromium tabs now wait, up to 3 seconds, for the requests and page updates (including child frames) the actions since your last look started, so a cell that clicks and then observes shows the result instead of the page before it; concurrent looks share the wait even across another action, clearing the request log does not discard pending work, and requests already open before the action do not delay the look (#19045 by @will-bogusz)InputBusysays which one: its agent id and omp process id (#18987 by @will-bogusz)browser.relayon) or over your Tern or cmux pane unless the open asks withapp: "relay"orapp: "tern"; they usebrowser.cdpUrlor omp's own Chromium (#18789 by @will-bogusz)xd://eval/browser: a short guide follows the first browser call as a hidden session message, outside cell output limits and shaping. Compaction or a rewind that drops or shortens it sends it again. Advisors keep reading the full guide first; real-session, borrowed-tab and page-integrity safety rules arrive before the first call (#18786 by @will-bogusz)computerwin.move()of a session now says in its output that the app gets no hover from it and that a control shown only on hover is pressed with a takeover click, instead of returning as a plain success (#19509 by @will-bogusz)commithelper now also runs Git hooks defined in config (hook.<name>.event, Git 2.54+), alongside.git/hooks/ones, asgit commitdoes. With such hooks configured, including in your global Git config, commits made through omp can now be slower or rejected where they used to pass (#19443)Fixed
!command); they now report the credential erroryield(directly or throughxd://yield) without finishing the delegated task; eval now rejectsyieldand points to the direct tool (#19470 by @alphastorm)ready.logand wait patterns not matching^and$at line boundaries after earlier output (#19471 by @alphastorm)stat,diffandfindbuiltins ignoringTZwhen printing file times;find -lsnow prints times in the shell's time zone instead of UTC, andfind -newermtreads dates likeNov 15, 2023in the shell's time zone (#19482 by @alphastorm)writeandeditcalls no longer slow down as their arguments stream (#19494 by @H4vC)/commitand evalcommitcreating a single-parent commit during a merge; committing now refuses while a merge, cherry-pick, revert, rebase orgit amis in progress or the index has unresolved conflicts (#19474 by @alphastorm)apply_patchtool description calling it a shell command and forbidding absolute paths (#19325, #19468 by @alphastorm)rgbuiltin ignoring--globpatterns that contain a slash (#19476 by @alphastorm)commitonly(...)failing when one of its paths or globs matched no changed file; unmatched ones are now reported (#19472 by @alphastorm)confirmorpromptas still open, and failing page calls on it, after the user clicked it away or another CDP client answered it (#19481 by @will-bogusz)computersetValueon a<select>in Chrome or an Electron app on macOS failing with an empty list of options; it now chooses the option the value names (#19224 by @will-bogusz)UserIntervenedthe same way (#19296 by @will-bogusz)site:-scoped queries to Claude with the domain and path still in the text, and tells Claude to always search, so those lookups and queries it thought it could answer from memory no longer fail with "returned no sources" (#18125 by @will-bogusz)too_many_requests, …) by name instead of "returned no sources" (#18125 by @will-bogusz)agent()calls withoutschema) no longer get ayieldtool; their final message is the result.yieldwithoutdatano longer uses the subagent's last message as the result; it only finalizes incremental sections that were already submitted.computerkey press that opens a macOS popup button's or menu button's menu in the background now fails listing the items and naming the press or click that chooses one, instead of leaving the menu open with your keyboard (#19179 by @will-bogusz)Refusal (reasoning_extraction)because of that carried reasoning, the turn is retried once without it, and the session stops sending it to Claude (#7241) (#17610 by @will-bogusz)reasoning_extractionrefusal stops making every following turn of the session likely to be refused too (#19419 by @flosrn)/commit's status scan hanging when a background git process kept the output pipe open (#19478 by @alphastorm)--continueopening a new empty session when the terminal's last session was started in another directory and closed before its first message; it now resumes the current directory's newest session (#19495).pre-commithook's stdout (e.g. thepre-commitframework's per-hook report) missing from the error raised by the evalcommitnamespace andomp commit(#19449)/commitnow shows the subject a commit-msg hook wrote onto the commit, instead of the subject from before the hook ran (#19448 by @eggpeat)^modelreferences followed by sentence punctuation now register their model agent instead of remaining literal (#19445 by @yuzu-octopus)/copynow copies what a mid-turnout://reply showed, instead of the reply tool's own "Shown to the user." result, and lists the reply's code blocks in the turn's blocks (#19551 by @bannert1337)opencode-go; it now follows the credential row the session router actually serves from (#19456) (#19492 by @thekitaev).fill,typeandpressaimed at a field behind it now fail naming what covers it, asclickdoes, instead of typing into the dialog's button and reporting success (on Chromium tabs and Tern tabs without native DOM reads);pressalso refuses a disabled element (on Chromium tabs and Tern tabs without native DOM reads), and Chromiumdragrefuses a covered start or end (#19524 by @will-bogusz)loginrefusing providers with an at-your-own-risk warning (antigravity-cli,google-antigravity): the warning now reaches the host as anextension_ui_requestinput, and login continues on "yes" and cancels on anything else (#19523)node_modulesand.git/objects, which could use up the system's inotify watch limit and keep other programs such as systemd from adding watches (#19459)/restartnow re-execs through the launcher omp was started by, so an open session moves onto an upgraded binary instead of re-running a retained old install (#19529 by @yuzu-octopus)agent_endafter an over-quota 429: usage-limit bookkeeping is bounded so a stalled usage report can no longer hold the event (#19411, #19417 by @yuzu-octopus)@oh-my-pi/pi-mnemopi
Fixed
Invalid JSON) when clipping an oversized input cut an emoji or other astral-plane character in half (#19455)@oh-my-pi/pi-natives
Added
DesktopFocusReturn.keyboardsays what a macOS takeover did with the user's keyboard (handedBack,leftAsChosen,kept,frontOnly,notSent) (#19208 by @will-bogusz)desktopWindowRoster()to@oh-my-pi/pi-natives/desktop: every on-screen window on macOS, popup menus, pickers and system prompts included, with owner, layer and whetherlistWindowsreports it, in one pass and without a desktop session (#18925 by @will-bogusz)readTreeKeepingSparse()to seed a private commit index from HEAD while keeping the shared checkout's sparse exclusions;diffText({ cached: true }),commitCreate()andwriteTree()now preserve out-of-cone files and collapsed sparse-directory entries (#19453 by @yuzu-octopus)Changed
scrollnow goes the same way as on Windows and Linux: a positivedyscrolls down and a positivedxright, where macOS scrolled up and left (#19165 by @will-bogusz)InputBusynames the agent holding input/focus control and its omp process id (agent "Main.Scout" (omp pid 4242) owns input/focus control) when the holder's session setagentId, on macOS and Linux, and on Windows when that agent runs in the same omp; cross-process names retain the last 64 characters so long dotted ids keep their distinct suffixes (#18987 by @will-bogusz)computer.type()now types characters your keyboard layout has no key for, such asä,日本or emoji, on compositors that offer libei text (KDE Plasma 6.8 and later); elsewhere they are still refused before anything is typed (#18013 by @will-bogusz)Fixed
press()on a list row, cell or text that has no press action sending agents to a window click at the element'sbounds(): the refusal now names the element's ownclick()and lists only actionsperform()takes (#19063 by @will-bogusz)click(), which needs none (#19063 by @will-bogusz)computertool failing a click or press on an Activity Monitor process row withStaleRefwhenever the table refreshed sinceax()read it: the ref now follows its row when exactly one row still reads as it did in every cell, a click is refused when the app has sorted another row into its place, and the error otherwise shows the row then and now (#19048 by @will-bogusz)el.typeputting the agent's text into the field the user was typing in, and reporting success, while it took over the keyboard: GTK fields and Qt multi-line text edits now take printable ASCII text through accessibility, with no keys, focus change or takeover; an unfocused GTK single-line field requires an empty or fully selected value so its selection-on-focus setting cannot cause text loss. On Wayland those fields now take text instead of refusing. Other fields keep the key path, including Qt single-line fields where an accessibility write skips validation (#19035 by @will-bogusz)press()on a GTK 3 or Qt combo box opening its list as a popup that took the keyboard, so a user typing elsewhere lost their keys to it: that press is now refused with nothing sent, even when takeover is allowed, naming the options.setValue("Two")andpress({ menu: ["Two"] })choose a GTK 3 or Chromium combo box's option with no popup and check the choice took; Qt combo boxes have no such route and are refused (#19034 by @will-bogusz)ax()now shows a toggle in a SwiftUI settings form with the label beside it, andfind({ title })matches that label, instead of leaving the control unnamed (#19061 by @will-bogusz)computerinput leaving your window inactive after putting your app back in front, both after a takeover and after background input whose target brought itself forward: the title bar stayed inactive and web pages saw a blur until you clicked, while a takeover reported focus returned (#19208 by @will-bogusz)axSetValuerefusing every option of a Chromium<select>(Chrome, Electron) with an empty option list: options are matched by the text Chromium reports as their value, the menu is opened when the closed popup publishes only its chosen option, and the menu is closed afterwards with Escape only when, afterAXCancel, it neither reports closing nor leaves the window list within 300 ms, as a Chromium menu that ignores the cancel does, instead of being left open with your keystrokes when no option matched (#19224 by @will-bogusz)setValueon macOS sliders, steppers and other number-valued controls failing withAXError(-25201): the value is now written as a number and read back, and a value the control clamped or rounded is reported with what it now reads (#19405 by @will-bogusz)setValueon a macOS control whose value is not settable but that steps (a SwiftUISlider) now refuses naming theAXIncrement/AXDecrementactions it offers and the number it reads, so the next call can step it (#19405 by @will-bogusz)UserIntervenedwith nothing sent if you are still at it near the run's timeout (#19296 by @will-bogusz)press()and backgroundel.click()reporting success on a control a person could not click, such as a hover toolbar's button after the pointer left, a hidden control, or one of a window that shows a sheet; they now fail before anything is sent and name what the window shows there (#19258 by @will-bogusz)computerkey press (Space or an up or down arrow) on a macOS popup button or menu button leaving its menu open, which took your typing once omp stopped sending your keys back to your app; the menu now closes before the call returns, and holding those keys on such a control is refused before anything is sent (#19179 by @will-bogusz)ax(),find(),el.children()and focus reads freezing a Qt app with a very long table (up to 45 s, over 1 GB) or knocking it off the accessibility bus, and showing nothing of a GTK 3 list over 65,536 cells; they read only the children they can use, andax()says how many it left out (#19036 by @will-bogusz)capabilities()reportscapture: falseand the capture error says why (#18353 by @will-bogusz)UserIntervenedwhile you typed in your own app and the target briefly took the keyboard, such as when a pop-up button's menu opened (#19515 by @will-bogusz)commitCreate()dropping a rejectingpre-commit/commit-msghook's stdout from the error; hook output now arrives interleaved instderrand the message, as undergit commit(#19449)PUT N*,CUT N*) anchored on the first#[...]attribute or///doc comment of a Rust item now cover the item, instead of failing as a single-line block (#19464 by @acfrazier)find(),ax()and element reads taking seconds in large apps: each element's independent accessibility questions now go out together, sofind()takes about half as long in Kate, Dolphin and LibreOffice Writer (#19521 by @will-bogusz)Removed
wayland-pipewireCargo feature andOMP_NATIVE_FEATURES, which only existed to pass it (#18353 by @will-bogusz)@oh-my-pi/omp-stats
Fixed
@oh-my-pi/pi-tui
Added
subagent1–subagent6tokens (#18987 by @will-bogusz)wrapoverflow forTableColumnandrenderTableRowLines(), which word-wraps such cells over extra lines and keeps their hard line breaks;Tablerenders them the same way (#19506)Changed
StatusAccountIdentitynow carries an optionalcredentialIdso hosts can attribute the usage segment to a credential-scoped (API-key) account, and cached usage invalidates when it changes (#19456) (#19492 by @thekitaev).Fixed
/copycopies the prose a mid-turnout://reply showed, including a reply a command sent in its result, instead of the reply tool's own "Shown to the user." result (#19551 by @bannert1337)<br>lines; cells now wrap within their column (#19506)What's Changed
New Contributors
Full Changelog: https://github.com/can1357/oh-my-pi/compare/v18.22.0...v18.22.1
Configuration
📅 Schedule: (UTC)
🚦 Automerge: Enabled.
♻ Rebasing: Whenever PR is behind base branch, or you tick the rebase/retry checkbox.
🔕 Ignore: Close this PR and you won't be reminded about this update again.
This PR has been generated by Mend Renovate CLI.