Jev Browser Use

[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_388":389,"_390":391},"routes/_app.x.$handle.$name",{"_113":112,"_38":52,"_122":99,"_33":123,"_13":14,"_124":-5,"_11":125,"_126":127,"_128":129,"_130":131,"_132":133,"_26":-5,"_134":135,"_136":137,"_138":31,"_139":140,"_141":68,"_142":-5,"_143":-5,"_144":145},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":18,"_19":-5},"project",{"_50":51,"_38":52,"_45":53,"_54":-5,"_55":31,"_56":57,"_58":59,"_60":61,"_62":63,"_19":64,"_65":66,"_67":68,"_69":-5,"_70":-5,"_71":72,"_73":74,"_75":76,"_77":78,"_79":80,"_81":-5,"_82":83,"_84":83,"_85":86,"_87":86,"_88":86,"_89":83,"_90":83,"_91":92,"_93":94,"_95":96,"_97":31},"version",33,"changes",[25],"aiChatInfo",{"_20":21,"_22":-5,"_23":-5,"_24":-5},"userId","aiChatMessageId","5008e97b-e5b2-484d-8412-8b98a2265843","aiChatThreadId","aiChatThreadUserId","aiChatThreadTitle",{"_26":27,"_28":29,"_30":31,"_32":31,"_33":34,"_35":34},"file",{"_36":37,"_38":39,"_40":41,"_42":44,"_45":46,"_47":14,"_48":49,"_30":31},"previous",{"_36":37,"_38":39,"_40":41,"_42":43},"deleted",false,"deletedHere","path","tests/decide.ts","latestPath","fileId","f9ab0b50-b209-11f1-b19e-1607ee4eb77e","name","decide.ts","fileType","script","content","/**\n * Exercises `decide()` against a fixture page state, with no browser involved.\n *\n * This is the cheap half of the loop to test and the expensive half to get\n * wrong: if stage 1 can't rank the goal link to the top of a Wikipedia article,\n * nothing downstream matters. The fixture deliberately contains the goal\n * (\"Eiffel Tower\") alongside ~20 plausible distractors, so the expected answer\n * is checkable by eye.\n *\n * Run with `run_file`. Costs two Jev calls (~$0.0004).\n */\nimport { decide, harvestTexts, type PageState } from \"../decide.ts\";\n\nconst page: PageState = {\n url: \"https://en.wikipedia.org/wiki/Rubber_duck\",\n title: \"Rubber duck - Wikipedia\",\n text:\n \"A rubber duck is a toy shaped like a stylised yellow-billed duck, \" +\n \"usually made of rubber or rubber-like material such as vinyl plastic. \" +\n \"It is a popular bathtub toy. The rubber duck has become a cultural icon, \" +\n \"appearing in art, music, and software engineering (see rubber duck debugging).\",\n candidates: [\n { n: 0, tag: \"a\", label: \"Toy\", href: \"/wiki/Toy\", x: 100, y: 200 },\n { n: 1, tag: \"a\", label: \"Duck\", href: \"/wiki/Duck\", x: 100, y: 220 },\n { n: 2, tag: \"a\", label: \"Rubber\", href: \"/wiki/Rubber\", x: 100, y: 240 },\n { n: 3, tag: \"a\", label: \"Plastic\", href: \"/wiki/Plastic\", x: 100, y: 260 },\n { n: 4, tag: \"a\", label: \"Bathtub\", href: \"/wiki/Bathtub\", x: 100, y: 280 },\n { n: 5, tag: \"a\", label: \"Cultural icon\", href: \"/wiki/Cultural_icon\", x: 100, y: 300 },\n { n: 6, tag: \"a\", label: \"Rubber duck debugging\", href: \"/wiki/Rubber_duck_debugging\", x: 100, y: 320 },\n { n: 7, tag: \"a\", label: \"Sesame Street\", href: \"/wiki/Sesame_Street\", x: 100, y: 340 },\n { n: 8, tag: \"a\", label: \"Ernie\", href: \"/wiki/Ernie\", x: 100, y: 360 },\n { n: 9, tag: \"a\", label: \"Vulcanization\", href: \"/wiki/Vulcanization\", x: 100, y: 380 },\n { n: 10, tag: \"a\", label: \"Charles Goodyear\", href: \"/wiki/Charles_Goodyear\", x: 100, y: 400 },\n { n: 11, tag: \"a\", label: \"Paris\", href: \"/wiki/Paris\", x: 100, y: 420 },\n { n: 12, tag: \"a\", label: \"Eiffel Tower\", href: \"/wiki/Eiffel_Tower\", x: 100, y: 440 },\n { n: 13, tag: \"a\", label: \"France\", href: \"/wiki/France\", x: 100, y: 460 },\n { n: 14, tag: \"a\", label: \"Tourism\", href: \"/wiki/Tourism\", x: 100, y: 480 },\n { n: 15, tag: \"a\", label: \"Gustave Eiffel\", href: \"/wiki/Gustave_Eiffel\", x: 100, y: 500 },\n { n: 16, tag: \"a\", label: \"World's fair\", href: \"/wiki/World%27s_fair\", x: 100, y: 520 },\n { n: 17, tag: \"a\", label: \"Landmark\", href: \"/wiki/Landmark\", x: 100, y: 540 },\n { n: 18, tag: \"a\", label: \"Iron\", href: \"/wiki/Iron\", x: 100, y: 560 },\n { n: 19, tag: \"a\", label: \"Tokyo Tower\", href: \"/wiki/Tokyo_Tower\", x: 100, y: 580 },\n { n: 20, tag: \"input\", label: \"Search Wikipedia\", editable: true, x: 100, y: 40 },\n ],\n};\n\nconst task = \"Reach the article about the Eiffel Tower, using only links on the page.\";\n\nconsole.log(\"harvestTexts:\", harvestTexts(task));\n\nconst started = performance.now();\nconst decision = await decide(page, task, []);\n\nconsole.log(`\\ndecided in ${decision.ms}ms (wall ${Math.round(performance.now() - started)}ms)`);\nconsole.log(`model ${decision.model} · ${decision.considered} candidates scored`);\nconsole.log(\n `usage ${decision.usage.input_tokens} in / ${decision.usage.output_tokens} out` +\n ` → $${(decision.usage.input_tokens * 0.042 / 1e6).toFixed(6)}`,\n);\n\nconsole.log(`\\naction: ${decision.action}`);\nconsole.log(`target: ${decision.target ? `\"${decision.target.label}\"` : \"(none)\"}`);\nconsole.log(`text: ${decision.text ?? \"(none)\"}`);\nconsole.log(\n `picked ${decision.probability?.toFixed(3) ?? \"?\"}` +\n ` · confidence ${decision.confidence?.toFixed(3) ?? \"?\"}` +\n ` · goal_reached ${decision.goalReached.toFixed(3)}`,\n);\n\nconsole.log(\"\\nstage 1, best first:\");\nfor (const { candidate, score } of decision.ranked.slice(0, 8)) {\n console.log(` ${score.toFixed(3)} ${candidate.label}`);\n}\n\n// The one assertion worth making by hand: on this fixture the goal link should\n// be the top-ranked candidate, and the run should not declare itself done on\n// the very first page.\nconst top = decision.ranked[0];\nconsole.log(`\\ntop-ranked is the goal link: ${top?.candidate.label === \"Eiffel Tower\"}`);\nconsole.log(`did not stop immediately: ${decision.action !== \"done\"}`);","/**\n * Exercises `decide()` against a fixture page state, with no browser involved.\n *\n * This is the cheap half of the loop to test and the expensive half to get\n * wrong: if stage 1 can't rank the goal link to the top of a Wikipedia article,\n * nothing downstream matters. The fixture deliberately contains the goal\n * (\"Eiffel Tower\") alongside ~20 plausible distractors, so the expected answer\n * is checkable by eye.\n *\n * Run with `run_file`. Costs two Jev calls (~$0.0004).\n */\nimport { decide, harvestTexts, type PageState } from \"../decide.ts\";\n\nconst page: PageState = {\n url: \"https://en.wikipedia.org/wiki/Rubber_duck\",\n title: \"Rubber duck - Wikipedia\",\n text:\n \"A rubber duck is a toy shaped like a stylised yellow-billed duck, \" +\n \"usually made of rubber or rubber-like material such as vinyl plastic. \" +\n \"It is a popular bathtub toy. The rubber duck has become a cultural icon, \" +\n \"appearing in art, music, and software engineering (see rubber duck debugging).\",\n candidates: [\n { n: 0, tag: \"a\", label: \"Toy\", href: \"/wiki/Toy\", x: 100, y: 200 },\n { n: 1, tag: \"a\", label: \"Duck\", href: \"/wiki/Duck\", x: 100, y: 220 },\n { n: 2, tag: \"a\", label: \"Rubber\", href: \"/wiki/Rubber\", x: 100, y: 240 },\n { n: 3, tag: \"a\", label: \"Plastic\", href: \"/wiki/Plastic\", x: 100, y: 260 },\n { n: 4, tag: \"a\", label: \"Bathtub\", href: \"/wiki/Bathtub\", x: 100, y: 280 },\n { n: 5, tag: \"a\", label: \"Cultural icon\", href: \"/wiki/Cultural_icon\", x: 100, y: 300 },\n { n: 6, tag: \"a\", label: \"Rubber duck debugging\", href: \"/wiki/Rubber_duck_debugging\", x: 100, y: 320 },\n { n: 7, tag: \"a\", label: \"Sesame Street\", href: \"/wiki/Sesame_Street\", x: 100, y: 340 },\n { n: 8, tag: \"a\", label: \"Ernie\", href: \"/wiki/Ernie\", x: 100, y: 360 },\n { n: 9, tag: \"a\", label: \"Vulcanization\", href: \"/wiki/Vulcanization\", x: 100, y: 380 },\n { n: 10, tag: \"a\", label: \"Charles Goodyear\", href: \"/wiki/Charles_Goodyear\", x: 100, y: 400 },\n { n: 11, tag: \"a\", label: \"Paris\", href: \"/wiki/Paris\", x: 100, y: 420 },\n { n: 12, tag: \"a\", label: \"Eiffel Tower\", href: \"/wiki/Eiffel_Tower\", x: 100, y: 440 },\n { n: 13, tag: \"a\", label: \"France\", href: \"/wiki/France\", x: 100, y: 460 },\n { n: 14, tag: \"a\", label: \"Tourism\", href: \"/wiki/Tourism\", x: 100, y: 480 },\n { n: 15, tag: \"a\", label: \"Gustave Eiffel\", href: \"/wiki/Gustave_Eiffel\", x: 100, y: 500 },\n { n: 16, tag: \"a\", label: \"World's fair\", href: \"/wiki/World%27s_fair\", x: 100, y: 520 },\n { n: 17, tag: \"a\", label: \"Landmark\", href: \"/wiki/Landmark\", x: 100, y: 540 },\n { n: 18, tag: \"a\", label: \"Iron\", href: \"/wiki/Iron\", x: 100, y: 560 },\n { n: 19, tag: \"a\", label: \"Tokyo Tower\", href: \"/wiki/Tokyo_Tower\", x: 100, y: 580 },\n { n: 20, tag: \"input\", label: \"Search Wikipedia\", editable: true, x: 100, y: 40 },\n ],\n};\n\nconst task = \"Reach the article about the Eiffel Tower, using only links on the page.\";\n\nconsole.log(\"harvestTexts:\", harvestTexts(task));\n\nconst started = performance.now();\nconst decision = await decide(page, task, []);\n\nconsole.log(`\\ndecided in ${decision.ms}ms (wall ${Math.round(performance.now() - started)}ms)`);\nconsole.log(`model ${decision.model} · ${decision.considered} candidates scored`);\nconsole.log(\n `usage ${decision.usage.input_tokens} in / ${decision.usage.output_tokens} out` +\n ` → $${(decision.usage.input_tokens * 0.042 / 1e6).toFixed(6)}`,\n);\n\nconsole.log(`\\naction: ${decision.action}`);\nconsole.log(`target: ${decision.target ? `\"${decision.target.label}\"` : \"(none)\"}`);\nconsole.log(`text: ${decision.text ?? \"(none)\"}`);\nconsole.log(\n `picked ${decision.probability?.toFixed(3) ?? \"?\"}` +\n ` · confidence ${decision.confidence?.toFixed(3) ?? \"?\"}` +\n ` · goal_reached ${decision.goalReached.toFixed(3)}`,\n);\n\nconsole.log(\"\\nstage 1, best first (score = goal + (1-goal)*closer):\");\nfor (const { candidate, goal, closer, score } of decision.ranked.slice(0, 8)) {\n console.log(\n ` ${score.toFixed(3)} goal ${goal.toFixed(2)} closer ${closer.toFixed(2)} ${candidate.label}`,\n );\n}\n\n// The one assertion worth making by hand: on this fixture the goal link should\n// be the top-ranked candidate, and the run should not declare itself done on\n// the very first page.\nconst top = decision.ranked[0];\nconsole.log(`\\ntop-ranked is the goal link: ${top?.candidate.label === \"Eiffel Tower\"}`);\nconsole.log(`did not stop immediately: ${decision.action !== \"done\"}`);","createdAt",["D",1789589208539],"branchVersion","branchId","789b925f-1f04-4496-b61e-c55aa361585d","id","3c438ea3-80e8-4cac-9708-6511f055598d","jev-browser-use",["D",1789588833583],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","restricted","description","Describe a task; Jev decides each step in a KERNEL cloud browser","imageUrl","https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/601425b5-d08c-486a-cae2-cbfbdf0cc801/square512?exp=1789617338&sig=2ead8d8c519518e683133ec8521dd1ecf259ec89747e28d5016d932e5cb051d3","a0bf3b31-15a5-4d5c-880e-4b1e22c9bc18","pinnedFileId","01a0abce-8571-76e1-8061-dc657cf63b40","pinPreview",true,"pinMigrated","screenshotVersion","likeCount",0,"orgId","b2fd4880-5c2e-11f0-9801-5764bf98a432","tursoName","3c438ea380e84cac97086511f055598d","tursoHostname","3c438ea380e84cac97086511f055598d-val-town.aws-us-east-2.turso.io","tursoMeasuredAt",["D",1789590099209],"pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_113":112,"_121":117},"org",{"_113":112,"_114":115,"_116":117,"_19":-5,"_118":119,"_120":68},"branch",{"_13":98,"_38":99,"_50":49,"_100":51,"_101":102,"_103":104,"_105":57,"_106":107,"_108":109,"_110":52,"_111":112},"isOwned",282,"main","projectId","forkedBranchId","e9a8eeca-33f2-423b-8459-435fd862f66d","forkedBranchVersion",74,"forkedProjectPrivacy","forkedProjectId","c242175a-bb41-49a1-ae69-d9c2930a22f1","forkedProjectOrgId","e81e7335-6d4c-4764-a902-d0962bbfa9f7","projectName","userHandle","kernel","handle","displayName","KERNEL","avatar","https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/772b82b9-bcf2-4ee8-2b88-3bbabc7f9801/square512?exp=1789617338&sig=8036a510406896f080b04062f07d06bd7b73dbf8658146551bbca65814f60dcc","tier","business","canRestrictVals","profileImageUrl","branchName","","adoptOrgId",{"_50":51,"_38":52,"_45":383,"_54":-5,"_55":31,"_56":57,"_58":59,"_60":61,"_62":63,"_19":64,"_65":66,"_67":68,"_69":-5,"_70":-5,"_71":72,"_73":74,"_75":76,"_77":78,"_79":384,"_81":-5,"_82":83,"_84":83,"_85":86,"_87":86,"_88":86,"_89":83,"_90":83,"_91":385,"_93":386,"_95":387,"_97":31},"tree",[261,270,272,273,274,275,276,277,278,279,280,281],"files",[261,262,263,264,265,266,267,268,269,270,271,272,273,274,275,276,277,278,279,280,281],"counts",{"_251":72,"_252":172,"_253":72,"_254":255,"_256":257},"contents",[156,157,158,159,160,161,162,163,164,165,166,167],"upstreamProject",{"_50":107,"_38":152,"_60":153,"_62":-5,"_56":57,"_91":154},"upstreamBranch",{"_50":102,"_45":146,"_147":148,"_38":99,"_100":107,"_13":104,"_101":-5,"_103":-5,"_149":150,"_54":-5,"_97":31,"_56":57,"_111":151,"_73":109,"_110":152},"emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},["D",1789580290788],"updatedAt",["D",1789591366294],"lockfileKey","1789591367061-esm.town_v_stevekrouse_typesafe-typewriter@74-main","stevekrouse","typesafe-typewriter","16 TypeSafe gut-checks on your writing, on every keystroke",{"_113":151,"_121":155},"https://images.clerk.dev/uploaded/img_2PqHa2Gsy93xQrjh2w78Xu0cChW.jpeg",{"_50":246,"_38":247,"_36":248,"_171":-5,"_48":49,"_47":172,"_173":68,"_45":249,"_13":72,"_40":244,"_42":-5,"_30":31,"_176":-5,"_33":250},{"_50":240,"_38":241,"_36":242,"_171":-5,"_48":49,"_47":172,"_173":68,"_45":243,"_13":72,"_40":244,"_42":-5,"_30":31,"_176":-5,"_33":245},{"_50":233,"_38":234,"_36":235,"_171":-5,"_48":49,"_47":236,"_173":31,"_45":237,"_13":172,"_40":26,"_42":238,"_30":31,"_176":-5,"_33":239},{"_50":227,"_38":228,"_36":229,"_171":-5,"_48":49,"_47":172,"_173":31,"_45":230,"_13":72,"_40":26,"_42":231,"_30":31,"_176":-5,"_33":232},{"_50":220,"_38":221,"_36":222,"_171":-5,"_48":49,"_47":223,"_173":31,"_45":224,"_13":72,"_40":41,"_42":225,"_30":31,"_176":-5,"_33":226},{"_50":212,"_38":213,"_36":214,"_171":-5,"_48":49,"_47":215,"_173":31,"_45":216,"_13":217,"_40":41,"_42":218,"_30":31,"_176":-5,"_33":219},{"_50":205,"_38":39,"_36":206,"_171":-5,"_48":49,"_47":207,"_173":31,"_45":208,"_13":209,"_40":41,"_42":210,"_30":31,"_176":-5,"_33":211},{"_50":198,"_38":199,"_36":66,"_171":-5,"_48":49,"_47":200,"_173":31,"_45":201,"_13":172,"_40":202,"_42":203,"_30":31,"_176":52,"_33":204},{"_50":191,"_38":192,"_36":193,"_171":-5,"_48":49,"_47":194,"_173":31,"_45":195,"_13":172,"_40":41,"_42":196,"_30":31,"_176":-5,"_33":197},{"_50":185,"_38":186,"_36":187,"_171":-5,"_48":49,"_47":172,"_173":31,"_45":188,"_13":72,"_40":41,"_42":189,"_30":31,"_176":-5,"_33":190},{"_50":178,"_38":179,"_36":180,"_171":-5,"_48":49,"_47":181,"_173":31,"_45":182,"_13":72,"_40":41,"_42":183,"_30":31,"_176":-5,"_33":184},{"_50":168,"_38":169,"_36":170,"_171":-5,"_48":49,"_47":172,"_173":68,"_45":174,"_13":72,"_40":41,"_42":175,"_30":31,"_176":-5,"_33":177},"8b664cdb-aa06-4f54-a1b8-2b0844cc483e","typesafe.ts","01a0abce-8571-76e1-8061-e98d955a6ea6","parentId",1,"latest",["D",1789588833583],"/**\n * Minimal TypeSafe (System One) client.\n * Docs: https://docs.typesafe.ai/api\n *\n * One POST, one state, N typed questions, N typed answers.\n */\n\nconst ENDPOINT = \"https://api.typesafe.ai/v1/systemone\";\n\n/** Override with the TYPESAFE_MODEL env var to swap models without a deploy. */\nexport const MODEL = Deno.env.get(\"TYPESAFE_MODEL\") ?? \"jev-latest\";\n\nexport type NoulQuestion = {\n type: \"noul\";\n instructions: string;\n criteria?: { true?: string; false?: string };\n};\n\nexport type ChoiceQuestion = {\n type: \"choice\";\n instructions: string;\n criteria: Record;\n};\n\nexport type ScoreQuestion = {\n type: \"score\";\n instructions: string;\n criteria: string[];\n};\n\nexport type Question = NoulQuestion | ChoiceQuestion | ScoreQuestion;\n\nexport type NoulAnswer = { type: \"noul\"; noul: number };\n\nexport type ChoiceAnswer = {\n type: \"choice\";\n choice: string;\n probabilities: Record;\n confidence: number;\n};\n\nexport type ScoreAnswer = {\n type: \"score\";\n score: number;\n legend: Record;\n probabilities: Record;\n confidence: number;\n};\n\nexport type Answer = NoulAnswer | ChoiceAnswer | ScoreAnswer;\n\nexport type SystemOneResponse = {\n model: string;\n answers: Record;\n usage: { input_tokens: number; output_tokens: number };\n};\n\n/** The env var this client needs. Exported so the UI can name it. */\nexport const API_KEY_VAR = \"TYPESAFE_API_KEY\";\n\n/** Where to get one. */\nexport const API_KEY_CONSOLE_URL = \"https://console.typesafe.ai/settings/keys\";\n\n/**\n * Whether this val has a key to call TypeSafe with. Trims rather than merely\n * testing presence: a var that exists holding an empty string is the normal\n * state of a fresh remix, and is just as unusable as a missing one.\n */\nexport function isConfigured(): boolean {\n return !!Deno.env.get(API_KEY_VAR)?.trim();\n}\n\n/** Ask a batch of typed questions about a single piece of state. */\nexport async function ask(\n state: unknown,\n questions: Record,\n { timeoutMs = 10_000 }: { timeoutMs?: number } = {},\n): Promise {\n const apiKey = Deno.env.get(API_KEY_VAR)?.trim();\n if (!apiKey) throw new Error(`${API_KEY_VAR} is not set on this val`);\n\n const res = await fetch(ENDPOINT, {\n method: \"POST\",\n headers: {\n Authorization: `Bearer ${apiKey}`,\n \"Content-Type\": \"application/json\",\n },\n body: JSON.stringify({ state, model: MODEL, questions }),\n signal: AbortSignal.timeout(timeoutMs),\n });\n\n if (!res.ok) {\n const body = await res.text();\n throw new Error(`TypeSafe ${res.status}: ${body.slice(0, 300)}`);\n }\n\n return await res.json() as SystemOneResponse;\n}","label",[169],"2b7d95d5-b229-41cb-a219-7cb28ebc1702","sessions.ts","979add00-b209-11f1-8f64-1607ee4eb77e",14,["D",1789588969304],"/**\n * The registry of live browsers, and the clock that bills them.\n *\n * A Kernel browser keeps running (and billing) after the HTTP request that\n * created it returns, so \"who owes what\" can't live in the request that made\n * it. One row per session in SQLite, and `accountTime` moves the elapsed\n * seconds from the session onto the budget ledger. Called on every step and\n * once more when the session closes, so the ledger tracks a browser that's\n * sitting idle in a tab as well as one mid-run.\n */\nimport { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\nimport { chargeKernelSeconds } from \"./budget.ts\";\n\n/** How many browsers this demo will run at once. */\nexport const MAX_LIVE_SESSIONS = 2;\n\nexport type SessionStatus = \"live\" | \"done\" | \"stopped\" | \"failed\";\n\nexport type Session = {\n id: string;\n kernelId: string;\n task: string;\n startUrl: string;\n startedAt: string;\n /** Seconds already moved onto the budget ledger for this session. */\n chargedSeconds: number;\n steps: number;\n status: SessionStatus;\n};\n\nconst SCHEMA = `CREATE TABLE IF NOT EXISTS sessions (\n id TEXT PRIMARY KEY,\n kernel_id TEXT NOT NULL,\n task TEXT NOT NULL,\n start_url TEXT NOT NULL DEFAULT '',\n started_at TEXT NOT NULL,\n charged_seconds REAL NOT NULL DEFAULT 0,\n steps INTEGER NOT NULL DEFAULT 0,\n status TEXT NOT NULL DEFAULT 'live'\n)`;\n\nlet schema: Promise | null = null;\n\nfunction init(): Promise {\n if (!schema) {\n schema = sqlite.execute(SCHEMA).catch((err) => {\n schema = null;\n throw err;\n });\n }\n return schema;\n}\n\nfunction toSession(row: Record): Session {\n return {\n id: String(row.id),\n kernelId: String(row.kernel_id),\n task: String(row.task),\n startUrl: String(row.start_url ?? \"\"),\n startedAt: String(row.started_at),\n chargedSeconds: Number(row.charged_seconds ?? 0),\n steps: Number(row.steps ?? 0),\n status: String(row.status) as SessionStatus,\n };\n}\n\nexport async function createSession(\n kernelId: string,\n task: string,\n startUrl: string,\n): Promise {\n await init();\n const session: Session = {\n id: crypto.randomUUID(),\n kernelId,\n task,\n startUrl,\n startedAt: new Date().toISOString(),\n chargedSeconds: 0,\n steps: 0,\n status: \"live\",\n };\n await sqlite.execute({\n sql: `INSERT INTO sessions\n (id, kernel_id, task, start_url, started_at, charged_seconds, steps, status)\n VALUES (?, ?, ?, ?, ?, 0, 0, 'live')`,\n args: [\n session.id,\n session.kernelId,\n session.task,\n session.startUrl,\n session.startedAt,\n ],\n });\n return session;\n}\n\nexport async function getSession(id: string): Promise {\n await init();\n const { rows } = await sqlite.execute({\n sql: \"SELECT * FROM sessions WHERE id = ?\",\n args: [id],\n });\n return rows[0] ? toSession(rows[0] as Record) : null;\n}\n\n/** Live browsers, for the concurrency guard. */\nexport async function liveSessionCount(): Promise {\n await init();\n const { rows } = await sqlite.execute(\n \"SELECT COUNT(*) AS n FROM sessions WHERE status = 'live'\",\n );\n return Number(rows[0]?.n ?? 0);\n}\n\n/**\n * Move this session's elapsed seconds onto the budget ledger, exactly once\n * each. Returns the seconds charged in this call.\n *\n * The subtraction is why `charged_seconds` exists: a browser is billed for its\n * whole life, so charging `now - startedAt` every step would charge the same\n * second twice.\n */\nexport async function accountTime(session: Session): Promise {\n const elapsed = (Date.now() - new Date(session.startedAt).getTime()) / 1000;\n const delta = Math.max(0, elapsed - session.chargedSeconds);\n if (delta <= 0) return 0;\n await chargeKernelSeconds(delta);\n await sqlite.execute({\n sql: \"UPDATE sessions SET charged_seconds = ? WHERE id = ?\",\n args: [session.chargedSeconds + delta, session.id],\n });\n return delta;\n}\n\nexport async function bumpSteps(id: string): Promise {\n await sqlite.execute({\n sql: \"UPDATE sessions SET steps = steps + 1 WHERE id = ?\",\n args: [id],\n });\n}\n\nexport async function closeSession(\n id: string,\n status: SessionStatus,\n): Promise {\n await sqlite.execute({\n sql: \"UPDATE sessions SET status = ? WHERE id = ?\",\n args: [status, id],\n });\n}",[179],"808e7952-d3ab-4a2b-81a0-d514e5e77a0f","rate-limit.ts","01a0abce-8571-76e1-8061-e5086f72225b",["D",1789588833583],"/**\n * Best-effort in-memory rate limit so a public demo can't drain the API budget.\n * Each isolate keeps its own window, so the effective limit is a bit looser\n * than the numbers below. Good enough for a demo; swap for sqlite if it matters.\n */\n\nconst hits = new Map();\nlet globalHits: number[] = [];\n\nexport type LimitResult =\n | { ok: true }\n | { ok: false; reason: \"per-key\" | \"global\"; retryAfterMs: number };\n\nexport function checkRateLimit(\n key: string,\n { perKey = 600, global = 6000, windowMs = 60_000 } = {},\n): LimitResult {\n const now = Date.now();\n const cutoff = now - windowMs;\n\n const keyHits = (hits.get(key) ?? []).filter((t) => t > cutoff);\n globalHits = globalHits.filter((t) => t > cutoff);\n\n if (keyHits.length >= perKey) {\n return { ok: false, reason: \"per-key\", retryAfterMs: keyHits[0] + windowMs - now };\n }\n if (globalHits.length >= global) {\n return { ok: false, reason: \"global\", retryAfterMs: globalHits[0] + windowMs - now };\n }\n\n keyHits.push(now);\n globalHits.push(now);\n hits.set(key, keyHits);\n\n // Crude memory guard for a long-lived isolate.\n if (hits.size > 5000) hits.clear();\n\n return { ok: true };\n}",[186],"ca8de7ef-5622-4ebd-91af-23a3ca69aedd","kernel.ts","01a0abcf-a6b5-7413-8f55-6a07a5e01800",16,["D",1789589027405],"/**\n * Kernel client — the cloud Chromium the agent drives.\n *\n * Adapted from @stevekrouse/browser-tester's `lib/browser.ts`, which is the\n * same API base, Bearer auth, error handling, and replay helpers. What's added\n * here is what browser *use* needs rather than browser *testing*:\n *\n * - `createBrowser` asks for kiosk mode and a fixed viewport, and returns the\n * live view URL the demo embeds in an iframe.\n * - The computer controls (real OS mouse and keyboard) that the agent acts\n * with, so the cursor moves along a Bézier path and typing is human-paced.\n *\n * Docs: https://www.kernel.sh/docs\n */\nconst API = \"https://api.onkernel.com\";\n\n/**\n * Read the key lazily, not at module load. A missing key should be a 503 the\n * page can explain, not an import-time throw that takes the whole val down.\n */\nfunction key(): string {\n const k = Deno.env.get(API_KEY_VAR)?.trim();\n if (!k) throw new Error(`${API_KEY_VAR} is not set on this val`);\n return k;\n}\n\n/** The env var this client needs. Exported so the UI can name it. */\nexport const API_KEY_VAR = \"KERNEL_API_KEY\";\n\n/** Where to get one. */\nexport const API_KEY_CONSOLE_URL = \"https://dashboard.onkernel.com/settings/api-keys\";\n\nexport function isConfigured(): boolean {\n return !!Deno.env.get(API_KEY_VAR)?.trim();\n}\n\nexport type KernelResponse = {\n status: number;\n headers: Headers;\n text: () => Promise;\n arrayBuffer: () => Promise;\n};\n\nexport async function kernelFetch(\n path: string,\n method = \"GET\",\n body?: unknown,\n): Promise {\n const res = await fetch(API + path, {\n method,\n headers: {\n Authorization: `Bearer ${key()}`,\n ...(body ? { \"Content-Type\": \"application/json\" } : {}),\n },\n body: body ? JSON.stringify(body) : undefined,\n });\n return {\n status: res.status,\n headers: res.headers,\n text: () => res.text(),\n arrayBuffer: () => res.arrayBuffer(),\n };\n}\n\n/** POST that expects a JSON body back, with the status in the error. */\nasync function postJson(path: string, body?: unknown): Promise {\n const r = await kernelFetch(path, \"POST\", body);\n const text = await r.text();\n if (r.status >= 300) {\n throw new Error(`Kernel ${path} failed (${r.status}): ${text.slice(0, 300)}`);\n }\n return text ? JSON.parse(text) as T : ({} as T);\n}\n\nexport type KernelBrowser = {\n session_id: string;\n browser_live_view_url: string;\n cdp_ws_url?: string;\n};\n\n/**\n * Create a browser session.\n *\n * Kiosk mode is not cosmetic. It removes the tab strip and toolbar, so the\n * window *is* the page viewport and `getBoundingClientRect()` coordinates are\n * already screen coordinates — which is what the computer controls click. With\n * browser chrome in the way every click would land a toolbar-height too high.\n * It also makes the live view a clean full-bleed rectangle for the video.\n *\n * `timeout_seconds` is Kernel's idle auto-delete ceiling, and it's the thing\n * that bounds cost when someone walks away from an open tab.\n */\nexport async function createBrowser(\n opts: {\n name?: string;\n startUrl?: string;\n width?: number;\n height?: number;\n idleTimeoutSec?: number;\n } = {},\n): Promise {\n const {\n name,\n startUrl,\n width = 1024,\n height = 768,\n idleTimeoutSec = 240,\n } = opts;\n return await postJson(\"/browsers\", {\n headless: false,\n kiosk_mode: true,\n viewport: { width, height },\n timeout_seconds: idleTimeoutSec,\n ...(name ? { name } : {}),\n // Boot straight at the start URL rather than the new-tab home page.\n ...(startUrl ? { start_url: startUrl } : {}),\n });\n}\n\n/** Execute Playwright/TypeScript against the browser, with page/context/browser in scope. */\nexport async function runPlaywright(\n sessionId: string,\n code: string,\n timeoutSec = 45,\n): Promise<{ success: boolean; result?: unknown; error?: string; stdout?: string }> {\n return await postJson(`/browsers/${sessionId}/playwright/execute`, {\n code,\n timeout_sec: timeoutSec,\n });\n}\n\n// ── Computer controls ───────────────────────────────────────────\n// A real OS mouse and keyboard, not synthetic DOM events. `smooth` is Kernel's\n// default and draws a human-like Bézier path, which is most of why the demo is\n// watchable rather than a series of jumps.\n\n/** Move the cursor. Smooth by default, so the path is visible in the live view. */\nexport async function moveMouse(\n sessionId: string,\n x: number,\n y: number,\n durationMs?: number,\n): Promise {\n await postJson(`/browsers/${sessionId}/computer/move_mouse`, {\n x,\n y,\n smooth: true,\n ...(durationMs ? { duration_ms: durationMs } : {}),\n });\n}\n\nexport async function clickMouse(\n sessionId: string,\n x: number,\n y: number,\n): Promise {\n await postJson(`/browsers/${sessionId}/computer/click_mouse`, { x, y });\n}\n\n/** Type literal text, human-paced. */\nexport async function typeText(sessionId: string, text: string): Promise {\n await postJson(`/browsers/${sessionId}/computer/type_text`, { text });\n}\n\nexport async function pressKey(sessionId: string, keys: string[]): Promise {\n await postJson(`/browsers/${sessionId}/computer/press_key`, { keys });\n}\n\n/** Capture a screenshot; returns raw PNG bytes. */\nexport async function screenshot(sessionId: string): Promise {\n const r = await kernelFetch(`/browsers/${sessionId}/computer/screenshot`, \"POST\", {});\n if (r.status !== 200) {\n throw new Error(`screenshot failed (${r.status}): ${await r.text()}`);\n }\n return r.arrayBuffer();\n}\n\n/** Delete a browser session. Frees the concurrency slot and stops the meter. */\nexport async function destroyBrowser(sessionId: string): Promise {\n const r = await kernelFetch(`/browsers/${sessionId}`, \"DELETE\");\n if (r.status !== 204 && r.status !== 200) {\n throw new Error(`destroyBrowser failed (${r.status}): ${await r.text()}`);\n }\n}\n\nexport function sleep(ms: number): Promise {\n return new Promise((resolve) => setTimeout(resolve, ms));\n}",[192],"bf249aa8-130d-441a-a374-20cd5c2ddf69","index.ts",18,["D",1789589059287],"http","import { hotReload, parseVal, serveImmutableFile } from \"https://esm.town/v/std/utils/index.ts\";\nimport { Hono } from \"npm:hono\";\nimport { canStartSession, getBudget, noteSession } from \"./budget.ts\";\nimport { act, describeAction, perceive } from \"./browser-use.ts\";\nimport { BudgetSpentError, decide } from \"./decide.ts\";\nimport { Root } from \"./frontend/root.tsx\";\nimport {\n API_KEY_CONSOLE_URL as KERNEL_CONSOLE_URL,\n API_KEY_VAR as KERNEL_VAR,\n createBrowser,\n destroyBrowser,\n isConfigured as kernelConfigured,\n} from \"./kernel.ts\";\nimport { checkRateLimit } from \"./rate-limit.ts\";\nimport {\n accountTime,\n bumpSteps,\n closeSession,\n createSession,\n getSession,\n liveSessionCount,\n MAX_LIVE_SESSIONS,\n type Session,\n type SessionStatus,\n} from \"./sessions.ts\";\nimport {\n API_KEY_CONSOLE_URL as TYPESAFE_CONSOLE_URL,\n API_KEY_VAR as TYPESAFE_VAR,\n isConfigured as typesafeConfigured,\n} from \"./typesafe.ts\";\n\n/** How many steps one run may take. Each step is two Jev calls and ~4s of browser. */\nconst MAX_STEPS = 12;\n\n/** Task and URL caps, so a pasted novel can't blow the Jev request budget. */\nconst MAX_TASK_CHARS = 500;\n\n/** Where a run starts when you don't say. The blog's own wikiracing example. */\nconst DEFAULT_START_URL = \"https://en.wikipedia.org/wiki/Rubber_duck\";\n\n/** One-click demos. Wikiracing is TypeSafe's own showcase task for Jev. */\nconst PRESETS = [\n {\n label: \"Wikiracing\",\n task: \"Reach the article about the Eiffel Tower, using only links on the page.\",\n startUrl: \"https://en.wikipedia.org/wiki/Rubber_duck\",\n },\n {\n label: \"Hacker News\",\n task: \"Open the comments for the top story on the front page.\",\n startUrl: \"https://news.ycombinator.com\",\n },\n {\n label: \"Val Town\",\n task: \"Find the pricing page and tell me what the cheapest paid plan costs.\",\n startUrl: \"https://www.val.town\",\n },\n];\n\nconst app = new Hono();\n\n/**\n * What the browser needs in order to explain a val that's missing a key:\n * whether the gap exists, and where to fill it. The env var editor URL is\n * derived from the val itself, so a fork sends people to its own settings.\n *\n * Nothing secret goes over the wire — only whether the gap exists.\n */\nfunction setupState() {\n const val = parseVal().links.self.val;\n const one = (keyVar: string, consoleUrl: string, configured: boolean) => ({\n configured,\n keyVar,\n envUrl: `${val}/environment-variables?key=${keyVar}`,\n consoleUrl,\n });\n return {\n typesafe: one(TYPESAFE_VAR, TYPESAFE_CONSOLE_URL, typesafeConfigured()),\n kernel: one(KERNEL_VAR, KERNEL_CONSOLE_URL, kernelConfigured()),\n };\n}\n\nfunction ipOf(c: { req: { header: (k: string) => string | undefined } }): string {\n return c.req.header(\"x-forwarded-for\")?.split(\",\")[0]?.trim() ||\n c.req.header(\"x-real-ip\") || \"unknown\";\n}\n\n/** Accept a bare host too, so \"val.town\" works. Empty falls back to the default. */\nfunction normalizeUrl(raw: unknown): string {\n const s = typeof raw === \"string\" ? raw.trim() : \"\";\n if (!s) return DEFAULT_START_URL;\n const withScheme = /^https?:\\/\\//i.test(s) ? s : `https://${s}`;\n try {\n return new URL(withScheme).toString();\n } catch {\n return DEFAULT_START_URL;\n }\n}\n\n/** Stop billing: move the last of the elapsed time onto the ledger, then delete. */\nasync function finish(session: Session, status: SessionStatus): Promise {\n if (session.status !== \"live\") return;\n await accountTime(session).catch((err) => console.error(\"accountTime failed:\", err));\n await closeSession(session.id, status);\n await destroyBrowser(session.kernelId).catch((err) =>\n console.error(\"destroyBrowser failed:\", err)\n );\n}\n\n// ── Shell ───────────────────────────────────────────────────────\n// Never cached, so it can stamp the current version into each asset URL.\napp.get(\"/\", (c) => c.html(Root()));\n\n// The version-addressed asset URLs Root() stamps, cached immutably.\napp.get(\"/__immutable/*\", (c) => serveImmutableFile(c.req.path));\n\napp.get(\"/source\", (c) => c.redirect(parseVal().links.self.val));\n\n// ── API ─────────────────────────────────────────────────────────\napp.get(\"/api/config\", async (c) =>\n c.json({\n setup: setupState(),\n presets: PRESETS,\n maxSteps: MAX_STEPS,\n maxLiveSessions: MAX_LIVE_SESSIONS,\n budget: await getBudget().catch(() => null),\n }));\n\n/** Start a run: one browser, one task. */\napp.post(\"/api/session\", async (c) => {\n if (!typesafeConfigured() || !kernelConfigured()) {\n return c.json({\n error: \"this val is missing an API key\",\n reason: \"unconfigured\",\n setup: setupState(),\n }, 503);\n }\n\n const body = await c.req.json().catch(() => null);\n const task = typeof body?.task === \"string\"\n ? body.task.trim().slice(0, MAX_TASK_CHARS)\n : \"\";\n if (!task) return c.json({ error: \"describe a task first\" }, 400);\n const startUrl = normalizeUrl(body?.startUrl);\n\n // A browser is the expensive thing here, so this limit is much tighter than\n // a per-keystroke demo's would be: ~$0.008/minute each.\n const limit = checkRateLimit(ipOf(c), {\n perKey: 6,\n global: 60,\n windowMs: 60_000,\n });\n if (!limit.ok) {\n return c.json({\n error: \"rate limited\",\n reason: limit.reason,\n retryAfterMs: limit.retryAfterMs,\n }, 429);\n }\n\n const gate = await canStartSession(MAX_LIVE_SESSIONS, await liveSessionCount());\n if (!gate.ok) {\n return c.json({\n error: gate.reason === \"budget\"\n ? \"demo budget spent\"\n : `${MAX_LIVE_SESSIONS} browsers are already running — stop one first`,\n reason: gate.reason,\n budget: gate.state,\n }, 429);\n }\n\n const browser = await createBrowser({\n name: `jev-${Date.now()}`,\n startUrl,\n });\n const session = await createSession(browser.session_id, task, startUrl);\n await noteSession();\n\n return c.json({\n sessionId: session.id,\n task,\n startUrl,\n // readOnly so a viewer can't fight the agent for the mouse, and so the\n // demo is a demo rather than a shared remote desktop.\n liveViewUrl: `${browser.browser_live_view_url}?readOnly=true`,\n budget: await getBudget(),\n });\n});\n\n/**\n * One turn of the loop: read the page, ask Jev, act.\n *\n * The client drives this rather than the server looping, so each request stays\n * short and the UI can show every step as it lands.\n */\napp.post(\"/api/step\", async (c) => {\n const body = await c.req.json().catch(() => null);\n const session = await getSession(typeof body?.sessionId === \"string\" ? body.sessionId : \"\");\n if (!session) return c.json({ error: \"no such session\" }, 404);\n if (session.status !== \"live\") {\n return c.json({ error: `run is already ${session.status}`, status: session.status }, 409);\n }\n\n // Bill the browser time that passed while it sat there.\n await accountTime(session);\n\n const history = Array.isArray(body?.history)\n ? body.history.slice(-6).map((h: unknown) => String(h).slice(0, 120))\n : [];\n\n if (session.steps >= MAX_STEPS) {\n await finish(session, \"done\");\n return c.json({\n step: session.steps,\n done: true,\n reason: \"step limit\",\n action: `stopped after ${MAX_STEPS} steps`,\n budget: await getBudget(),\n });\n }\n\n try {\n const page = await perceive(session.kernelId);\n const decision = await decide(page, session.task, history);\n if (decision.action !== \"done\") await act(session.kernelId, decision);\n await bumpSteps(session.id);\n\n const done = decision.action === \"done\";\n if (done) await finish(session, \"done\");\n\n return c.json({\n step: session.steps + 1,\n done,\n action: describeAction(decision),\n page: {\n url: page.url,\n title: page.title,\n candidates: page.candidates.length,\n },\n decision,\n budget: await getBudget(),\n });\n } catch (err) {\n // A spent budget ends the run: there is no point holding a browser open\n // that we've decided we can't afford to drive.\n if (err instanceof BudgetSpentError) {\n await finish(session, \"stopped\");\n return c.json({\n error: \"demo budget spent\",\n reason: \"budget\",\n budget: err.state,\n }, 429);\n }\n await finish(session, \"failed\");\n console.error(\"step failed:\", err);\n return c.json({\n error: err instanceof Error ? err.message : String(err),\n reason: \"upstream\",\n budget: await getBudget().catch(() => null),\n }, 500);\n }\n});\n\n/** Stop a run and delete its browser. */\napp.post(\"/api/session/:id/stop\", async (c) => {\n const session = await getSession(c.req.param(\"id\"));\n if (!session) return c.json({ error: \"no such session\" }, 404);\n await finish(session, \"stopped\");\n return c.json({ ok: true, budget: await getBudget() });\n});\n\n// Unwrap and rethrow Hono errors as the original error.\napp.onError((err) => Promise.reject(err));\n\nexport default hotReload(app.fetch);",[199],"a1ed574e-f2b2-4604-9552-3464d2e0a637","a8691e76-b209-11f1-b405-1607ee4eb77e",31,["D",1789589198804],5,"/**\n * The Jev half of the loop: given a text description of the page and a goal,\n * decide what to do next.\n *\n * Jev can't emit an action. It has no string output at all — only Choice,\n * Score, and Noul — so \"browser use\" here is not a model emitting tool calls.\n * It's the model *ranking and choosing* from a closed set that this file builds\n * out of the page, and then ordinary code doing the clicking. That constraint\n * is the whole design: every action is a value Jev was handed, so it cannot\n * invent a selector, hallucinate a URL, or make a type error.\n *\n * Two stages, which is what TypeSafe does for high-cardinality choices (their\n * wikiracing demo, \"Jev supports a cardinality up to 255... for the higher\n * cardinality choices, we do a 2 stage-system of scoring independently then\n * making an explicit choice\"):\n *\n * 1. Score — one Noul per candidate: \"does activating this move us closer?\"\n * Independent judgments, so they run in parallel and cost one call.\n * 2. Choose — one Choice over the top finalists, plus a way out (`__done__`).\n *\n * Stage 1 is a filter, not a decision. It never picks anything; it just makes\n * stage 2's choice set small enough to be a real choice.\n */\nimport {\n type BudgetState,\n estimateInputTokens,\n reconcile,\n reserve,\n} from \"./budget.ts\";\nimport { ask, type Answer, type Question, type SystemOneResponse } from \"./typesafe.ts\";\n\n/**\n * Candidates sent to stage 1. Beyond this we'd rather lose links than the call:\n * stage 1 asks two questions per candidate, and the documented request budget\n * is ~32k tokens.\n */\nconst MAX_CANDIDATES = 150;\n\n/** Finalists stage 2 chooses among. Small enough to be a comparison, not a scan. */\nconst FINALISTS = 24;\n\n/** Recent actions, so Jev can tell \"we're going in circles\" from \"we're progressing\". */\nconst HISTORY = 6;\n\nexport type Candidate = {\n /** Stable within one page state; the key Jev's answers come back under. */\n n: number;\n tag: string;\n label: string;\n href?: string;\n /** Viewport coordinates of the element's centre, for the computer controls. */\n x: number;\n y: number;\n /** True when this is a text field, i.e. clicking it isn't enough. */\n editable?: boolean;\n};\n\nexport type PageState = {\n url: string;\n title: string;\n /** Visible text, flattened. Jev never sees a screenshot — only this. */\n text: string;\n candidates: Candidate[];\n};\n\n/**\n * One candidate's two independent judgments, plus the composite code makes of\n * them. `goal` and `closer` stay separate all the way to the UI: they answer\n * different questions, and the gap between them is what the score strip shows.\n */\nexport type Ranked = {\n candidate: Candidate;\n /** Noul: this control *is* what the goal names. */\n goal: number;\n /** Noul: this control is a useful next step, even if it isn't the goal. */\n closer: number;\n /** Probabilistic OR of the two: `goal + (1 - goal) * closer`. */\n score: number;\n};\n\nexport type Decision = {\n action: \"click\" | \"type\" | \"done\";\n target?: Candidate;\n text?: string;\n /** Jev's probability for the option it chose, and the answer's own confidence. */\n probability?: number;\n confidence?: number;\n /** Noul: how likely the page already satisfies the goal. */\n goalReached: number;\n /** Stage 1's output, best first. The UI shows the distribution. */\n ranked: Ranked[];\n /** How many candidates stage 1 saw. */\n considered: number;\n usage: { input_tokens: number; output_tokens: number };\n model: string;\n ms: number;\n};\n\n/** Thrown instead of calling Jev when the ledger is spent. */\nexport class BudgetSpentError extends Error {\n constructor(readonly state: BudgetState) {\n super(\"demo budget spent\");\n this.name = \"BudgetSpentError\";\n }\n}\n\n/**\n * Strings we could plausibly type, harvested from the task itself. Jev can't\n * write text, so anything typed has to come from the user's own words — quoted\n * phrases are the honest source, and a task with none simply can't type.\n */\nexport function harvestTexts(task: string): string[] {\n const out: string[] = [];\n const re = /\"([^\"\\n]{1,80})\"|'([^'\\n]{1,80})'|“([^”\\n]{1,80})”/g;\n for (const m of task.matchAll(re)) {\n const value = (m[1] ?? m[2] ?? m[3] ?? \"\").trim();\n if (value) out.push(value);\n }\n return [...new Set(out)];\n}\n\n/** How a candidate is described to Jev. The label is the only thing it can judge on. */\nfunction describe(c: Candidate): string {\n const href = c.href ? ` → ${c.href}` : \"\";\n return `\"${c.label}\" (${c.tag}${href})`;\n}\n\nfunction noulOf(answer: Answer | undefined): number {\n return answer?.type === \"noul\" ? answer.noul : 0;\n}\n\n/** Reserve, call, reconcile. Every Jev call in this file goes through here. */\nasync function askReserved(\n state: unknown,\n questions: Record,\n): Promise {\n const reservation = await reserve(\n estimateInputTokens(JSON.stringify(state).length, questions),\n );\n if (!reservation.ok) throw new BudgetSpentError(reservation.state);\n try {\n const res = await ask(state, questions);\n await reconcile(reservation, res.usage.input_tokens).catch((err) =>\n console.error(\"budget reconcile failed:\", err)\n );\n return res;\n } catch (err) {\n // A call we watched fail wasn't billed, so hand the reservation back.\n await reconcile(reservation, 0).catch(() => {});\n throw err;\n }\n}\n\nexport async function decide(\n page: PageState,\n task: string,\n history: string[],\n): Promise {\n const started = performance.now();\n const considered = page.candidates.slice(0, MAX_CANDIDATES);\n const recent = history.slice(-HISTORY);\n\n // ── Stage 1: score every candidate independently ────────────────\n // Two questions per candidate, not one. A single \"does this move us closer?\"\n // conflates the goal itself with a step toward it, and measurably does: on a\n // fixture where the goal link sat beside \"Gustave Eiffel\", the one-question\n // version ranked the biographer above the tower and clicked him. Asking the\n // two things separately and combining them in code fixes that, and costs one\n // call either way — questions in a request are evaluated in parallel.\n const scoreQuestions: Record = {};\n for (const c of considered) {\n scoreQuestions[`g${c.n}`] = {\n type: \"noul\",\n instructions:\n `This control IS the goal — activating it lands on exactly the page or thing the goal names. The control is ${describe(c)}.`,\n };\n scoreQuestions[`c${c.n}`] = {\n type: \"noul\",\n instructions:\n `This control is a useful next step toward the goal, even though it is not the goal itself. The control is ${describe(c)}.`,\n };\n }\n\n const scoreState = {\n goal: task,\n current_page: { url: page.url, title: page.title },\n page_text: page.text,\n recent_actions: recent,\n };\n\n const stage1 = await askReserved(scoreState, scoreQuestions);\n\n // Combining the two answers is code's job, not the model's. A probabilistic\n // OR is the right shape: a candidate is promising if it either *is* the goal\n // or clearly leads toward it. When the goal link isn't on the page at all,\n // the second term is the only one that fires — which is the multi-hop case.\n const ranked: Ranked[] = considered\n .map((candidate) => {\n const goal = noulOf(stage1.answers[`g${candidate.n}`]);\n const closer = noulOf(stage1.answers[`c${candidate.n}`]);\n return { candidate, goal, closer, score: goal + (1 - goal) * closer };\n })\n .sort((a, b) => b.score - a.score);\n\n const finalists = ranked.slice(0, FINALISTS);\n\n // ── Stage 2: choose among the finalists ─────────────────────────\n const criteria: Record = {\n __done__: \"The goal is already achieved on this page — stop and report success.\",\n };\n for (const f of finalists) criteria[`c${f.candidate.n}`] = describe(f.candidate);\n\n const texts = harvestTexts(task);\n const questions: Record = {\n goal_reached: {\n type: \"noul\",\n instructions:\n \"The page currently displayed already satisfies the goal, with nothing further to do.\",\n },\n next: {\n type: \"choice\",\n instructions:\n \"Which single control should be activated next to make progress toward the goal?\",\n criteria,\n },\n };\n\n // Only asked when it can matter, and only when there is something to type.\n if (texts.length) {\n questions.needs_text = {\n type: \"noul\",\n instructions:\n \"The control chosen above needs text typed into it before it does anything useful.\",\n };\n questions.text = {\n type: \"choice\",\n instructions:\n \"Which of these strings should be typed into the chosen control?\",\n criteria: Object.fromEntries([\n ...texts.map((t) => [t, t] as const),\n [\"__none__\", \"None of these — type nothing.\"] as const,\n ]),\n };\n }\n\n const stage2 = await askReserved(\n {\n goal: task,\n current_page: { url: page.url, title: page.title },\n page_text: page.text,\n recent_actions: recent,\n finalists: finalists.map((f) => ({\n control: describe(f.candidate),\n is_the_goal: Number(f.goal.toFixed(2)),\n moves_closer: Number(f.closer.toFixed(2)),\n promising: Number(f.score.toFixed(2)),\n })),\n },\n questions,\n );\n\n const answers = stage2.answers;\n const goalReached = noulOf(answers.goal_reached);\n const pick = answers.next?.type === \"choice\" ? answers.next : undefined;\n const chosenKey = pick?.choice ?? \"__done__\";\n const chosen = finalists.find((f) => `c${f.candidate.n}` === chosenKey)?.candidate;\n\n const needsText = noulOf(answers.needs_text) >= 0.5;\n const textPick = answers.text?.type === \"choice\" ? answers.text.choice : \"__none__\";\n\n // Code, not the model, decides what a choice means. `__done__`, a confident\n // \"already satisfied\", or a finalist that vanished all stop the loop.\n const action: Decision[\"action\"] = goalReached >= 0.6 || !chosen\n ? \"done\"\n : needsText && textPick !== \"__none__\"\n ? \"type\"\n : \"click\";\n\n return {\n action,\n target: action === \"done\" ? undefined : chosen,\n text: action === \"type\" ? textPick : undefined,\n probability: pick?.probabilities?.[chosenKey],\n confidence: pick?.confidence,\n goalReached,\n ranked,\n considered: considered.length,\n usage: {\n input_tokens: stage1.usage.input_tokens + stage2.usage.input_tokens,\n output_tokens: stage1.usage.output_tokens + stage2.usage.output_tokens,\n },\n model: stage2.model,\n ms: Math.round(performance.now() - started),\n };\n}",[39],"da6c4d4d-590a-4599-8b60-f22b731231c2","budget.ts","01a0abce-8571-76e1-8061-db4c242eaf56",32,["D",1789589198882],2,"/**\n * A hard ceiling on what this demo can spend — at TypeSafe *and* at Kernel.\n *\n * Two meters, one ledger:\n *\n * - Jev bills $0.042 per 1M input tokens and nothing for output, so a step\n * costs only what you send it: a few thousand tokens, i.e. ~$0.0002.\n * - Kernel bills a headful browser at $0.0001333336/second, which is\n * $0.008/minute. Per second, a browser costs 20x what a Jev call costs.\n * The browser is what actually drains a budget, not the model.\n *\n * The running total lives in this val's SQLite database: in-memory counters\n * don't survive Val Town's serverless instances, so a cap that only lives in\n * memory isn't a cap.\n *\n * Jev is reserve-then-reconcile: charge a conservative estimate before the\n * call, so a burst of concurrent requests can't all read the same total and\n * slip through, then correct the ledger with the real `usage.input_tokens`\n * from the response. The estimate only has to be close, because the correction\n * lands right after.\n *\n * Kernel time is charged as wall-clock since the session started, which\n * deliberately over-counts: Kernel's Standby Mode stops billing 5 seconds\n * after the last request, while this clock keeps running. Over-counting is the\n * safe direction for a hard cap, and it means an abandoned tab can't quietly\n * spend the budget.\n */\nimport { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\n\n/** Jev list price, i.e. $42 per billion input tokens. Output tokens are free. */\nexport const USD_PER_INPUT_TOKEN = 0.042 / 1_000_000;\n\n/** Kernel headful browser, per second. See https://www.kernel.sh/docs/info/pricing */\nexport const USD_PER_KERNEL_SECOND = 0.0001333336;\n\n/** Lifetime budget for the whole demo. Override with BUDGET_USD. */\nexport const BUDGET_USD = Number(Deno.env.get(\"BUDGET_USD\") ?? \"2\");\n\n/**\n * Calibrated against real responses. Prose runs about 4 characters per token;\n * question JSON is denser — lots of punctuation and short keys — and comes in\n * nearer 2.7.\n */\nconst STATE_CHARS_PER_TOKEN = 4;\nconst QUESTION_CHARS_PER_TOKEN = 2.7;\n\n/** Cheap pre-flight token estimate for one call. Not exact, and doesn't need to be. */\nexport function estimateInputTokens(\n stateChars: number,\n questions: unknown,\n): number {\n return Math.ceil(stateChars / STATE_CHARS_PER_TOKEN) +\n Math.ceil(JSON.stringify(questions).length / QUESTION_CHARS_PER_TOKEN);\n}\n\nexport type BudgetState = {\n budgetUsd: number;\n spentUsd: number;\n remainingUsd: number;\n /** Split out so the UI can show which meter is actually moving. */\n jevUsd: number;\n kernelUsd: number;\n inputTokens: number;\n kernelSeconds: number;\n calls: number;\n sessions: number;\n exhausted: boolean;\n};\n\n/**\n * A claim on the budget. `state` is the ledger *before* the call, which is what\n * you want to show when refusing one.\n */\nexport type Reservation = {\n ok: boolean;\n state: BudgetState;\n estimateTokens: number;\n};\n\nconst SCHEMA = `CREATE TABLE IF NOT EXISTS budget (\n id INTEGER PRIMARY KEY CHECK (id = 1),\n input_tokens INTEGER NOT NULL DEFAULT 0,\n kernel_seconds REAL NOT NULL DEFAULT 0,\n calls INTEGER NOT NULL DEFAULT 0,\n sessions INTEGER NOT NULL DEFAULT 0,\n updated_at TEXT\n)`;\n\nlet schema: Promise | null = null;\n\n/**\n * A remix inherits the source val's database *schema*, and `CREATE TABLE IF NOT\n * EXISTS` will not add columns to a table that's already there. So remixing a\n * val whose ledger had fewer columns fails on the very first read — which is\n * exactly what happened when this val was remixed from the typewriter demo.\n * Adding the columns explicitly, and swallowing \"duplicate column\", works on a\n * fresh database and on an inherited one.\n */\nconst MIGRATIONS = [\n \"ALTER TABLE budget ADD COLUMN kernel_seconds REAL NOT NULL DEFAULT 0\",\n \"ALTER TABLE budget ADD COLUMN sessions INTEGER NOT NULL DEFAULT 0\",\n];\n\nfunction init(): Promise {\n if (!schema) {\n schema = (async () => {\n await sqlite.execute(SCHEMA);\n for (const sql of MIGRATIONS) {\n await sqlite.execute(sql).catch(() => {}); // already there\n }\n })().catch((err) => {\n schema = null; // let the next request try again\n throw err;\n });\n }\n return schema;\n}\n\nfunction summarize(\n inputTokens: number,\n kernelSeconds: number,\n calls: number,\n sessions: number,\n): BudgetState {\n const jevUsd = inputTokens * USD_PER_INPUT_TOKEN;\n const kernelUsd = kernelSeconds * USD_PER_KERNEL_SECOND;\n const spentUsd = jevUsd + kernelUsd;\n return {\n budgetUsd: BUDGET_USD,\n spentUsd,\n remainingUsd: Math.max(0, BUDGET_USD - spentUsd),\n jevUsd,\n kernelUsd,\n inputTokens,\n kernelSeconds,\n calls,\n sessions,\n exhausted: spentUsd >= BUDGET_USD,\n };\n}\n\ntype Delta = { tokens?: number; seconds?: number; calls?: number; sessions?: number };\n\n/** Move the ledger. Tokens are clamped at zero so a downward correction can't go negative. */\nasync function add(delta: Delta): Promise {\n await init();\n const { tokens = 0, seconds = 0, calls = 0, sessions = 0 } = delta;\n await sqlite.execute({\n sql: `INSERT INTO budget (id, input_tokens, kernel_seconds, calls, sessions, updated_at)\n VALUES (1, MAX(0, ?), MAX(0, ?), ?, ?, ?)\n ON CONFLICT (id) DO UPDATE SET\n input_tokens = MAX(0, input_tokens + ?),\n kernel_seconds = MAX(0, kernel_seconds + ?),\n calls = calls + ?,\n sessions = sessions + ?,\n updated_at = excluded.updated_at`,\n args: [\n tokens, seconds, calls, sessions, new Date().toISOString(),\n tokens, seconds, calls, sessions,\n ],\n });\n}\n\nexport async function getBudget(): Promise {\n await init();\n const { rows } = await sqlite.execute(\n \"SELECT input_tokens, kernel_seconds, calls, sessions FROM budget WHERE id = 1\",\n );\n const row = rows[0];\n return summarize(\n Number(row?.input_tokens ?? 0),\n Number(row?.kernel_seconds ?? 0),\n Number(row?.calls ?? 0),\n Number(row?.sessions ?? 0),\n );\n}\n\n/** Claim room in the budget for one call, or refuse it. */\nexport async function reserve(estimateTokens: number): Promise {\n const state = await getBudget();\n const estimateUsd = estimateTokens * USD_PER_INPUT_TOKEN;\n if (state.exhausted || state.spentUsd + estimateUsd > BUDGET_USD) {\n return { ok: false, state, estimateTokens };\n }\n await add({ tokens: estimateTokens, calls: 1 });\n return { ok: true, state, estimateTokens };\n}\n\n/** Swap a reservation's estimate for what the call actually cost. */\nexport async function reconcile(\n reservation: Reservation,\n actualTokens: number,\n): Promise {\n if (!reservation.ok) return;\n const delta = actualTokens - reservation.estimateTokens;\n if (delta !== 0) await add({ tokens: delta });\n}\n\n/** Charge browser time. Called as a session is accounted, and once more when it closes. */\nexport async function chargeKernelSeconds(seconds: number): Promise {\n if (seconds > 0) await add({ seconds });\n}\n\n/** Count a browser against the ledger's session tally. */\nexport async function noteSession(): Promise {\n await add({ sessions: 1 });\n}\n\n/**\n * Would a new browser fit? Refused on either meter, so a spent budget stops\n * handing out browsers rather than only refusing steps.\n */\nexport async function canStartSession(\n maxLive: number,\n live: number,\n): Promise<{ ok: true } | { ok: false; reason: \"budget\" | \"concurrency\"; state: BudgetState }> {\n const state = await getBudget();\n if (state.exhausted) return { ok: false, reason: \"budget\", state };\n if (live >= maxLive) return { ok: false, reason: \"concurrency\", state };\n return { ok: true };\n}",[213],"92e50c69-8ee3-4d17-9813-e12fe5cb8848","browser-use.ts","ba961748-b209-11f1-a2a2-1607ee4eb77e",17,["D",1789589027970],"/**\n * The Kernel half of the loop: turn the live page into text Jev can judge, and\n * then carry out the action Jev chose.\n *\n * Two halves that have to agree on one thing — coordinates.\n *\n * - `perceive()` runs a snippet inside the browser's own VM (Kernel's\n * Playwright execution endpoint) and returns the page as *data*: url,\n * title, visible text, and every visible interactive element with the\n * viewport coordinates of its centre. Jev never sees a screenshot; this is\n * the only view of the page it gets.\n * - `act()` uses Kernel's computer controls to move a real OS cursor and\n * click, so the live view shows the work happening.\n *\n * They agree because the browser runs in kiosk mode: the window is the page\n * viewport, so `getBoundingClientRect()` is already screen space.\n */\nimport type { Candidate, Decision, PageState } from \"./decide.ts\";\nimport {\n clickMouse,\n moveMouse,\n pressKey,\n runPlaywright,\n sleep,\n typeText,\n} from \"./kernel.ts\";\n\n/** Visible-text budget handed to Jev. The docs put the request cap near 32k tokens. */\nconst TEXT_CHARS = 2500;\n\n/** Interactive elements collected per page, before `decide.ts` trims to its own cap. */\nconst MAX_ELEMENTS = 400;\n\n/**\n * The snippet that runs in the browser. `page` is in scope and the returned\n * value comes back as `result`.\n *\n * The `closest()` filter is what makes this usable on a real site: without it\n * a Wikipedia article is mostly sidebar, and stage 1 would spend its whole\n * budget scoring navigation chrome.\n */\nconst PERCEIVE = `\nconst raw = await page.evaluate(() => {\n const CHROME = 'nav,header,footer,aside,[role=\"navigation\"],[role=\"banner\"],[role=\"contentinfo\"],[role=\"search\"]';\n const SEL = 'a[href],button,input,textarea,select,[role=\"button\"],[role=\"link\"],[role=\"tab\"],[contenteditable=\"true\"]';\n const seen = new Set();\n const out = [];\n for (const el of document.querySelectorAll(SEL)) {\n if (el.closest(CHROME)) continue;\n const r = el.getBoundingClientRect();\n if (r.width < 8 || r.height < 8) continue;\n if (r.bottom <= 0 || r.top >= window.innerHeight) continue;\n if (r.right <= 0 || r.left >= window.innerWidth) continue;\n const cs = getComputedStyle(el);\n if (cs.visibility === 'hidden' || cs.display === 'none' || Number(cs.opacity) < 0.05) continue;\n const label = String(\n el.innerText || el.value || el.getAttribute('aria-label') ||\n el.getAttribute('placeholder') || el.getAttribute('title') ||\n el.getAttribute('alt') || ''\n ).replace(/\\\\s+/g, ' ').trim().slice(0, 90);\n if (!label) continue;\n const key = label + '|' + Math.round(r.left) + '|' + Math.round(r.top);\n if (seen.has(key)) continue;\n seen.add(key);\n const tag = el.tagName.toLowerCase();\n out.push({\n tag,\n label,\n href: (el.getAttribute('href') || '').slice(0, 140),\n editable: tag === 'input' || tag === 'textarea' || el.isContentEditable === true,\n x: Math.round(r.left + r.width / 2),\n y: Math.round(r.top + r.height / 2),\n });\n if (out.length >= ${MAX_ELEMENTS}) break;\n }\n return {\n url: location.href,\n title: document.title,\n text: document.body.innerText.replace(/\\\\s+/g, ' ').trim().slice(0, ${TEXT_CHARS}),\n candidates: out,\n };\n});\nreturn raw;\n`;\n\ntype RawPage = {\n url: string;\n title: string;\n text: string;\n candidates: Omit[];\n};\n\n/** Read the current page as data. */\nexport async function perceive(sessionId: string): Promise {\n const res = await runPlaywright(sessionId, PERCEIVE);\n if (!res.success) {\n throw new Error(res.error || \"could not read the page\");\n }\n const raw = res.result as RawPage;\n return {\n url: raw?.url ?? \"\",\n title: raw?.title ?? \"\",\n text: raw?.text ?? \"\",\n // `n` is assigned here rather than in the browser: it's the key Jev's\n // answers come back under, so it has to be stable and sequential.\n candidates: (raw?.candidates ?? []).map((c, n) => ({ ...c, n })),\n };\n}\n\n/** How long to let a click settle before reading the page again. */\nconst SETTLE_MS = 1200;\n\n/**\n * Carry out one decision.\n *\n * Clicking always moves the cursor first. That's not just for the video — a\n * bare click at coordinates the cursor never visited is the kind of thing\n * hover-sensitive menus ignore.\n */\nexport async function act(sessionId: string, decision: Decision): Promise {\n const target = decision.target;\n if (!target) return;\n\n await moveMouse(sessionId, target.x, target.y, 420);\n await clickMouse(sessionId, target.x, target.y);\n\n if (decision.action === \"type\" && decision.text) {\n await sleep(250);\n await typeText(sessionId, decision.text);\n // Enter is the assumption: a field we chose to type into is nearly always\n // a search box or a form's only input. A field that needs a separate\n // submit click gets one on the next step.\n await pressKey(sessionId, [\"Enter\"]);\n }\n\n await sleep(SETTLE_MS);\n}\n\n/** One line describing what happened, for the step log and Jev's own history. */\nexport function describeAction(decision: Decision): string {\n if (decision.action === \"done\") return \"stop — goal reached\";\n const target = decision.target;\n const what = target ? `\"${target.label}\"` : \"nothing\";\n return decision.action === \"type\"\n ? `type \"${decision.text}\" into ${what}`\n : `click ${what}`;\n}",[221],"dbe10f1b-b21f-423f-8c1d-d5683ff88a2e","README.md","01a0abce-8571-76e1-8061-d5a5e82286cc",["D",1789588833583],"# TypeSafe Typewriter\n\n16 typed questions about whatever you're typing, re-asked on **every keystroke**.\n\n[Try it](url:index.ts)\n\n[TypeSafe](https://docs.typesafe.ai)'s Jev is a \"System One\" model. You send it a\npiece of state and a map of typed questions, and it returns typed answers:\n\n| Question | Returns |\n| --- | --- |\n| `choice` | one option from a list you define, plus a probability for every option |\n| `score` | a position along an ordered rubric you define, plus the distribution |\n| `noul` | the probability that a statement is true |\n\nNo text generation, no parsing, no recovering a value from prose. Every answer is\nconstrained to the options you supplied, and comes back with a probability\ndistribution your code can threshold, sort, or route on.\n\nThe interesting property is that it's **fast** — fast enough to sit inside a UI\nloop. So this demo asks all 16 questions in a single API call and re-asks them on\nevery keystroke.\n\n## What it demonstrates\n\n- **Fan-out is nearly free.** All 16 questions go out in one request. Adding\n questions barely moves the response time, because each one is evaluated in\n parallel against the same state.\n- **The distribution is the product.** The bars aren't decoration — they're the\n full probability vector. A `score` answer of 2.77 means \"77% Warm, 23%\n Neutral\", and the panel shows exactly that.\n- **Code owns the rendering.** The model returns numbers. The bars are CSS. There\n is no prompt that says \"output a chart\".\n- **Speed changes the design.** At a few hundred milliseconds this is a live\n instrument panel, not a form you submit. A keystroke re-asks all 16 questions.\n\n## How it works\n\n```mermaid\nsequenceDiagram\n participant B as Browser\n participant H as index.ts (Hono)\n participant T as TypeSafe\n B->>H: POST /api/judge { text }\n H->>H: reserve ~900 tokens against the $2 budget\n H->>T: POST /v1/systemone { state, 16 questions }\n T-->>H: 11 noul + 4 score + 1 choice\n H->>H: reconcile the ledger with usage.input_tokens\n H-->>B: answers, usage, budget, serverMs\n B->>B: redraw 16 meters\n```\n\nThe client debounces 120ms, cancels the pending request when you type again\n(`AbortController`), and keeps showing the previous answers while the next batch\nis in flight — so the panel never flickers to empty.\n\n## The questions\n\nThree groups, all defined in [`questions.ts`](url:questions.ts):\n\n- **What it is** — asks a question, asks for action, makes a promise, states an\n opinion, checkable claim, names specifics\n- **How it lands** — reads AI-written, passive-aggressive, expresses emotion,\n would start an argument, would get a reply, tone, conviction, urgency\n- **Who it's for** — audience, formality\n\nAdding one is a single entry in that file. The browser renders whatever the API\nreports, so there's nothing else to update.\n\n## Files\n\nSee [`AGENTS.md`](url:AGENTS.md) for the file map.\n\n## Cost and caps\n\nJev bills **$0.042 per 1M input tokens** and nothing at all for output, so a\ncall costs only what you send it. A 16-question call on a short sentence bills\n~900 input tokens — about **$0.00004**, or 25,000 calls to the dollar.\n\nThat cheapness is the risk. At a 120ms debounce a single open tab can ask a few\nhundred times a minute, so a public link is a spend you can't see. `POST\n/api/judge` is capped twice over, and the two caps do different jobs:\n\n| Cap | Lives in | Stops |\n| --- | --- | --- |\n| 600/min per IP, 6000/min globally | `rate-limit.ts`, in memory | a burst from one tab or one scraper |\n| **$2.00 lifetime** | `budget.ts`, in SQLite | the demo ever costing more than $2, however it's hit |\n\nThe budget is the real cap. It's one row in the val's SQLite database, so it\nsurvives isolates, cold starts and deploys — an in-memory counter wouldn't be a\ncap at all. Each call reserves a token estimate before it goes out (so a burst of\nconcurrent requests can't all read the same total and slip through) and then\ncorrects the ledger with the `usage.input_tokens` that came back. A call we\nwatched fail hands its reservation back.\n\nWhen the budget is spent, `/api/judge` answers 429 with the ledger in the body,\nand the client stops asking on its own — the panel freezes on the last answers\nand says what the $2 bought, rather than filling with errors.\n\nSet `TYPESAFE_BUDGET_USD` to change the cap. Current spend is in the app's footer,\nin the response bodies, and in `SELECT * FROM budget`.\n\n## Notes\n\n- Text is capped at 4000 characters.\n- `TYPESAFE_API_KEY` is the only required environment variable. Without it the\n page says so up front — `/api/questions` reports whether a key exists, so the\n panel shows what's missing and asks nothing, instead of failing on the first\n keystroke.",[228],"6eaf2403-b94a-43d6-b8ac-7daf11eed9ca","AGENTS.md","01a0abce-8571-76e1-8061-cdf4a9f6fc29",25,["D",1789589133731],"# jev-browser-use\n\nDescribe a task. Jev picks every next move in a live cloud browser you watch in\nan iframe.\n\n## File map\n\n| Path | Type | Purpose |\n| --- | --- | --- |\n| `index.ts` | http | Hono app: shell, assets, `GET /api/config`, `POST /api/session`, `POST /api/step`, `POST /api/session/:id/stop` |\n| `decide.ts` | script | The Jev half: two-stage score-then-choose over the page's controls |\n| `browser-use.ts` | script | The Kernel half: `perceive()` reads the page as data, `act()` clicks/types |\n| `kernel.ts` | script | Kernel REST client — browsers, live view, computer controls |\n| `budget.ts` | script | SQLite ledger: the hard $2 cap, across both Jev tokens and Kernel seconds |\n| `sessions.ts` | script | Live-browser registry + the clock that bills elapsed seconds |\n| `rate-limit.ts` | script | Best-effort in-memory per-IP + global burst control |\n| `typesafe.ts` | script | Minimal TypeSafe (System One) REST client + answer types |\n| `tests/decide.ts` | script | Runs `decide()` against a fixture page state — no browser needed |\n| `frontend/root.tsx` | script | HTML shell (Hono JSX); stamps immutable asset URLs |\n| `frontend/index.tsx` | script | React entrypoint |\n| `frontend/components/App.tsx` | script | State machine, controls, step log, `?present=1` mode |\n| `frontend/components/Stage.tsx` | script | The square stage: task, live view iframe, caption, score strip |\n| `frontend/lib/api.ts` | script | Typed client for the val's own API |\n| `frontend/favicon.svg` | file | Icon |\n\n## The loop\n\n```\nperceive() → decide() → act() → repeat\n Kernel Jev Kernel\n```\n\n`POST /api/step` runs exactly one turn, and the **browser** drives the loop by\ncalling it repeatedly. That keeps each request short and lets the UI render\nevery step as it lands — which is the whole point of the demo.\n\n## Notes\n\n- **Jev cannot emit an action.** It has no string output at all — only `choice`,\n `score`, `noul`. So this is not a model emitting tool calls: `decide.ts` builds\n a closed set out of the page, Jev ranks and chooses within it, and ordinary\n code does the clicking. That's the constraint the whole design hangs on, and\n it's why the model can't hallucinate a selector or invent a URL.\n- **Two stages, per TypeSafe's own wikiracing demo** (\"Jev supports a cardinality\n up to 255... for the higher cardinality choices, we do a 2 stage-system of\n scoring independently then making an explicit choice\"). Stage 1 is one `noul`\n per candidate — a filter, it never picks. Stage 2 is one `choice` over the top\n `FINALISTS` plus `__done__`. Constants live at the top of `decide.ts`.\n- **Jev never sees the screen.** It gets `{ url, title, visible_text, candidates }`\n as text, plus the goal and recent actions. No screenshots — matching the blog's\n note that the Doom demo runs \"on structured state as a data structure with\n text, not on images (yet…)\".\n- **Typing is the one thing Jev can't source.** It can't write text, so\n `harvestTexts()` pulls quoted phrases out of the user's own task and Jev\n *chooses* which to type. A task with no quoted phrase simply can't type.\n- **Kiosk mode is load-bearing, not cosmetic.** It removes the browser chrome so\n the window *is* the page viewport, which makes `getBoundingClientRect()`\n coordinates equal the screen coordinates the computer controls click. With\n chrome in the way every click lands a toolbar-height too high.\n- **Clicking always moves the cursor first**, via `moveMouse` with Kernel's\n default Bézier path. It's not just for the video — hover-sensitive menus\n ignore a click at coordinates the cursor never visited.\n- **Two caps, two jobs.** `rate-limit.ts` is burst control (6/min per IP,\n 60/min globally — much tighter than a per-keystroke demo's, because a browser\n is ~$0.008/minute). `budget.ts` is the hard cap: **$2.00 lifetime**, one row in\n SQLite so it survives isolates and deploys. Set `BUDGET_USD` to move it.\n- **The browser is what costs money, not the model.** A Jev step is ~$0.0002; a\n minute of headful browser is ~$0.008. `sessions.ts` charges wall-clock elapsed\n time, which deliberately over-counts (Kernel's Standby Mode stops billing after\n 5s idle) because over-counting is the safe direction for a hard cap.\n- **Browsers must be deleted, not abandoned.** `finish()` accounts the last\n seconds and deletes the session. It's called on done, on stop, on step-limit,\n on a spent budget, and on error. The client also stops the session in a\n `finally`, so a loop that exits early can't leave a browser billing. Kernel's\n `timeout_seconds` (240s idle) is the backstop.\n- **`MAX_LIVE_SESSIONS` is 2.** Kernel's Developer tier allows 5 concurrent\n browsers; 2 is a spend decision, not a technical one.\n- `data-*` hooks for Playwright: `data-stage`, `data-task`, `data-caption`,\n `data-status`, `data-scores`, `data-live-view`, `data-log`, `data-log-entry`,\n `data-task-input`, `data-url-input`, `data-run`, `data-stop`,\n `data-budget-spent`.\n- `?present=1` renders only the square stage, sized to the viewport. That's the\n mode to screen-record.\n- Env: `TYPESAFE_API_KEY` (https://console.typesafe.ai/settings/keys) and\n `KERNEL_API_KEY` (https://dashboard.onkernel.com/settings/api-keys). Both are\n required; `/api/config` reports which are missing and the page says so up front.",[234],"a6c4385d-8c86-4d90-bb2f-e40fa8f5ad86","tests","01a0abce-8571-76e1-8061-ca995244ffd6",["D",1789588833583],"directory",[241],"547cc93c-0b46-4b3d-b204-3199dbf1b11e","frontend","01a0abce-8571-76e1-8061-c0e52b5be3b5",["D",1789588833583],[247],"pulls","branches","forks","envVars",{"_258":217,"_260":72},"connections",{"_258":72,"_259":72},"count","unfilledCount","emptyCount",{"_50":246,"_38":247,"_36":248,"_171":-5,"_48":49,"_47":172,"_173":68,"_45":380,"_13":72,"_40":244,"_42":-5,"_30":31,"_176":-5,"_283":381,"_33":382},{"_50":376,"_38":367,"_36":361,"_171":248,"_48":49,"_47":172,"_173":68,"_45":377,"_13":72,"_40":244,"_42":-5,"_30":31,"_176":-5,"_283":378,"_33":379},{"_50":368,"_38":369,"_36":370,"_171":361,"_48":49,"_47":371,"_173":31,"_45":372,"_13":172,"_40":41,"_42":373,"_30":31,"_176":-5,"_283":374,"_33":375},{"_50":358,"_38":359,"_36":360,"_171":361,"_48":49,"_47":362,"_173":31,"_45":363,"_13":72,"_40":41,"_42":364,"_30":31,"_176":-5,"_283":365,"_33":366},{"_50":354,"_38":353,"_36":347,"_171":248,"_48":49,"_47":172,"_173":68,"_45":355,"_13":72,"_40":244,"_42":-5,"_30":31,"_176":-5,"_283":356,"_33":357},{"_50":344,"_38":345,"_36":346,"_171":347,"_48":49,"_47":348,"_173":31,"_45":349,"_13":72,"_40":41,"_42":350,"_30":31,"_176":-5,"_283":351,"_33":352},{"_50":336,"_38":337,"_36":338,"_171":248,"_48":49,"_47":339,"_173":68,"_45":340,"_13":172,"_40":26,"_42":341,"_30":31,"_176":-5,"_283":342,"_33":343},{"_50":328,"_38":329,"_36":330,"_171":248,"_48":49,"_47":331,"_173":68,"_45":332,"_13":172,"_40":41,"_42":333,"_30":31,"_176":-5,"_283":334,"_33":335},{"_50":320,"_38":321,"_36":322,"_171":248,"_48":49,"_47":323,"_173":31,"_45":324,"_13":172,"_40":41,"_42":325,"_30":31,"_176":-5,"_283":326,"_33":327},{"_50":240,"_38":241,"_36":242,"_171":-5,"_48":49,"_47":172,"_173":68,"_45":317,"_13":72,"_40":244,"_42":-5,"_30":31,"_176":-5,"_283":318,"_33":319},{"_50":313,"_38":39,"_36":37,"_171":242,"_48":49,"_47":14,"_173":31,"_45":314,"_13":172,"_40":41,"_42":44,"_30":31,"_176":-5,"_283":315,"_33":316},{"_50":233,"_38":234,"_36":235,"_171":-5,"_48":49,"_47":236,"_173":31,"_45":310,"_13":172,"_40":26,"_42":238,"_30":31,"_176":-5,"_283":311,"_33":312},{"_50":227,"_38":228,"_36":229,"_171":-5,"_48":49,"_47":172,"_173":31,"_45":307,"_13":72,"_40":26,"_42":231,"_30":31,"_176":-5,"_283":308,"_33":309},{"_50":220,"_38":221,"_36":222,"_171":-5,"_48":49,"_47":223,"_173":31,"_45":304,"_13":72,"_40":41,"_42":225,"_30":31,"_176":-5,"_283":305,"_33":306},{"_50":212,"_38":213,"_36":214,"_171":-5,"_48":49,"_47":215,"_173":31,"_45":301,"_13":217,"_40":41,"_42":218,"_30":31,"_176":-5,"_283":302,"_33":303},{"_50":205,"_38":39,"_36":206,"_171":-5,"_48":49,"_47":207,"_173":31,"_45":298,"_13":209,"_40":41,"_42":210,"_30":31,"_176":-5,"_283":299,"_33":300},{"_50":198,"_38":199,"_36":66,"_171":-5,"_48":49,"_47":200,"_173":31,"_45":295,"_13":172,"_40":202,"_42":203,"_30":31,"_176":52,"_283":296,"_33":297},{"_50":191,"_38":192,"_36":193,"_171":-5,"_48":49,"_47":194,"_173":31,"_45":292,"_13":172,"_40":41,"_42":196,"_30":31,"_176":-5,"_283":293,"_33":294},{"_50":185,"_38":186,"_36":187,"_171":-5,"_48":49,"_47":172,"_173":31,"_45":289,"_13":72,"_40":41,"_42":189,"_30":31,"_176":-5,"_283":290,"_33":291},{"_50":178,"_38":179,"_36":180,"_171":-5,"_48":49,"_47":181,"_173":31,"_45":286,"_13":72,"_40":41,"_42":183,"_30":31,"_176":-5,"_283":287,"_33":288},{"_50":168,"_38":169,"_36":170,"_171":-5,"_48":49,"_47":172,"_173":68,"_45":282,"_13":72,"_40":41,"_42":175,"_30":31,"_176":-5,"_283":284,"_33":285},["D",1789588833583],"children",[],[169],["D",1789588969304],[],[179],["D",1789588833583],[],[186],["D",1789589027405],[],[192],["D",1789589059287],[],[199],["D",1789589198804],[],[39],["D",1789589198882],[],[213],["D",1789589027970],[],[221],["D",1789588833583],[],[228],["D",1789589133731],[],[234],"fca81fc9-0eec-4681-8d25-8b5de29b80f1",["D",1789589208539],[],[241,39],["D",1789588833583],[271],[241],"79588ce1-a520-4f24-bd1d-1a35911e5912","root.tsx","01a0abce-8571-76e1-8061-bf005da2f3ad",19,["D",1789589059333],"/** @jsxImportSource npm:hono@4/jsx */\nimport { raw } from \"npm:hono@4/html\";\nimport { immutableFileUrl } from \"https://esm.town/v/std/utils/index.ts\";\n\n/**\n * Inlined before Twind loads, so the first paint isn't unstyled — Twind is\n * fetched from a CDN at runtime. Also holds the keyframes, which Twind's\n * Tailwind v2 theme has no way to express.\n *\n * Dark by design. The thing this page is for is being screen-recorded, and\n * white-on-near-black stays legible at Twitter's compression and at a glance\n * from across a room.\n */\nconst CSS = `\nhtml, body { background: #0b0b0d; color: #fff; }\n@keyframes captionIn {\n from { opacity: 0; transform: translateY(8px); }\n to { opacity: 1; transform: none; }\n}\n.caption-in { animation: captionIn 260ms ease-out; }\n@keyframes breathe {\n 0%, 100% { opacity: 1; }\n 50% { opacity: 0.3; }\n}\n.breathe { animation: breathe 1.1s ease-in-out infinite; }\n`;\n\n// The HTML shell for the app. It is never cached, so immutableFileUrl can stamp\n// the current val version into each asset URL — the browser caches those\n// immutably, and publishing the val changes the URLs, refetching exactly once.\nexport function Root() {\n return (\n <>\n {raw(\"\")}\n \n \n \n \n Jev Browser Use\n \n \n \n \n \n \n \n \n

Diff · v33