@mirasim/computer-use 0.1.0-rc.2 → 0.1.0-rc.4

package/dist/platform/browser/driver.js

@@ -1,3 +1,3 @@
-import{Type as i}from"typebox";import{SurfaceUnreadableError as d,text as h}from"../types.js";import{NO_TAB_OPEN_NOTE as c,NO_TAB_YET_NOTE as u}from"./messages.js";import{PASTE_VERB as b,SELECT_OPTION_VERB as p}from"../chromium/actions-value.js";import{SELECT_TEXT_VERB as w}from"../chromium/select-text.js";import{DIALOG_VERB as f,DRAG_VERB as g,UPLOAD_VERB as m}from"../chromium/actions.js";import{RUN_JS_CODE_DESC as v}from"../chromium/run-js.js";import{BrowserDriverVerbs as k}from"./driver-verbs.js";import{previewPage as y}from"../chromium/preview.js";import{withPageCapture as S}from"../chromium/shot.js";import{DOWNLOAD_VERB as T}from"./downloads.js";import{bridgeRequest as O,bridgeTabs as E,restoreBridgeTab as _}from"./bridge-tabs.js";const P=5e3,N=1500;async function R(l){if(!l.url())return"";let e;try{return await Promise.race([l.title().catch(()=>""),new Promise(t=>{e=setTimeout(()=>t(""),N),e.unref?.()})])}finally{e&&clearTimeout(e)}}class G extends k{unloadedSelection=null;async activeTarget(){return this.unloadedSelection?.id??super.activeTarget()}async ownershipResources(){return this.unloadedSelection?[{key:`chromium:${this.unloadedSelection.targetId}`,mode:"exclusive"}]:super.ownershipResources()}async prepareTarget(){const e=this.unloadedSelection;if(!e)return;const{page:t,viewport:s}=await _(this.browser,e.tabId,e.targetId,e.browserSession);this.tabIds.set(t,e.id),this.active=t,this.unloadedSelection=null,this.resetObservationState();const r=await t.evaluate("({ width: innerWidth, height: innerHeight })");if(r.width<=0||r.height<=0){const a=await this.perform("viewport",s??{});if(!a.ok)throw new d(`The restored tab has no usable viewport: ${a.note??"resize failed"}`);this.appendNote("Chrome restored this background tab with a zero-size viewport; applied a temporary viewport, cleared when this driver disconnects.")}this.appendNote("Chrome had unloaded this tab; restored the same tab in the background. Read its current state before continuing.")}page(){if(this.unloadedSelection)throw new d(`Tab ${this.unloadedSelection.id} has not been restored; no other page was observed or operated.`);return super.page()}adoptWhenReal(e){this.unloadedSelection||super.adoptWhenReal(e)}previewTarget(){return this.unloadedSelection||!this.browser?.isConnected()||!this.active||this.active.isClosed()?Promise.resolve(null):y(this.active,this.cdpUrl,"browser")}platform="browser";async turnEnded(e){if(this.attached||e.keepBrowserTabs===!0)return;const t=await this.ownPages().catch(()=>[]);for(const s of t)await s.close().catch(()=>{});t.length&&(this.closedItsLastTab=!0,this.resetObservationState())}extraVerbs(){return[{kind:"open_url",schema:i.Object({kind:i.Literal("open_url"),url:i.String({description:"Full URL to open. Only URLs that appeared on a page or were given in the task \u2014 never guess domains."})})},{kind:"activate",schema:i.Object({kind:i.Literal("activate")},{description:"Bring this browser tab and its window to the foreground. This changes the user's foreground window: explain the effect and obtain permission unless it is already authorized. focus() only selects the target. Embedded writes normally wait for the person to yield the window; activate() is an explicit foreground handoff, revoked by subsequent human input. Do not use it to bypass an input wait without the user handing over control."})},{kind:"back",schema:i.Object({kind:i.Literal("back")})},{kind:"reload",schema:i.Object({kind:i.Literal("reload")})},{kind:"new_tab",schema:i.Object({kind:i.Literal("new_tab"),url:i.Optional(i.String({description:"URL to open in it; omit for a blank tab"}))},{description:"Open a NEW tab of this session's own and make it the target \u2014 the current tab keeps its state, and both are listed as targets afterwards."})},{kind:"close_tab",schema:i.Object({kind:i.Literal("close_tab"),id:i.Optional(i.String({description:"Tab id from computer.surfaces(); omit to close the one you are on"})),all:i.Optional(i.Boolean({description:"Close every tab this session opened or claimed \u2014 the tidy-up at the end of a job. Never the user's own, whatever else is set. Refused together with id."})),confirmed:i.Optional(i.Boolean({description:"Only for a tab marked (user's): you ASKED the user in the conversation and they said yes. Never set it because closing would be convenient \u2014 the first refusal tells you to ask, and asking is the point."}))},{description:"Close a tab. This session's own always; one of the user's (marked (user's) in the listing) only after you asked them in the conversation and they agreed \u2014 then pass confirmed: true."})},{kind:"viewport",schema:i.Object({kind:i.Literal("viewport"),width:i.Optional(i.Number({description:"Viewport width in CSS pixels (default 1200)"})),height:i.Optional(i.Number({description:"Viewport height in CSS pixels \u2014 this is the one that decides how much of a long list exists (default 1700)"})),reset:i.Optional(i.Boolean({description:"Put the tab back to the real window size"}))},{description:"Give THIS tab a viewport of your own size, so a virtualized list (a feed, a search page, a long transcript) mounts more of itself per read. Set it BEFORE opening the page: rows mount against the viewport in effect at load."})},{kind:"run_js",readOnly:!0,schema:i.Object({kind:i.Literal("run_js"),code:i.String({description:v})},{description:"Ask the PAGE something no observation carries \u2014 read its data in bulk, its stored state, or an API it exposes. This is a stranger's site holding the user's logged-in session: read and compute here, act through the interaction verbs."})},g,f,m,T,p,w,b]}async observeElements(e={}){const t=await this.noTabNote();return t?this.noTabObservation(t,e.prefix):super.observeElements(e)}async observeImage(e={}){const t=await this.noTabNote();return t?this.noTabObservation(t,e.prefix):super.observeImage(e)}async noTabNote(){return this.closedItsLastTab?c:this.unloadedSelection?null:(await this.ensure(),this.active?null:u)}noTabObservation(e,t){const s=this.takePendingNote();return h(`${t?`${t}
+import{Type as i}from"typebox";import{SurfaceUnreadableError as d,text as h}from"../types.js";import{NO_TAB_OPEN_NOTE as c,NO_TAB_YET_NOTE as u}from"./messages.js";import{PASTE_VERB as b,SELECT_OPTION_VERB as p}from"../chromium/actions-value.js";import{SELECT_TEXT_VERB as w}from"../chromium/select-text.js";import{DIALOG_VERB as f,DRAG_VERB as g,UPLOAD_VERB as m}from"../chromium/actions.js";import{RUN_JS_CODE_DESC as v}from"../chromium/run-js.js";import{BrowserDriverVerbs as y}from"./driver-verbs.js";import{previewPage as k}from"../chromium/preview.js";import{withPageCapture as S}from"../chromium/shot.js";import{DOWNLOAD_VERB as T}from"./downloads.js";import{bridgeRequest as O,bridgeTabs as E,restoreBridgeTab as _}from"./bridge-tabs.js";const P=5e3,N=1500;async function R(l){if(!l.url())return"";let e;try{return await Promise.race([l.title().catch(()=>""),new Promise(t=>{e=setTimeout(()=>t(""),N),e.unref?.()})])}finally{e&&clearTimeout(e)}}class G extends y{unloadedSelection=null;async activeTarget(){return this.unloadedSelection?.id??super.activeTarget()}async ownershipResources(){return this.unloadedSelection?[{key:`chromium:${this.unloadedSelection.targetId}`,mode:"exclusive"}]:super.ownershipResources()}async prepareTarget(){const e=this.unloadedSelection;if(!e)return;const{page:t,viewport:s}=await _(this.browser,e.tabId,e.targetId,e.browserSession);this.tabIds.set(t,e.id),this.active=t,this.unloadedSelection=null,this.resetObservationState();const o=await t.evaluate("({ width: innerWidth, height: innerHeight })");if(o.width<=0||o.height<=0){const a=await this.perform("viewport",s??{});if(!a.ok)throw new d(`The restored tab has no usable viewport: ${a.note??"resize failed"}`);this.appendNote("Chrome restored this background tab with a zero-size viewport; applied a temporary viewport, cleared when this driver disconnects.")}this.appendNote("Chrome had unloaded this tab; restored the same tab in the background. Read its current state before continuing.")}page(){if(this.unloadedSelection)throw new d(`Tab ${this.unloadedSelection.id} has not been restored; no other page was observed or operated.`);return super.page()}adoptWhenReal(e){this.unloadedSelection||super.adoptWhenReal(e)}previewTarget(){return this.unloadedSelection||!this.browser?.isConnected()||!this.active||this.active.isClosed()?Promise.resolve(null):k(this.active,this.cdpUrl,"browser")}platform="browser";async turnEnded(e){if(this.attached||e.keepBrowserTabs===!0)return;const t=await this.ownPages().catch(()=>[]);for(const s of t)await s.close().catch(()=>{});t.length&&(this.closedItsLastTab=!0,this.resetObservationState())}extraVerbs(){return[{kind:"open_url",schema:i.Object({kind:i.Literal("open_url"),url:i.String({description:"Full URL to open. Only URLs that appeared on a page or were given in the task \u2014 never guess domains."})})},{kind:"activate",schema:i.Object({kind:i.Literal("activate")},{description:"Bring this browser tab and its window to the foreground. This changes the user's foreground window: explain the effect and obtain permission unless it is already authorized. focus() only selects the target. Embedded writes normally wait for the person to yield the window; activate() is an explicit foreground handoff, revoked by subsequent human input. Do not use it to bypass an input wait without the user handing over control."})},{kind:"back",schema:i.Object({kind:i.Literal("back")})},{kind:"reload",schema:i.Object({kind:i.Literal("reload")})},{kind:"new_tab",schema:i.Object({kind:i.Literal("new_tab"),url:i.Optional(i.String({description:"URL to open in it; omit for a blank tab"}))},{description:"Open a NEW tab of this session's own and make it the target \u2014 the current tab keeps its state, and both are listed as targets afterwards."})},{kind:"close_tab",schema:i.Object({kind:i.Literal("close_tab"),id:i.Optional(i.String({description:"Tab id from computer.surfaces(); omit to close the one you are on"})),all:i.Optional(i.Boolean({description:"Close every tab this session opened or claimed \u2014 the tidy-up at the end of a job. Never the user's own, whatever else is set. Refused together with id."})),confirmed:i.Optional(i.Boolean({description:"Only for a tab marked (user's): you ASKED the user in the conversation and they said yes. Never set it because closing would be convenient \u2014 the first refusal tells you to ask, and asking is the point."}))},{description:"Close a tab: the tabs you open outlive your turn and nothing closes them for you, so follow the tab policy the host states (if it states none, close your own once the job is done), and leave open a page that is waiting for the person (a sign-in, a security-key or 2FA prompt, a form for them to confirm). This session's own always; one of the user's (marked (user's) in the listing) only after you asked them in the conversation and they agreed \u2014 then pass confirmed: true."})},{kind:"viewport",schema:i.Object({kind:i.Literal("viewport"),width:i.Optional(i.Number({description:"Viewport width in CSS pixels (default 1200)"})),height:i.Optional(i.Number({description:"Viewport height in CSS pixels \u2014 this is the one that decides how much of a long list exists (default 1700)"})),reset:i.Optional(i.Boolean({description:"Put the tab back to the real window size"}))},{description:"Give THIS tab a viewport of your own size, so a virtualized list (a feed, a search page, a long transcript) mounts more of itself per read. Set it BEFORE opening the page: rows mount against the viewport in effect at load."})},{kind:"run_js",readOnly:!0,schema:i.Object({kind:i.Literal("run_js"),code:i.String({description:v})},{description:"Ask the PAGE something no observation carries \u2014 read its data in bulk, its stored state, or an API it exposes. This is a stranger's site holding the user's logged-in session: read and compute here, act through the interaction verbs."})},g,f,m,T,p,w,b]}async observeElements(e={}){const t=await this.noTabNote();return t?this.noTabObservation(t,e.prefix):super.observeElements(e)}async observeImage(e={}){const t=await this.noTabNote();return t?this.noTabObservation(t,e.prefix):super.observeImage(e)}async noTabNote(){return this.closedItsLastTab?c:this.unloadedSelection?null:(await this.ensure(),this.active?null:u)}noTabObservation(e,t){const s=this.takePendingNote();return h(`${t?`${t}
 `:""}${s?`NOTE: ${s}.
-`:""}${e}.`,{verdict:"unknown"})}staleRefCause(){return"it predates a navigation, a reload or a tab switch"}promptFragment(){return"Grounding: observations are text lists of interactive elements with refs, for the CURRENT viewport only. Act by ref from the latest observation. When what you want is NOT in the list: if you can name it, w.find('<name>') searches EVERY control on the page and hands back a handle \u2014 act on it directly, the click scrolls it into view, so do not scroll first. Scroll only to discover what you cannot name. Observations may show only changes since the previous one; refs stay valid across observations (until navigation). Use w.text() to read full page text."}async listTargets(){await this.ensure();const{unloaded:e,browserSession:t}=!this.tab&&!this.attached&&!this.managedProfileDir?await E(this.browser):{unloaded:[],browserSession:null};this.noteBridgeSession(t);const s=new Set(await this.ownPages()),r=await this.visiblePages(),a=await Promise.all(r.map(async o=>({id:this.tabId(o),kind:"tab",title:await R(o)||o.url(),active:!this.unloadedSelection&&o===this.active,own:s.has(o)})));for(const o of e){const n=`unloaded:${o.tabId}:${o.browserSession}`;a.push({id:n,kind:"tab",title:`${o.title||o.url} [unloaded \u2014 use this handle to restore in background]`,active:this.unloadedSelection?.id===n,own:!1,unloaded:!0})}return a}noteBridgeSession(e){const t=super.noteBridgeSession(e);return t&&(this.unloadedSelection=null),t}async actionResources(e,t){if(e!=="close_tab")return[];const s=typeof t.id=="string"?t.id.replace(/^browser\//,""):null,r=t.all?await this.ownPages():s?(await this.visiblePages()).filter(a=>this.tabId(a)===s):[this.page()];return Promise.all(r.filter(a=>!a.isClosed()).map(a=>this.pageOwnership(a)))}async focusTarget(e){await this.ensure();const t=/^unloaded:(\d+):(.*)$/.exec(e);if(t&&!this.tab&&!this.attached&&!this.managedProfileDir&&this.browser.version()==="mirasim-bridge"){const r=Number(t[1]),a=t[2],{targetId:o}=await O(this.browser,"Mirasim.tabTarget",{tabId:r,browserSession:a});return this.unloadedSelection={id:e,tabId:r,targetId:o,browserSession:a},this.closedItsLastTab=!1,{ok:!0,note:"Selected unloaded tab; restoration follows target ownership."}}let s={ok:!1,note:`no target with id ${e}; call computer.surfaces() for current ids`};for(const r of await this.visiblePages())if(!(r.isClosed()||this.tabId(r)!==e)){this.unloadedSelection=null,this.active=r,this.closedItsLastTab=!1,s={ok:!0};break}return this.journal?.publishAction("focus",{target:e},null,s),s}async close(){this.unloadedSelection=null,await this.downloads.close(),await this.journal?.drain(P);for(const e of this.overridden)e.isClosed()||await S(e,()=>this.cdpFor(e).then(t=>t.send("Emulation.clearDeviceMetricsOverride"))).catch(()=>{});if(this.overridden.clear(),!this.managedProfileDir&&!this.tab&&this.browser)for(const e of await this.ownPages().catch(()=>[]))e.url()==="about:blank"&&await e.close().catch(()=>{});await this.browser?.close().catch(()=>{}),this.browser=null,this.resetObservationState(),this.resetPageState(),await this.disconnected()}journalVerdict(e){this.journal?.publishVerdict(e)}renderHeadLines(e,t,s=!1,r){const a=super.renderHeadLines(e,t,s,r);if(e.url==="about:blank"&&e.elements.length===0&&!e.text){const o=t.some(n=>n.title!=="about:blank");a.push(o?"NOTE: this tab is blank \u2014 your page may be in another tab; computer.surfaces() lists it \u2014 act through its handle.":"NOTE: this tab is blank and nothing has been opened in it \u2014 navigate (openUrl({url})) instead of observing again.")}return a}async blankSurfaceHint(){try{return this.page().url()!=="about:blank"?"":(await this.ownPages()).some(t=>t!==this.page()&&t.url()!=="about:blank")?"the active tab is a blank page \u2014 your page may be in another tab; computer.surfaces() lists it \u2014 act through its handle":""}catch{return""}}}export{G as BrowserDriver};
+`:""}${e}.`,{verdict:"unknown"})}staleRefCause(){return"it predates a navigation, a reload or a tab switch"}promptFragment(){return"Grounding: observations are text lists of interactive elements with refs, for the CURRENT viewport only. Act by ref from the latest observation. When what you want is NOT in the list: if you can name it, w.find('<name>') searches EVERY control on the page and hands back a handle \u2014 act on it directly, the click scrolls it into view, so do not scroll first. Scroll only to discover what you cannot name. Observations may show only changes since the previous one; refs stay valid across observations (until navigation). Use w.text() to read full page text."}async listTargets(){await this.ensure();const{unloaded:e,browserSession:t}=!this.tab&&!this.attached&&!this.managedProfileDir?await E(this.browser):{unloaded:[],browserSession:null};this.noteBridgeSession(t);const s=new Set(await this.ownPages()),o=await this.visiblePages(),a=await Promise.all(o.map(async r=>({id:this.tabId(r),kind:"tab",title:await R(r)||r.url(),active:!this.unloadedSelection&&r===this.active,own:s.has(r)})));for(const r of e){const n=`unloaded:${r.tabId}:${r.browserSession}`;a.push({id:n,kind:"tab",title:`${r.title||r.url} [unloaded \u2014 use this handle to restore in background]`,active:this.unloadedSelection?.id===n,own:!1,unloaded:!0})}return a}noteBridgeSession(e){const t=super.noteBridgeSession(e);return t&&(this.unloadedSelection=null),t}async actionResources(e,t){if(e!=="close_tab")return[];const s=typeof t.id=="string"?t.id.replace(/^browser\//,""):null,o=t.all?await this.ownPages():s?(await this.visiblePages()).filter(a=>this.tabId(a)===s):[this.page()];return Promise.all(o.filter(a=>!a.isClosed()).map(a=>this.pageOwnership(a)))}async focusTarget(e){await this.ensure();const t=/^unloaded:(\d+):(.*)$/.exec(e);if(t&&!this.tab&&!this.attached&&!this.managedProfileDir&&this.browser.version()==="mirasim-bridge"){const o=Number(t[1]),a=t[2],{targetId:r}=await O(this.browser,"Mirasim.tabTarget",{tabId:o,browserSession:a});return this.unloadedSelection={id:e,tabId:o,targetId:r,browserSession:a},this.closedItsLastTab=!1,{ok:!0,note:"Selected unloaded tab; restoration follows target ownership."}}let s={ok:!1,note:`no target with id ${e}; call computer.surfaces() for current ids`};for(const o of await this.visiblePages())if(!(o.isClosed()||this.tabId(o)!==e)){this.unloadedSelection=null,this.active=o,this.closedItsLastTab=!1,s={ok:!0};break}return this.journal?.publishAction("focus",{target:e},null,s),s}async close(){this.unloadedSelection=null,await this.downloads.close(),await this.journal?.drain(P);for(const e of this.overridden)e.isClosed()||await S(e,()=>this.cdpFor(e).then(t=>t.send("Emulation.clearDeviceMetricsOverride"))).catch(()=>{});if(this.overridden.clear(),!this.managedProfileDir&&!this.tab&&this.browser)for(const e of await this.ownPages().catch(()=>[]))e.url()==="about:blank"&&await e.close().catch(()=>{});await this.browser?.close().catch(()=>{}),this.browser=null,this.resetObservationState(),this.resetPageState(),await this.disconnected()}journalVerdict(e){this.journal?.publishVerdict(e)}renderHeadLines(e,t,s=!1,o){const a=super.renderHeadLines(e,t,s,o);if(e.url==="about:blank"&&e.elements.length===0&&!e.text){const r=t.some(n=>n.title!=="about:blank");a.push(r?"NOTE: this tab is blank \u2014 your page may be in another tab; computer.surfaces() lists it \u2014 act through its handle.":"NOTE: this tab is blank and nothing has been opened in it \u2014 navigate (openUrl({url})) instead of observing again.")}return a}async blankSurfaceHint(){try{return this.page().url()!=="about:blank"?"":(await this.ownPages()).some(t=>t!==this.page()&&t.url()!=="about:blank")?"the active tab is a blank page \u2014 your page may be in another tab; computer.surfaces() lists it \u2014 act through its handle":""}catch{return""}}}export{G as BrowserDriver};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mirasim/computer-use",
-  "version": "0.1.0-rc.2",
+  "version": "0.1.0-rc.4",
   "description": "Computer use for any agent: an MCP server that drives browsers, macOS apps, Windows apps and Electron apps through their accessibility trees.",
   "type": "module",
   "keywords": [
@@ -70,9 +70,9 @@
     "access": "public"
   },
   "optionalDependencies": {
-    "@mirasim/computer-use-darwin-arm64": "0.1.0-rc.2",
-    "@mirasim/computer-use-darwin-x64": "0.1.0-rc.2",
-    "@mirasim/computer-use-win32-x64": "0.1.0-rc.2",
-    "@mirasim/computer-use-win32-arm64": "0.1.0-rc.2"
+    "@mirasim/computer-use-darwin-arm64": "0.1.0-rc.4",
+    "@mirasim/computer-use-darwin-x64": "0.1.0-rc.4",
+    "@mirasim/computer-use-win32-x64": "0.1.0-rc.4",
+    "@mirasim/computer-use-win32-arm64": "0.1.0-rc.4"
   }
 }