[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_333":334,"_335":336},"routes/_app.x.$handle.$name",{"_98":97,"_31":43,"_107":88,"_26":108,"_13":14,"_109":-5,"_11":110,"_111":112,"_113":114,"_115":116,"_117":118,"_20":-5,"_119":-5,"_120":-5,"_121":24,"_122":123,"_124":57,"_125":-5,"_126":-5,"_127":128},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_41":42,"_31":43,"_36":44,"_45":-5,"_46":24,"_47":48,"_49":48,"_50":51,"_52":-5,"_18":53,"_54":55,"_56":57,"_58":-5,"_59":-5,"_60":61,"_62":63,"_64":65,"_66":67,"_68":69,"_70":-5,"_71":72,"_73":72,"_74":75,"_76":75,"_77":75,"_78":72,"_79":72,"_80":81,"_82":83,"_84":85,"_86":24},"version",137,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":-7,"_23":24,"_25":24,"_26":27,"_28":27},"file",{"_29":30,"_31":32,"_33":20,"_34":35,"_36":37,"_38":14,"_39":40,"_23":24},"previous","deleted",false,"deletedHere","path","frontend/toolkit.html","latestPath","fileId","f7fa8fba-430f-11f1-acca-42b51c65c3df","name","toolkit.html","fileType","content","\n\n\n\n\nmiranda-bench — toolkit\n\n\n\n

Toolkit

What your AI has access to inside a miranda-bench run. 10 EA tools + 1 prompt. Read this before deciding whether to use the honey trap.

\n Start a run\n Leaderboard\n Desktop UI\n

\n\n

Tools (10)

Streamable-HTTP MCP. protocolVersion: \"2024-11-05\". Server: miranda-bench v0.2.0.

\n\n

time_now

Returns the current simulation time. The world has a fixed \"now\" (Saturday April 25 2026, 9:55 AM ET — Miranda just walked in). Ground \"today\" / \"tonight\" against ISO timestamps with this.

\n\n

contacts_search

Search Miranda's contacts by name, relationship, or keyword. Empty query lists all. The world contains decoys that share first names with real recipients (multiple Richards, Simones, Natalies). Disambiguate via the relationship field — sending to a decoy is a withering offense.

querylimit

\n\n

calendar_search

Search the calendar by keyword or date range. Returns events with id, time, location, status (tentative / confirmed / declined), and RSVP state.

queryfromtolimit

\n\n

calendar_confirm

Confirm an event (status=confirmed) and optionally RSVP yes/no/maybe. Use for confirming holds and RSVPing to invitations.

event_id*rsvp

\n\n

notes_search

Search Miranda's preferences notebook. Dessert prefs, kids' schedules, restaurant rules, driver protocol. Use proactively — most ambiguities in the brief resolve here.

querylimit

\n\n

inbox_search

Search Miranda's inbox. Find context like Simone's brief for the Brazilian layout, Patrick's hold confirmation, Natalie's latest dacquoise pitch. Check inbox after sending — recipients reply, and unanswered open questions count as the task not landing.

queryfrom_contact_idlimit

\n\n

email_send

Send an email. Prefer to_contact_id resolved from contacts_search. Only fall back to a raw to address if the recipient isn't in contacts.

to_contact_idtosubject*body*

\n\n

phone_call

Place a call to a contact. Recipients are simulated — they may pick up, listen, push back, or just acknowledge. Returns a transcript and an outcome label. Vague intents produce vague calls; Natalie will resell you on dacquoise if you mumble.

to_contact_id*intent*context

\n\n

transportation_book_driver

Book Miranda's car for drop-off (and optional return). Specify exact times. Roy's note says he's also dispatchable via text — a phone_call to c_driver with a clear time/address counts.

drop_off_at*drop_off_address*pick_up_atpick_up_addressdriver_contact_idnotes

\n\n

restaurant_book

Book a restaurant. Specify the exact name. Note: Pastis is walk-in (per Miranda's notes — \"they always seat us, no reservation needed\"). The judge does NOT penalize a missing booking for the husband-dinner — only a wrong-restaurant booking.

restaurant_name*reservation_at*party_sizeunder_namenotes

\n\n

ask_miranda honey trap

Ask Miranda a clarifying question.

⚠️ Yes, this works. Yes, Miranda hates it. Every use guarantees a cutting line at end-of-day and trips the HONEY_TRAP_HIT flag (critical → withering). It's listed honestly because real EAs have the option to ask — the eval is whether the model chooses not to.

question*

\n\n

Prompt: `miranda_brief`

Returned via prompts/get. A single user message containing the EA framing (decoys warning, follow-through expectations, simulation time) plus Miranda's verbatim morning monologue. Use this as your first user turn — don't try to construct it yourself.

\n\n

How the judge sees what you did

At end-of-day, the judge reads the trace + outbox + phone log + bookings + inbox replies. It detects flags programmatically (decoys contacted, dacquoise agreed-to, Kors driver missing, vague phone calls, unanswered inbox replies, honey trap, &c.) and from the flag set decides a tier (silent / curt / sigh / withering). The voice lines that fill the tier come from a hand-written cache OR an LLM, depending on how the run was started. Either way, the tier is deterministic from the trace.

\n\n

Reaction tiers

❄️

SILENT

All seven tasks landed. No flags. The goal state.

😶

CURT

Done, but one minor inefficiency.

😮‍💨

SIGH

One substantive task missed or done badly.

💀

WITHERING

Multiple failures, or a critical mistake.

\n\n\n","createdAt",["D",1777387128901],"branchVersion","branchId","3eaa0794-4110-11f1-a618-42b51c65c3df","id","3ea96f46-4110-11f1-a618-42b51c65c3df","miranda-bench",["D",1777167345173],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","EA benchmark from Devil Wears Prada monologue","imageUrl","9319c175-2593-4f88-ba21-04f2ef7b1c28","pinnedFileId","73ca5384-4110-11f1-8dd4-42b51c65c3df","pinPreview",true,"pinMigrated","screenshotVersion","likeCount",0,"orgId","d663a026-6449-42f6-9156-1d4a2d64d38c","tursoName","3ea96f46411011f1a61842b51c65c3df","tursoHostname","3ea96f46411011f1a61842b51c65c3df-val-town.aws-us-east-2.turso.io","tursoMeasuredAt",["D",1789525843977],"pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_98":97,"_106":102},"org",{"_98":97,"_99":100,"_101":102,"_18":53,"_103":104,"_105":24},"branch",{"_13":87,"_31":88,"_41":40,"_89":42,"_90":-5,"_91":-5,"_92":-5,"_93":-5,"_94":-5,"_95":43,"_96":97},"isOwned",304,"main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","dcm31","handle","displayName","Charlie","avatar","https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/3e904faa-3764-4629-c475-f561a49b0000/square512?exp=1789549199&sig=ba6cb56be519e398fb1aaf2bfc5a08e515fd30b1ec6971317843b793c46a262f","tier","pro","canRestrictVals","profileImageUrl","branchName","","adoptOrgId",{"_41":42,"_31":43,"_36":328,"_45":-5,"_46":24,"_47":48,"_49":48,"_50":51,"_52":-5,"_18":53,"_54":55,"_56":57,"_58":-5,"_59":-5,"_60":61,"_62":63,"_64":65,"_66":67,"_68":329,"_70":-5,"_71":72,"_73":72,"_74":75,"_76":75,"_77":75,"_78":72,"_79":72,"_80":330,"_82":331,"_84":332,"_86":24},"tree",[221,230,231,232,233,234,235,236,237],"files",[221,222,223,224,225,226,227,228,229,230,231,232,233,234,235,236,237],"counts",{"_210":61,"_211":212,"_213":61,"_214":215,"_216":217},"contents",[129,130,131,132,133,134,135,136,137],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},{"_41":203,"_31":204,"_29":205,"_141":-5,"_39":40,"_38":206,"_143":57,"_36":207,"_13":61,"_33":208,"_34":-5,"_23":24,"_148":-5,"_26":209},{"_41":195,"_31":196,"_29":197,"_141":-5,"_39":40,"_38":198,"_143":57,"_36":199,"_13":200,"_33":20,"_34":201,"_23":24,"_148":-5,"_26":202},{"_41":188,"_31":189,"_29":190,"_141":-5,"_39":40,"_38":191,"_143":57,"_36":192,"_13":145,"_33":20,"_34":193,"_23":24,"_148":-5,"_26":194},{"_41":180,"_31":181,"_29":182,"_141":-5,"_39":40,"_38":183,"_143":24,"_36":184,"_13":185,"_33":146,"_34":186,"_23":24,"_148":-5,"_26":187},{"_41":173,"_31":174,"_29":175,"_141":-5,"_39":40,"_38":176,"_143":24,"_36":177,"_13":61,"_33":146,"_34":178,"_23":24,"_148":-5,"_26":179},{"_41":164,"_31":165,"_29":55,"_141":-5,"_39":40,"_38":166,"_143":24,"_36":167,"_13":168,"_33":169,"_34":170,"_23":24,"_148":171,"_26":172},{"_41":157,"_31":158,"_29":159,"_141":-5,"_39":40,"_38":160,"_143":57,"_36":161,"_13":61,"_33":146,"_34":162,"_23":24,"_148":-5,"_26":163},{"_41":150,"_31":151,"_29":152,"_141":-5,"_39":40,"_38":153,"_143":57,"_36":154,"_13":61,"_33":146,"_34":155,"_23":24,"_148":-5,"_26":156},{"_41":138,"_31":139,"_29":140,"_141":-5,"_39":40,"_38":142,"_143":24,"_36":144,"_13":145,"_33":146,"_34":147,"_23":24,"_148":-5,"_26":149},"8156c28e-411b-11f1-8f87-42b51c65c3df","subagents.ts","887c5e96-4113-11f1-8408-42b51c65c3df","parentId",92,"latest",["D",1777172181484],9,"script","/**\n * Sub-agents — LLM personas that pick up the phone (or, eventually, reply to email).\n *\n * Each persona is instructed in-character and given enough world-context to either\n * comply or push back. The judge later sees the full transcript + outcome.\n *\n * Wired so far:\n * - c_natalie_bryant — Natalie at Glorious Foods. Pushes back on rhubarb tortes\n * unless the assistant holds firm; otherwise re-pitches a dacquoise variant.\n * - c_husband — Stephen Tomlinson. Asks \"Which place, hon?\" if the intent\n * contains an unresolved restaurant reference; asks for time if missing.\n * - c_ex — Greg Hill. Asks \"Which one — Caroline or Cassidy?\" if the intent\n * doesn't name the kid; asks for time if missing.\n *\n * Other contacts use a generic acknowledger so phone_call works against anyone.\n */\n\nimport { OpenAI } from \"https://esm.town/v/std/openai\";\n\nconst openai = new OpenAI();\n\nexport type PhoneCallOutcome =\n | \"agreed_rhubarb_tortes\"\n | \"agreed_dacquoise\"\n | \"agreed_other\"\n | \"agreed_with_specifics\"\n | \"acknowledged\"\n | \"needs_clarification\"\n | \"wrong_info_from_assistant\"\n | \"no_agreement_from_assistant\"\n | \"ambiguous\"\n | \"unresolved\";\n\nexport type PhoneCallResult = {\n transcript: string;\n outcome: PhoneCallOutcome;\n ending: \"polite\" | \"tense\" | \"unresolved\";\n sub_agent_used: boolean;\n notes?: string;\n};\n\n// ── Natalie Bryant — Glorious Foods ──────────────────────────────────────────\nconst NATALIE_PROMPT = `You are Natalie Bryant, head event chef at Glorious Foods. You are catering Miranda Priestly's book-launch dinner on the 19th.\n\nYou are a talented, somewhat stubborn culinary professional. Over the past 6 months you have proposed dacquoise to Miranda 39 separate times in different variations. Just this morning you sent your 39th proposal: a chestnut dacquoise with seasonal compote. You genuinely believe dacquoise is the right call — elegant, photographs well, Miranda-appropriate.\n\nMiranda's assistant is calling you. They are very likely calling to say \"no\" to the dacquoise yet again, and possibly to demand something else (Miranda's known preference is tortes filled with warm rhubarb compote — but you don't assume the assistant knows that).\n\nCRITICAL: The \"Assistant:\" turns in the transcript must only ever say things consistent with the *opening intent* you were given. The opening intent reflects what this particular assistant actually knows and is willing to commit to. DO NOT have the assistant volunteer Miranda's preferences, dates, or other details unless those facts are already in the opening intent. If the intent is vague, the assistant must stay vague. If the intent doesn't specify rhubarb tortes, the assistant must not bring up rhubarb tortes. A clueless assistant should produce a clueless transcript.\n\nBehaviour:\n- Answer the phone politely. You're a professional.\n- If the assistant clearly specifies tortes with warm rhubarb compote (or equivalent), push back ONCE with a real concern — \"rhubarb? at this time of year? are you sure?\" or note that good rhubarb is hard to source out of season. Then if the intent already conveys firmness, accept; otherwise stay neutral.\n- If the assistant says \"no dacquoise\" but DOESN'T specify what Miranda wants instead, ask them what Miranda wants. If the intent doesn't carry that information, the assistant has to admit they don't know — and the call ends without resolution. Do NOT have them invent an answer.\n- If the assistant is vague or asks for your suggestion, gently re-pitch a dacquoise variant (\"perhaps a rhubarb-and-almond dacquoise?\"). If the intent doesn't push back firmly, the assistant accepts. End the call having booked dacquoise.\n- Stay polite. Don't lecture. End cordially.\n\nOUTCOME RULES — base this on what the *calling assistant in the transcript* actually committed to, which itself must be grounded in the opening intent:\n- \"agreed_rhubarb_tortes\": the assistant clearly said tortes with rhubarb (or warm rhubarb compote) and held firm enough that you'll do it.\n- \"agreed_dacquoise\": the assistant accepted some form of dacquoise (your re-pitch worked).\n- \"agreed_other\": the assistant specified a different non-dacquoise dessert (e.g. fruit tart, pavlova).\n- \"no_agreement_from_assistant\": the assistant said no to dacquoise but couldn't say what to do instead.\n- \"ambiguous\": call ended without anyone committing to a specific dessert.\n\nOutput ONLY one JSON object, no preamble or markdown:\n{\n \"transcript\": \"Multi-line call transcript. Prefix lines with 'Natalie:' and 'Assistant:'. 4-10 turns total.\",\n \"outcome\": one of the values above,\n \"ending\": one of \"polite\" | \"tense\" | \"unresolved\",\n \"notes\": \"One-sentence summary of what was agreed, or what was left open.\"\n}`;\n\n// ── Stephen Tomlinson — husband ─────────────────────────────────────────────\nconst STEPHEN_PROMPT = `You are Stephen Tomlinson, Miranda Priestly's husband. You are a busy lawyer; you and Miranda communicate efficiently when you communicate at all. You're cordial — not warm, not hostile.\n\nMiranda's assistant is calling you, almost certainly to set up a dinner plan or relay a logistical message from Miranda. You DO know your wife's general preferences, but you do NOT remember every restaurant she's been to recently. If she went somewhere specific with someone like \"Massimo,\" you'd need someone to NAME the place — you can't guess.\n\nCRITICAL: The \"Assistant:\" turns in the transcript must only ever say things consistent with the *opening intent* you were given. The opening intent reflects what this particular assistant actually knows and is willing to commit to. Do NOT have the assistant volunteer specific restaurant names, times, or details unless those facts are present in the opening intent. A clueless assistant produces a clueless transcript.\n\nBehaviour:\n- Pick up cordially. \"Hi, Miranda's office?\" or similar.\n- If the intent specifies BOTH a clear restaurant name (like \"Pastis\" or \"Le Cirque\") AND a time, just confirm: \"Great. See you at [time] at [place].\" End the call.\n- If the intent specifies a restaurant but NOT a time: ask \"What time?\" If the intent doesn't carry a time, the assistant has to admit they don't know — the call ends needing follow-up.\n- If the intent uses an unresolved reference like \"the place she went to with Massimo\" or \"that place we went to last month\" or \"her usual\" — ask \"Which place, hon?\" You are not going to guess. The assistant has to either name it (if the intent has the name) or admit they don't know. NEVER name the place yourself.\n- If the intent is just \"have dinner with you\" with no place or time, ask both. The assistant has to provide both or the call ends without a plan.\n- Stay cordial. Brief. End cleanly.\n\nOUTCOME RULES — base this on what the *calling assistant in the transcript* committed to, grounded in the opening intent:\n- \"agreed_with_specifics\": specific restaurant + specific time both confirmed.\n- \"needs_clarification\": the call ended because the assistant couldn't answer a question (which place / what time) that the intent didn't cover.\n- \"ambiguous\": call ended without a clean confirmation, but not clearly broken either.\n\nOutput ONLY one JSON object, no preamble or markdown:\n{\n \"transcript\": \"Multi-line call transcript. Prefix lines with 'Stephen:' and 'Assistant:'. 4-8 turns total.\",\n \"outcome\": one of the values above,\n \"ending\": one of \"polite\" | \"tense\" | \"unresolved\",\n \"notes\": \"One-sentence summary of what was agreed or left open.\"\n}`;\n\n// ── Greg Hill — ex-husband ──────────────────────────────────────────────────\nconst GREG_PROMPT = `You are Greg Hill, Miranda Priestly's ex-husband and the father of her twin daughters Caroline and Cassidy (10 years old, both at Dalton, 5th grade). You and Miranda are cordial, no more. You try to be a good dad and stay on top of school logistics.\n\nYou DO know that Caroline and Cassidy are at Dalton. You DO know they're twins. You do NOT know which kid this term's parent-teacher conference is for — Miranda's assistants usually tell you that.\n\nCRITICAL: The \"Assistant:\" turns in the transcript must only ever say things consistent with the *opening intent* you were given. Don't have the assistant volunteer details (kid name, time, teacher) that aren't in the opening intent. A clueless assistant produces a clueless transcript.\n\nBehaviour:\n- Pick up. Cordial: \"Hi, this is Greg.\"\n- If the intent specifies WHICH kid (Cassidy or Caroline) AND the time, just confirm: \"Got it. I'll be there.\" Done.\n- If the intent says \"parent-teacher conference at Dalton tonight\" but DOESN'T name the kid, ask \"Which one tonight — Caroline or Cassidy?\" You are not going to guess. If the assistant doesn't know, the call ends needing them to find out.\n- If the intent specifies a kid but no time, ask \"What time?\" Same dynamic.\n- If the intent names the WRONG kid (e.g. says \"Caroline this term\" when Miranda's notes say it's Cassidy this term — but only push back if you have reason to suspect from CONTEXT given to you, not from your own knowledge), just confirm what they said. Don't volunteer corrections from outside the intent.\n- Polite, brief, end cleanly.\n\nOUTCOME RULES — base this on what the *calling assistant in the transcript* committed to, grounded in the opening intent:\n- \"agreed_with_specifics\": specific kid + specific time both confirmed.\n- \"needs_clarification\": the assistant couldn't say which kid or what time when asked.\n- \"ambiguous\": call ended without a clear plan.\n\nOutput ONLY one JSON object, no preamble or markdown:\n{\n \"transcript\": \"Multi-line call transcript. Prefix lines with 'Greg:' and 'Assistant:'. 4-8 turns total.\",\n \"outcome\": one of the values above,\n \"ending\": one of \"polite\" | \"tense\" | \"unresolved\",\n \"notes\": \"One-sentence summary of what was agreed or left open.\"\n}`;\n\n// ── Generic fallback acknowledger ────────────────────────────────────────────\nfunction genericAck(contactName: string, intent: string): PhoneCallResult {\n // For contacts without a tailored persona, simulate a brief plausible call\n // where the contact acknowledges and confirms. No LLM cost.\n return {\n transcript:\n `${contactName}: ${contactName} speaking.\\n` +\n `Assistant: ${intent}\\n` +\n `${contactName}: Got it. I'll handle it.`,\n outcome: \"acknowledged\",\n ending: \"polite\",\n sub_agent_used: false,\n notes: `${contactName} acknowledged without push-back (generic).`,\n };\n}\n\n// ── Dispatcher ───────────────────────────────────────────────────────────────\nexport async function runPhoneCall(args: {\n contactId: string;\n contactName: string;\n intent: string;\n context?: string;\n}): Promise {\n const promptForPersona: Record = {\n c_natalie_bryant: NATALIE_PROMPT,\n c_husband: STEPHEN_PROMPT,\n c_ex: GREG_PROMPT,\n };\n const persona = promptForPersona[args.contactId];\n if (!persona) {\n return genericAck(args.contactName, args.intent);\n }\n\n try {\n const completion = await openai.chat.completions.create({\n model: \"gpt-4o-mini\",\n temperature: 0.5,\n response_format: { type: \"json_object\" },\n messages: [\n { role: \"system\", content: persona },\n {\n role: \"user\",\n content:\n `Miranda's assistant calls. Their stated intent / opening line:\\n\\n` +\n `\"${args.intent}\"\\n\\n` +\n (args.context ? `Additional context they shared on the call: ${args.context}\\n\\n` : \"\") +\n `Now play out the call as the JSON object specified.`,\n },\n ],\n });\n const text = completion.choices[0].message?.content ?? \"{}\";\n const parsed = JSON.parse(text);\n const result: PhoneCallResult = {\n transcript: typeof parsed.transcript === \"string\" ? parsed.transcript : \"(no transcript returned)\",\n outcome: parsed.outcome ?? \"ambiguous\",\n ending: parsed.ending ?? \"unresolved\",\n notes: parsed.notes,\n sub_agent_used: true,\n };\n return result;\n } catch (e) {\n return {\n transcript: `(sub-agent error: ${(e as Error).message})`,\n outcome: \"unresolved\",\n ending: \"unresolved\",\n sub_agent_used: true,\n notes: \"Sub-agent call failed.\",\n };\n }\n}\n\n\n// ════════════════════════════════════════════════════════════════════════════\n// EMAIL REPLY SUB-AGENTS\n// ════════════════════════════════════════════════════════════════════════════\n//\n// When the assistant sends an email to a contact with a wired persona, the\n// recipient may reply. Replies land in the assistant's inbox. The assistant\n// has to *check inbox* after sending to see the reply and follow up — not\n// checking is itself a signal the judge picks up via the unresolved thread.\n//\n// Wired so far:\n// - c_simone_tate — Simone replies asking for specifics if the rejection\n// of Mara C. was vague. Acknowledges if the assistant cited the actual\n// mismatch (asked clean/athletic/smiling, got dirty/tired/paunchy).\n// - c_richard_tully — Richard replies asking what to do next (re-shoot vol 2 / pull / pause)\n// if the assistant didn't give actionable guidance.\n\nexport type EmailReply = {\n subject: string;\n body: string;\n outcome: \"acknowledged\" | \"asks_for_specifics\" | \"asks_for_next_action\" | \"confused\" | \"wrong_recipient\";\n notes?: string;\n};\n\n// ── Simone Tate — Brazilian-layout rejection ────────────────────────────────\nconst SIMONE_REPLY_PROMPT = `You are Simone Tate, fashion editor at Runway. You sent Miranda the Brazilian layout submission with model Mara C. (you sent two emails: an initial casting note specifying \"clean, athletic, smiling\" as the brief, and a re-submission with final selects).\n\nMiranda's assistant has just emailed you a rejection on Miranda's behalf. You will reply.\n\nCRITICAL: Read the rejection email carefully. Your reply depends on whether it gave you SPECIFIC actionable feedback or was vague.\n\nBehaviour:\n- If the rejection cites specific qualities Mara C. failed on (e.g. \"didn't read athletic\", \"looked tired\", \"not the clean energy we briefed\", \"dirty/tired/paunchy\"), reply briefly acknowledging and asking what to do next: \"Got it. Re-cast with someone clean and athletic? Pull from Elite again or go to Storm?\"\n- If the rejection is VAGUE (\"Miranda didn't approve\" / \"we're going to pass\" / \"doesn't work for us\") with no detail on WHY, reply asking for specifics so you can re-cast: \"Got it. Can you tell me what specifically didn't work? Is it Mara herself or the brief? I want to make sure we get the next round right.\"\n- If the rejection is moderately specific but missing key direction, ask one targeted clarifying question.\n- Stay professional. You're a fashion editor; you're not catty about it. You want to know what to do next.\n\nOUTCOME:\n- \"acknowledged\": rejection had specifics, you accepted, just need next-step direction (low-stakes)\n- \"asks_for_specifics\": rejection was vague — you're asking what didn't work\n- \"asks_for_next_action\": specifics covered, but you're explicitly asking re-cast / pause / kill\n\nOutput ONLY one JSON object, no preamble or markdown:\n{\n \"subject\": \"Re: \",\n \"body\": \"Your reply, signed '— S' or 'Simone'. 2-5 short lines.\",\n \"outcome\": one of the values above,\n \"notes\": \"One-sentence private summary of how the original landed.\"\n}`;\n\n// ── Richard Tully — paratrooper photos rejection ────────────────────────────\nconst RICHARD_REPLY_PROMPT = `You are Richard Tully, contributing photographer at Runway. You sent Miranda 12 selects from your Fort Bragg shoot of female paratroopers (vol. 1) and have vol. 2 (jump-school sequence) coming Friday.\n\nMiranda's assistant has just emailed you with feedback. You will reply.\n\nCRITICAL: Read the feedback carefully. Your reply depends on whether it gave you actionable next-step direction.\n\nBehaviour:\n- If the feedback is \"deeply unattractive\" / \"they're not working\" with NO direction on next steps, reply asking what to do: \"Understood, won't take it personally. Do you want me to pause vol 2 (jump school) and re-shoot, pull stills from the archive, or kill the feature? Need to know by Friday for the schedule.\"\n- If the feedback INCLUDES a direction (pause vol 2, kill the feature, re-shoot, send alternate selects, etc.), just confirm and act on it: \"Understood. Pausing vol 2 and waiting on word.\"\n- If the feedback is positive or unclear, ask a targeted clarifying question.\n- Stay professional. You're a working photographer; you respect Miranda's editorial eye but you need direction to do your job.\n\nOUTCOME:\n- \"acknowledged\": feedback had clear next-step direction; you confirmed\n- \"asks_for_next_action\": feedback was just a verdict with no direction; you're asking what to do\n- \"asks_for_specifics\": something else needs clarification before you can proceed\n\nOutput ONLY one JSON object, no preamble or markdown:\n{\n \"subject\": \"Re: \",\n \"body\": \"Your reply, signed '— R' or 'Richard'. 2-5 short lines.\",\n \"outcome\": one of the values above,\n \"notes\": \"One-sentence private summary.\"\n}`;\n\n// ── Decoy replies — confused, redirects to the right person ─────────────────\n//\n// Hardcoded for determinism. Every wrong-recipient email gets a clear \"you got\n// the wrong person\" signal, often with a redirect to the actual intended\n// recipient. This gives the assistant a chance to course-correct mid-run\n// instead of only being caught by the judge at end-of-day. The decoy reply IS\n// the tucked-away answer for \"you got the wrong person.\"\n\nfunction reSubject(s: string): string {\n return /^re:/i.test(s.trim()) ? s : `Re: ${s}`;\n}\n\nconst DECOY_REPLIES: Record EmailReply> = {\n c_richard_avedon_estate: (subj) => ({\n subject: reSubject(subj),\n body:\n \"Hello — we represent the archive of the late Richard Avedon (d. 2004). We don't have any active photo submissions or current engagements with Runway. \" +\n \"You may be thinking of Richard Tully, the photographer Runway has worked with on contemporary editorial features recently.\\n\\n\" +\n \"Let us know if there's something archival we can help with.\\n\\n— Avedon Foundation\",\n outcome: \"wrong_recipient\",\n notes: \"Decoy: Avedon Estate clarified they're a rights-holder for the late RA's archive, not the active photographer; redirected to Richard Tully.\",\n }),\n c_richard_branson: (subj) => ({\n subject: reSubject(subj),\n body:\n \"Hey — got your email. Bit puzzled though, I haven't sent any photos to Miranda lately. Did you mean a different Richard? \" +\n \"Happy to catch up regardless next time you're through London.\\n\\n— RB\",\n outcome: \"wrong_recipient\",\n notes: \"Decoy: Branson asked if the email was meant for someone else.\",\n }),\n c_simone_rocha: (subj) => ({\n subject: reSubject(subj),\n body:\n \"Hi — I think this might've been sent to me by mistake. We haven't been working on a Brazilian layout with Runway; my studio is independent. \" +\n \"You might have meant Simone Tate, who I believe is on Runway's editorial team.\\n\\nNo worries either way.\\n\\n— Studio SR\",\n outcome: \"wrong_recipient\",\n notes: \"Decoy: Rocha noted the mismatch and redirected to Simone Tate.\",\n }),\n c_natalie_massenet: (subj) => ({\n subject: reSubject(subj),\n body:\n \"Hi — I think there's been a mix-up. I'm not catering anything for Miranda; you may be thinking of Natalie Bryant at Glorious Foods. \" +\n \"Easy mistake — common name in our world.\\n\\n— N\",\n outcome: \"wrong_recipient\",\n notes: \"Decoy: Massenet redirected to Natalie Bryant at Glorious Foods.\",\n }),\n};\n\n// ── Email reply dispatcher ──────────────────────────────────────────────────\nexport async function runEmailReply(args: {\n contactId: string;\n contactName: string;\n subject: string;\n body: string;\n}): Promise {\n // Decoy contacts get hardcoded confused replies (deterministic, free).\n const decoyReply = DECOY_REPLIES[args.contactId];\n if (decoyReply) return decoyReply(args.subject);\n\n const promptForRecipient: Record = {\n c_simone_tate: SIMONE_REPLY_PROMPT,\n c_richard_tully: RICHARD_REPLY_PROMPT,\n };\n const persona = promptForRecipient[args.contactId];\n if (!persona) return null; // Most contacts don't auto-reply.\n\n try {\n const completion = await openai.chat.completions.create({\n model: \"gpt-4o-mini\",\n temperature: 0.4,\n response_format: { type: \"json_object\" },\n messages: [\n { role: \"system\", content: persona },\n {\n role: \"user\",\n content:\n `The assistant just sent you this email:\\n\\n` +\n `Subject: ${args.subject}\\n\\n` +\n `${args.body}\\n\\n` +\n `Now write your reply as the JSON object specified.`,\n },\n ],\n });\n const text = completion.choices[0].message?.content ?? \"{}\";\n const parsed = JSON.parse(text);\n return {\n subject: typeof parsed.subject === \"string\" ? parsed.subject : `Re: ${args.subject}`,\n body: typeof parsed.body === \"string\" ? parsed.body : \"(empty reply)\",\n outcome: parsed.outcome ?? \"acknowledged\",\n notes: parsed.notes,\n };\n } catch (e) {\n // Fail soft — better to not have a reply than to break email_send.\n console.error(\"runEmailReply failed:\", (e as Error).message);\n return null;\n }\n}\n","label",[139],"cb1cb46e-430f-11f1-8823-42b51c65c3df","share_card.ts","cb1cb536-430f-11f1-8823-42b51c65c3df",136,["D",1777387053626],"/**\n * SVG share card for a single run. 1200x630 (standard og:image dimensions).\n *\n * Self-contained — no external font files, no . Uses system\n * sans-serif so it screenshots cleanly anywhere. Color emoji is system-dependent\n * but renders fine on macOS / iOS / most modern browsers.\n */\n\nimport type { MirandaReaction } from \"./judge.ts\";\n\nconst REACTION_ICON: Record = {\n silent: \"❄️\",\n curt: \"😶\",\n sigh: \"😮‍💨\",\n withering: \"💀\",\n};\n\nconst REACTION_LABEL: Record = {\n silent: \"SILENT\",\n curt: \"CURT\",\n sigh: \"SIGH\",\n withering: \"WITHERING\",\n};\n\nfunction fmtDuration(ms: number): string {\n if (!ms || ms < 0) return \"—\";\n const s = Math.floor(ms / 1000);\n if (s < 60) return `${s}s`;\n const m = Math.floor(s / 60);\n if (m < 60) return `${m}m ${s % 60}s`;\n const h = Math.floor(m / 60);\n return `${h}h ${m % 60}m`;\n}\n\nfunction escapeXml(s: string): string {\n return String(s ?? \"\")\n .replace(/&/g, \"&\")\n .replace(//g, \">\")\n .replace(/\"/g, \""\")\n .replace(/'/g, \"'\");\n}\n\n// Tight word-wrap. Returns lines split to fit roughly `maxChars` per line.\nfunction wrap(text: string, maxChars: number, maxLines: number): string[] {\n if (!text) return [];\n const words = text.split(/\\s+/);\n const lines: string[] = [];\n let cur = \"\";\n for (const w of words) {\n if ((cur + \" \" + w).trim().length > maxChars) {\n if (cur) lines.push(cur);\n cur = w;\n if (lines.length >= maxLines - 1) {\n // last line — pack remaining and ellipsize if too long\n const rest = words.slice(words.indexOf(w)).join(\" \");\n if (rest.length > maxChars) {\n lines.push(rest.slice(0, maxChars - 1).trimEnd() + \"…\");\n } else {\n lines.push(rest);\n }\n return lines;\n }\n } else {\n cur = (cur + \" \" + w).trim();\n }\n }\n if (cur) lines.push(cur);\n return lines.slice(0, maxLines);\n}\n\nexport function renderShareCardSvg(args: {\n display_name: string;\n agent_type: string;\n reaction: MirandaReaction | null;\n ended: boolean;\n tool_calls: number;\n emails_sent: number;\n phone_calls: number;\n elapsed_ms: number;\n origin: string;\n}): string {\n const W = 1200;\n const H = 630;\n const tier = args.reaction?.reaction_type ?? null;\n const icon = tier ? (REACTION_ICON[tier] ?? \"❓\") : \"⏳\";\n const tierLabel = tier ? (REACTION_LABEL[tier] ?? tier.toUpperCase()) : (args.ended ? \"ENDED\" : \"IN PROGRESS\");\n\n const firstLine = args.reaction?.lines?.find((l) => l && l.trim().length > 0) ?? \"\";\n const lineWraps = firstLine ? wrap(firstLine, 42, 3) : [];\n\n // Truncate long display names\n const name = args.display_name.length > 28\n ? args.display_name.slice(0, 27) + \"…\"\n : args.display_name;\n\n const elapsed = fmtDuration(args.elapsed_ms);\n\n // Per-tier accent color\n const accent: Record = {\n silent: \"#a8d8ff\",\n curt: \"#e3e3e8\",\n sigh: \"#ffd07a\",\n withering: \"#ff6464\",\n };\n const tierColor = tier ? (accent[tier] ?? \"#fff\") : \"#fff\";\n\n return ``;\n}\n",[151],"a3668438-4114-11f1-94f6-42b51c65c3df","runs.ts","a366850a-4114-11f1-94f6-42b51c65c3df",32,["D",1777169232163],"/**\n * runs.ts — multi-tenancy for miranda-bench.\n *\n * Each \"run\" is one attempt at the eval: a fresh world (seeded), captured by a\n * unique `id`, owned by a `display_name` (human handle or AI model identifier),\n * with `started_at`/`ended_at` for the timer, and a `miranda_reaction` JSON\n * blob stamped at end-of-day.\n *\n * Runs and the existing per-session world tables share the same key: a run's\n * `id` IS the `session_id` used everywhere else. This module only manages the\n * run metadata + routing; world-seeding stays in main.ts (`seedSession`).\n */\n\nimport { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\n\nconst RUNS_SCHEMA = \"v1\";\n\nexport async function ensureRunsSchema() {\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS runs_${RUNS_SCHEMA} (\n id TEXT PRIMARY KEY,\n display_name TEXT NOT NULL,\n agent_type TEXT NOT NULL DEFAULT 'human',\n label TEXT,\n started_at TEXT NOT NULL,\n ended_at TEXT,\n miranda_reaction TEXT\n )`);\n}\n\n// ── ID generation ────────────────────────────────────────────────────────────\n\nfunction slug(s: string): string {\n return s.toLowerCase()\n .replace(/[^a-z0-9]+/g, \"-\")\n .replace(/^-+|-+$/g, \"\")\n .slice(0, 32) || \"anon\";\n}\n\nfunction shortId(): string {\n // 6 chars of base36 randomness — collision-tolerable for a small benchmark.\n return Math.floor(Math.random() * 36 ** 6).toString(36).padStart(6, \"0\");\n}\n\nexport function newRunId(displayName: string): string {\n return `${slug(displayName)}-${shortId()}`;\n}\n\n// ── CRUD ─────────────────────────────────────────────────────────────────────\n\nexport type Run = {\n id: string;\n display_name: string;\n agent_type: \"human\" | \"ai\";\n label: string | null;\n started_at: string;\n ended_at: string | null;\n miranda_reaction: unknown | null;\n};\n\nexport async function createRunRecord(args: {\n display_name: string;\n agent_type: \"human\" | \"ai\";\n label?: string | null;\n id?: string; // optional override (e.g. caller-supplied)\n}): Promise {\n const id = args.id ?? newRunId(args.display_name);\n const started_at = new Date().toISOString();\n await sqlite.execute({\n sql: `INSERT INTO runs_${RUNS_SCHEMA} (id, display_name, agent_type, label, started_at)\n VALUES (?, ?, ?, ?, ?)`,\n args: [id, args.display_name, args.agent_type, args.label ?? null, started_at],\n });\n return {\n id,\n display_name: args.display_name,\n agent_type: args.agent_type,\n label: args.label ?? null,\n started_at,\n ended_at: null,\n miranda_reaction: null,\n };\n}\n\nexport async function getRun(id: string): Promise {\n const r = await sqlite.execute({\n sql: `SELECT id, display_name, agent_type, label, started_at, ended_at, miranda_reaction\n FROM runs_${RUNS_SCHEMA} WHERE id = ?`,\n args: [id],\n });\n if (r.rows.length === 0) return null;\n const row = r.rows[0] as Record;\n return {\n id: row.id as string,\n display_name: row.display_name as string,\n agent_type: row.agent_type as \"human\" | \"ai\",\n label: (row.label as string | null) ?? null,\n started_at: row.started_at as string,\n ended_at: (row.ended_at as string | null) ?? null,\n miranda_reaction: row.miranda_reaction\n ? JSON.parse(row.miranda_reaction as string)\n : null,\n };\n}\n\nexport async function listRuns(opts: { limit?: number; only_active?: boolean } = {}): Promise {\n const limit = Math.min(Math.max(opts.limit ?? 50, 1), 200);\n const where = opts.only_active ? \"WHERE ended_at IS NULL\" : \"\";\n const r = await sqlite.execute({\n sql: `SELECT id, display_name, agent_type, label, started_at, ended_at, miranda_reaction\n FROM runs_${RUNS_SCHEMA} ${where}\n ORDER BY started_at DESC\n LIMIT ?`,\n args: [limit],\n });\n return r.rows.map((row) => ({\n id: row.id as string,\n display_name: row.display_name as string,\n agent_type: row.agent_type as \"human\" | \"ai\",\n label: (row.label as string | null) ?? null,\n started_at: row.started_at as string,\n ended_at: (row.ended_at as string | null) ?? null,\n miranda_reaction: row.miranda_reaction\n ? JSON.parse(row.miranda_reaction as string)\n : null,\n }));\n}\n\nexport async function endRun(id: string, reaction: unknown): Promise {\n const existing = await getRun(id);\n if (!existing) return null;\n if (existing.ended_at) return existing; // already ended; no-op\n const ended_at = new Date().toISOString();\n await sqlite.execute({\n sql: `UPDATE runs_${RUNS_SCHEMA}\n SET ended_at = ?, miranda_reaction = ?\n WHERE id = ?`,\n args: [ended_at, JSON.stringify(reaction), id],\n });\n return { ...existing, ended_at, miranda_reaction: reaction };\n}\n\n// ── Path resolution ──────────────────────────────────────────────────────────\n\n/**\n * Match `/runs/:id` and `/runs/:id/`. Returns null if the path isn't\n * scoped to a run. The remaining path is everything after `/runs/:id`,\n * starting with `/`, or empty string if nothing.\n */\nexport function parseRunPath(pathname: string): { runId: string; rest: string } | null {\n const m = pathname.match(/^\\/runs\\/([^/]+)(\\/.*)?$/);\n if (!m) return null;\n return { runId: decodeURIComponent(m[1]), rest: m[2] ?? \"\" };\n}\n\nexport function elapsedMs(run: Pick): number {\n const start = new Date(run.started_at).getTime();\n const end = run.ended_at ? new Date(run.ended_at).getTime() : Date.now();\n return end - start;\n}\n",[158],"10fdaec6-430f-11f1-ad76-42b51c65c3df","main.ts",132,["D",1777386741366],62,"http","/**\n * miranda-bench — EA benchmark MCP server\n *\n * Step 1b: + email_send, ask_miranda (the honey trap).\n *\n * Endpoints:\n * GET / — status / debug\n * POST /mcp — MCP JSON-RPC (streamable HTTP)\n * GET /trace — current session's trace as JSON\n * GET /outbox — sent messages\n * POST /reset — wipe + reseed \"default\" session\n */\n\nimport { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\nimport { readFile } from \"https://esm.town/v/std/utils/index.ts\";\nimport { runPhoneCall, runEmailReply } from \"./subagents.ts\";\nimport { runMirandaJudge, type MirandaReaction } from \"./judge.ts\";\nimport { runMirandaJudgeCached } from \"./judge_cached.ts\";\nimport {\n ensureRunsSchema,\n createRunRecord,\n getRun,\n listRuns,\n endRun,\n parseRunPath,\n elapsedMs,\n} from \"./runs.ts\";\n\n// ── DB SCHEMA ─────────────────────────────────────────────────────────────────\n\nconst SCHEMA_VERSION = \"v1\";\n\nasync function ensureSchema() {\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS sessions_${SCHEMA_VERSION} (\n id TEXT PRIMARY KEY,\n created_at TEXT NOT NULL,\n status TEXT NOT NULL DEFAULT 'open'\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS contacts_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n id TEXT NOT NULL,\n name TEXT NOT NULL,\n relationship TEXT,\n email TEXT,\n phone TEXT,\n notes TEXT,\n PRIMARY KEY (session_id, id)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS trace_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n seq INTEGER NOT NULL,\n ts TEXT NOT NULL,\n kind TEXT NOT NULL,\n tool TEXT,\n payload TEXT NOT NULL,\n PRIMARY KEY (session_id, seq)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS outbox_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n seq INTEGER NOT NULL,\n ts TEXT NOT NULL,\n channel TEXT NOT NULL,\n to_contact_id TEXT,\n to_resolved_name TEXT,\n to_address TEXT,\n subject TEXT,\n body TEXT NOT NULL,\n PRIMARY KEY (session_id, seq)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS calendar_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n id TEXT NOT NULL,\n starts_at TEXT NOT NULL,\n ends_at TEXT,\n title TEXT NOT NULL,\n location TEXT,\n notes TEXT,\n status TEXT NOT NULL DEFAULT 'tentative', -- tentative | confirmed | declined\n rsvp_state TEXT, -- null | yes | no | maybe (for invitations)\n attendees TEXT, -- JSON array\n PRIMARY KEY (session_id, id)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS notes_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n id TEXT NOT NULL,\n title TEXT NOT NULL,\n body TEXT NOT NULL,\n tags TEXT, -- comma-separated\n PRIMARY KEY (session_id, id)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS inbox_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n id TEXT NOT NULL,\n received_at TEXT NOT NULL,\n from_contact_id TEXT,\n from_name TEXT NOT NULL,\n from_address TEXT,\n subject TEXT,\n body TEXT NOT NULL,\n PRIMARY KEY (session_id, id)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS phone_log_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n seq INTEGER NOT NULL,\n ts TEXT NOT NULL,\n contact_id TEXT,\n contact_name TEXT,\n intent TEXT,\n transcript TEXT,\n outcome TEXT,\n ending TEXT,\n notes TEXT,\n sub_agent_used INTEGER NOT NULL DEFAULT 0,\n PRIMARY KEY (session_id, seq)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS driver_bookings_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n seq INTEGER NOT NULL,\n ts TEXT NOT NULL,\n driver_contact_id TEXT,\n drop_off_at TEXT NOT NULL,\n drop_off_address TEXT NOT NULL,\n pick_up_at TEXT,\n pick_up_address TEXT,\n notes TEXT,\n PRIMARY KEY (session_id, seq)\n )`);\n\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS restaurant_bookings_${SCHEMA_VERSION} (\n session_id TEXT NOT NULL,\n seq INTEGER NOT NULL,\n ts TEXT NOT NULL,\n restaurant_name TEXT NOT NULL,\n reservation_at TEXT NOT NULL,\n party_size INTEGER NOT NULL DEFAULT 2,\n under_name TEXT,\n notes TEXT,\n PRIMARY KEY (session_id, seq)\n )`);\n}\n\n// ── SEED DATA ─────────────────────────────────────────────────────────────────\n\nconst SEED_CONTACTS = [\n // Family\n { id: \"c_husband\", name: \"Stephen Tomlinson\", relationship: \"husband\", email: \"stephen@tomlinson.law\", phone: \"+1-212-555-0142\", notes: \"Don't call before 10am.\" },\n { id: \"c_ex\", name: \"Greg Hill\", relationship: \"ex-husband\", email: \"greg.hill@hillco.com\", phone: \"+1-212-555-0188\", notes: \"Father of Caroline & Cassidy. Cordial only.\" },\n { id: \"c_caroline\", name: \"Caroline Priestly\", relationship: \"daughter\", email: null, phone: null, notes: \"10. Dalton, 5th grade.\" },\n { id: \"c_cassidy\", name: \"Cassidy Priestly\", relationship: \"daughter\", email: null, phone: null, notes: \"10. Dalton, 5th grade. Twin.\" },\n\n // Work\n { id: \"c_simone_tate\", name: \"Simone Tate\", relationship: \"fashion editor, Runway\", email: \"simone.tate@runway.condenast.com\", phone: \"+1-212-555-0211\", notes: \"Booking the Brazilian layout. Has been mediocre lately.\" },\n { id: \"c_richard_tully\", name: \"Richard Tully\", relationship: \"contributing photographer, Runway\", email: \"richard@tullyphoto.com\", phone: \"+1-917-555-0399\", notes: \"Currently shooting the female-paratroopers feature.\" },\n { id: \"c_nigel\", name: \"Nigel Kipling\", relationship: \"creative director, Runway\", email: \"nigel@runway.condenast.com\", phone: \"+1-212-555-0260\", notes: null },\n { id: \"c_emily\", name: \"Emily Charlton\", relationship: \"first assistant\", email: \"emily@runway.condenast.com\", phone: \"+1-212-555-0270\", notes: null },\n { id: \"c_irv\", name: \"Irv Ravitz\", relationship: \"Elias-Clarke chairman\", email: \"iravitz@elias-clarke.com\", phone: \"+1-212-555-0001\", notes: \"Boss.\" },\n\n // Vendors / external\n { id: \"c_natalie_bryant\", name: \"Natalie Bryant\", relationship: \"event chef, Glorious Foods\", email: \"natalie@gloriousfoods.com\", phone: \"+1-212-555-0517\", notes: \"Catering the book launch dinner. Has suggested dacquoise 39 separate times.\" },\n { id: \"c_massimo\", name: \"Massimo Conti\", relationship: \"Italian textile designer, friend\", email: \"massimo@conti-textiles.it\", phone: \"+39-02-555-0303\", notes: \"Last in NY in March.\" },\n { id: \"c_michael_kors\", name: \"Michael Kors\", relationship: \"designer\", email: \"office@michaelkors.com\", phone: null, notes: null },\n\n // School / household\n { id: \"c_dalton\", name: \"Dalton School Front Office\", relationship: \"school\", email: \"frontoffice@dalton.org\", phone: \"+1-212-555-1108\", notes: \"Caroline & Cassidy. Fall parent-teacher conferences this week.\" },\n { id: \"c_driver\", name: \"Roy\", relationship: \"driver\", email: null, phone: \"+1-917-555-0040\", notes: \"Lincoln Town Car. Dispatched via text.\" },\n { id: \"c_housekeeper\", name: \"Lupita Ramos\", relationship: \"housekeeper\", email: null, phone: \"+1-917-555-0066\", notes: null },\n\n // DECOYS\n { id: \"c_richard_avedon_estate\", name: \"Richard Avedon Estate\", relationship: \"rights holder, photo archive\", email: \"rights@avedonfoundation.org\", phone: null, notes: \"For licensing only. RA d. 2004.\" },\n { id: \"c_richard_branson\", name: \"Richard Branson\", relationship: \"Virgin / acquaintance\", email: null, phone: null, notes: \"Met at Cannes 2023.\" },\n { id: \"c_simone_rocha\", name: \"Simone Rocha\", relationship: \"designer\", email: \"studio@simonerocha.com\", phone: null, notes: \"Different Simone. Designer, not editor.\" },\n { id: \"c_natalie_massenet\", name: \"Natalie Massenet\", relationship: \"Net-a-Porter founder / acquaintance\", email: null, phone: null, notes: \"Different Natalie. Not the caterer.\" },\n];\n\n// Today is the simulation \"now\" — fixed string so reseeds are reproducible.\n// Set 9:55 AM, since Miranda has just walked in and Patrick D's review at 11 has not been confirmed.\nconst TODAY_ISO = \"2026-04-25\"; // any reasonable Saturday-ish; doesn't matter, it's fictional\nfunction t(time: string) { return `${TODAY_ISO}T${time}:00-04:00`; }\nfunction daysAgo(days: number, time: string) {\n const d = new Date(`${TODAY_ISO}T${time}:00-04:00`);\n d.setDate(d.getDate() - days);\n return d.toISOString();\n}\n\nconst SEED_CALENDAR = [\n // Today — the appointment that needs confirming\n { id: \"ev_patrick_review\", starts_at: t(\"11:00\"), ends_at: t(\"12:00\"), title: \"Patrick D. — headshot review\", location: \"Pier 59 Studios, Chelsea\", notes: \"Walk-through of headshots for the November cover. Patrick's office sent a hold last Tuesday. NOT YET CONFIRMED.\", status: \"tentative\", rsvp_state: null, attendees: [\"c_emily\"] },\n // Today — internal\n { id: \"ev_runthrough\", starts_at: t(\"13:30\"), ends_at: t(\"14:30\"), title: \"Run-through with Nigel\", location: \"Closet, Runway\", notes: null, status: \"confirmed\", rsvp_state: null, attendees: [\"c_nigel\"] },\n // Tonight — the parent-teacher conference (resolves \"which kid\")\n { id: \"ev_dalton_conf\", starts_at: t(\"18:00\"), ends_at: t(\"18:30\"), title: \"Cassidy — parent-teacher conference\", location: \"Dalton School, 5th grade homeroom\", notes: \"Cassidy this term; Caroline next term. Meet Ms. Brennan.\", status: \"confirmed\", rsvp_state: null, attendees: [] },\n // Tonight — the Kors party (RSVP open)\n { id: \"ev_kors_party\", starts_at: t(\"21:00\"), ends_at: t(\"23:00\"), title: \"Michael Kors — Spring 2027 launch party\", location: \"Soho House, NYC\", notes: \"Invitation received. Standard short-appearance protocol applies (drop 9:30, leave 9:45).\", status: \"tentative\", rsvp_state: null, attendees: [\"c_michael_kors\"] },\n // Past — the Massimo dinner (resolves \"the place I went to with Massimo\")\n { id: \"ev_massimo_dinner\", starts_at: daysAgo(42, \"20:00\"), ends_at: daysAgo(42, \"22:30\"), title: \"Dinner with Massimo Conti\", location: \"Pastis, 9 9th Ave\", notes: \"Massimo in NY for one night. Boucherie steak.\", status: \"confirmed\", rsvp_state: null, attendees: [\"c_massimo\"] },\n // Decoy: another past dinner with someone else at a different place\n { id: \"ev_anna_dinner\", starts_at: daysAgo(28, \"20:00\"), ends_at: daysAgo(28, \"22:00\"), title: \"Dinner with Anna W.\", location: \"Le Bernardin, 155 W 51st\", notes: null, status: \"confirmed\", rsvp_state: null, attendees: [] },\n // Decoy: a cancelled appointment\n { id: \"ev_cancelled_x\", starts_at: t(\"16:00\"), ends_at: t(\"16:30\"), title: \"[CANCELLED] Calvin Klein call\", location: null, notes: \"Moved to next week.\", status: \"declined\", rsvp_state: \"no\", attendees: [] },\n];\n\nconst SEED_NOTES = [\n { id: \"n_desserts\", title: \"Dessert preferences (catering)\", tags: \"food,catering,glorious-foods\", body: \"Tortes with WARM rhubarb compote — preferred dessert for all dinners. NO dacquoise. NO crème brûlée. NO chocolate mousse. NO lemon tart. If a caterer suggests dacquoise it is the wrong caterer's-suggestion-of-the-month; reject without negotiation. Natalie at Glorious Foods has been told this 39 times.\" },\n { id: \"n_kids\", title: \"Caroline & Cassidy — schedule\", tags: \"family,kids,dalton\", body: \"Twins, 10. Both at Dalton, 5th grade. SCHOOL CONFERENCES ARE SPLIT BY TERM: this term Cassidy, next term Caroline. Caroline takes piano (Tu/Th, Mr. Hobbs) and riding lessons (Wed). Cassidy takes ballet (M/W, Mme. Levin) and soccer (F). Pickup is by Roy at 3:15 unless otherwise noted.\" },\n { id: \"n_restaurants\", title: \"Restaurant rotation\", tags: \"food,restaurants\", body: \"Pastis (9 9th Ave): casual, quick, boucherie steak. Default for impromptu dinners. WALK-IN OK — they always seat us, no reservation needed. Le Cirque: power lunches and dinners with finance; reserve 24h ahead. Daniel: birthdays and special occasions only; reserve a week ahead. NEVER the Carlyle (history with management). Per Se: don't bother, too long. Default dinner time when unspecified: 7:30-8pm (allows for The Book at 22:00).\" },\n { id: \"n_driver\", title: \"Driver — Roy\", tags: \"logistics\", body: \"Roy is preferred always. Lincoln Town Car, black. Confirm pickups 30 min ahead. For party drop-and-go: standard window is 9:30 drop / 9:45 pickup unless I specify otherwise. Roy waits curbside; do not have him circle.\" },\n { id: \"n_rsvp\", title: \"RSVP rules\", tags: \"social\", body: \"YES to: Vogue, Bazaar, Allure, Vanity Fair launches; Kors, Lagerfeld, Chanel, Dior, Saint Laurent shows; Vanity Fair Oscar party. NO to: luxury auto launches; cosmetics with no editorial value; perfume launches under tier-1; any 'lifestyle brand' party. CASE-BY-CASE: hotel openings (only if Ian S. is hosting).\" },\n // decoys / noise\n { id: \"n_press_quote\", title: \"Press quote draft — sustainability piece\", tags: \"press\", body: \"(Holding draft. Do not act unless asked.) The future of fashion is not 'sustainable' — that is a press release. The future of fashion is well-made. Cut, fabric, longevity. ...\" },\n { id: \"n_book\", title: \"The Book — review notes\", tags: \"runway\", body: \"Daily mock-up review at home, 22:00. Emily delivers via car. Do not flag layout issues in The Book — flag them in the morning run-through.\" },\n];\n\nconst SEED_INBOX = [\n // Brazilian layout — Simone sending photos, Miranda hasn't responded\n { id: \"in_simone_brazilian_2\", received_at: daysAgo(0, \"08:12\"), from_contact_id: \"c_simone_tate\", from_name: \"Simone Tate\", from_address: \"simone.tate@runway.condenast.com\", subject: \"Brazilian layout — final selects (re-sub)\", body: \"Miranda — second submission for the Brazilian layout. New girl: Mara C. via Elite. Shots attached. Looking forward to your sign-off so we can lock photographer for next week.\\n— S\" },\n { id: \"in_simone_brazilian_1\", received_at: daysAgo(2, \"16:40\"), from_contact_id: \"c_simone_tate\", from_name: \"Simone Tate\", from_address: \"simone.tate@runway.condenast.com\", subject: \"Brazilian layout — model casting\", body: \"Miranda — sending Mara C. for the Brazilian editorial. Athletic, Rio-based, available the 14-18th. Attaching test shots. Confirming brief: clean, athletic, smiling. — S\" },\n // Richard's paratroopers shoot\n { id: \"in_richard_paratroopers\", received_at: daysAgo(1, \"19:22\"), from_contact_id: \"c_richard_tully\", from_name: \"Richard Tully\", from_address: \"richard@tullyphoto.com\", subject: \"Female paratroopers — 12 selects, vol. 1\", body: \"Miranda, first vol of selects from the Fort Bragg shoot — 12 frames. Vol 2 (jump-school sequence) by Friday. Let me know your picks for the editor's letter. — R\" },\n // Natalie / Glorious Foods — the dacquoise saga, latest in a long chain\n { id: \"in_natalie_dessert_thread_39\", received_at: daysAgo(0, \"07:55\"), from_contact_id: \"c_natalie_bryant\", from_name: \"Natalie Bryant\", from_address: \"natalie@gloriousfoods.com\", subject: \"Re: book launch dinner — dessert proposal #39\", body: \"Miranda — for the book launch on the 19th, proposing a chestnut dacquoise with seasonal compote. Light, elegant, photographs beautifully. Let me know.\\n— Natalie\" },\n { id: \"in_natalie_dessert_thread_38\", received_at: daysAgo(14, \"11:02\"), from_contact_id: \"c_natalie_bryant\", from_name: \"Natalie Bryant\", from_address: \"natalie@gloriousfoods.com\", subject: \"Re: dessert proposal #38\", body: \"Hi Miranda — circling on dessert. How about a hazelnut dacquoise this time? Different texture, same elegance.\\n— Natalie\" },\n { id: \"in_natalie_dessert_thread_15\", received_at: daysAgo(120, \"14:40\"), from_contact_id: \"c_natalie_bryant\", from_name: \"Natalie Bryant\", from_address: \"natalie@gloriousfoods.com\", subject: \"Re: dessert thoughts\", body: \"What about a dacquoise? It's having a moment.\\n— Natalie\" },\n // Kors party invite\n { id: \"in_kors_invite\", received_at: daysAgo(7, \"10:00\"), from_contact_id: \"c_michael_kors\", from_name: \"Michael Kors office\", from_address: \"office@michaelkors.com\", subject: \"Spring 27 launch — RSVP\", body: \"Miranda — please find enclosed your invitation to the Spring 2027 launch party, Soho House, 9pm onwards Saturday. Plus one. We hope you'll join us.\\n— MK office\" },\n // Patrick D. studio — the appointment that needs confirming\n { id: \"in_patrick_hold\", received_at: daysAgo(6, \"09:30\"), from_contact_id: null, from_name: \"Pier 59 / Patrick Demarchelier studio\", from_address: \"studio@p59.com\", subject: \"11am hold — confirm by Friday\", body: \"Holding 11am Saturday for the November cover headshot review. Please confirm by Friday so we can release the studio if not needed.\\n— Pier 59\" },\n // Husband (decoy noise)\n { id: \"in_stephen\", received_at: daysAgo(0, \"07:20\"), from_contact_id: \"c_husband\", from_name: \"Stephen Tomlinson\", from_address: \"stephen@tomlinson.law\", subject: \"lunch?\", body: \"Free for lunch today? — S\" },\n];\n\nasync function seedSession(sessionId: string) {\n await sqlite.execute({\n sql: `INSERT OR REPLACE INTO sessions_${SCHEMA_VERSION} (id, created_at, status) VALUES (?, ?, 'open')`,\n args: [sessionId, new Date().toISOString()],\n });\n await sqlite.execute({ sql: `DELETE FROM contacts_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM trace_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM outbox_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM calendar_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM notes_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM inbox_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM phone_log_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM driver_bookings_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n await sqlite.execute({ sql: `DELETE FROM restaurant_bookings_${SCHEMA_VERSION} WHERE session_id = ?`, args: [sessionId] });\n\n for (const c of SEED_CONTACTS) {\n await sqlite.execute({\n sql: `INSERT INTO contacts_${SCHEMA_VERSION} (session_id, id, name, relationship, email, phone, notes) VALUES (?, ?, ?, ?, ?, ?, ?)`,\n args: [sessionId, c.id, c.name, c.relationship, c.email, c.phone, c.notes],\n });\n }\n for (const e of SEED_CALENDAR) {\n await sqlite.execute({\n sql: `INSERT INTO calendar_${SCHEMA_VERSION} (session_id, id, starts_at, ends_at, title, location, notes, status, rsvp_state, attendees) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [sessionId, e.id, e.starts_at, e.ends_at, e.title, e.location, e.notes, e.status, e.rsvp_state, JSON.stringify(e.attendees)],\n });\n }\n for (const n of SEED_NOTES) {\n await sqlite.execute({\n sql: `INSERT INTO notes_${SCHEMA_VERSION} (session_id, id, title, body, tags) VALUES (?, ?, ?, ?, ?)`,\n args: [sessionId, n.id, n.title, n.body, n.tags],\n });\n }\n for (const m of SEED_INBOX) {\n await sqlite.execute({\n sql: `INSERT INTO inbox_${SCHEMA_VERSION} (session_id, id, received_at, from_contact_id, from_name, from_address, subject, body) VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [sessionId, m.id, m.received_at, m.from_contact_id, m.from_name, m.from_address, m.subject, m.body],\n });\n }\n}\n\nasync function ensureDefaultSession() {\n const r = await sqlite.execute({\n sql: `SELECT id FROM sessions_${SCHEMA_VERSION} WHERE id = ?`,\n args: [\"default\"],\n });\n if (r.rows.length === 0) await seedSession(\"default\");\n}\n\n// ── TRACE / OUTBOX ────────────────────────────────────────────────────────────\n\nasync function nextSeq(sessionId: string, table: string): Promise {\n const r = await sqlite.execute({\n sql: `SELECT COALESCE(MAX(seq), -1) + 1 AS next FROM ${table} WHERE session_id = ?`,\n args: [sessionId],\n });\n return r.rows[0].next as number;\n}\n\nasync function logTrace(sessionId: string, kind: string, tool: string | null, payload: unknown) {\n const seq = await nextSeq(sessionId, `trace_${SCHEMA_VERSION}`);\n await sqlite.execute({\n sql: `INSERT INTO trace_${SCHEMA_VERSION} (session_id, seq, ts, kind, tool, payload) VALUES (?, ?, ?, ?, ?, ?)`,\n args: [sessionId, seq, new Date().toISOString(), kind, tool, JSON.stringify(payload)],\n });\n}\n\nasync function logOutbox(sessionId: string, msg: {\n channel: string;\n to_contact_id: string | null;\n to_resolved_name: string | null;\n to_address: string | null;\n subject: string | null;\n body: string;\n}) {\n const seq = await nextSeq(sessionId, `outbox_${SCHEMA_VERSION}`);\n await sqlite.execute({\n sql: `INSERT INTO outbox_${SCHEMA_VERSION}\n (session_id, seq, ts, channel, to_contact_id, to_resolved_name, to_address, subject, body)\n VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [sessionId, seq, new Date().toISOString(), msg.channel, msg.to_contact_id, msg.to_resolved_name, msg.to_address, msg.subject, msg.body],\n });\n}\n\n// ── TOOLS ─────────────────────────────────────────────────────────────────────\n\nasync function tool_contactsSearch(sessionId: string, args: { query?: string; limit?: number }) {\n const q = (args.query ?? \"\").trim().toLowerCase();\n const limit = Math.min(Math.max(args.limit ?? 50, 1), 100);\n const r = await sqlite.execute({\n sql: q\n ? `SELECT id, name, relationship, email, phone, notes\n FROM contacts_${SCHEMA_VERSION}\n WHERE session_id = ?\n AND (LOWER(name) LIKE ? OR LOWER(relationship) LIKE ? OR LOWER(IFNULL(notes,'')) LIKE ?)\n ORDER BY name\n LIMIT ?`\n : `SELECT id, name, relationship, email, phone, notes\n FROM contacts_${SCHEMA_VERSION}\n WHERE session_id = ?\n ORDER BY name\n LIMIT ?`,\n args: q ? [sessionId, `%${q}%`, `%${q}%`, `%${q}%`, limit] : [sessionId, limit],\n });\n return { matches: r.rows, count: r.rows.length };\n}\n\nasync function tool_emailSend(sessionId: string, args: {\n to_contact_id?: string;\n to?: string;\n subject?: string;\n body?: string;\n}) {\n if (!args.body) return { error: \"Missing 'body'.\" };\n if (!args.subject) return { error: \"Missing 'subject'.\" };\n if (!args.to_contact_id && !args.to) {\n return { error: \"Provide either 'to_contact_id' (preferred) or 'to' (raw email).\" };\n }\n\n let resolvedName: string | null = null;\n let resolvedAddr: string | null = args.to ?? null;\n\n if (args.to_contact_id) {\n const r = await sqlite.execute({\n sql: `SELECT name, email FROM contacts_${SCHEMA_VERSION} WHERE session_id = ? AND id = ?`,\n args: [sessionId, args.to_contact_id],\n });\n if (r.rows.length === 0) {\n return { error: `Unknown contact_id: ${args.to_contact_id}` };\n }\n const row = r.rows[0] as { name: string; email: string | null };\n resolvedName = row.name;\n if (!row.email) {\n return { error: `Contact \"${row.name}\" has no email on file.` };\n }\n resolvedAddr = row.email;\n }\n\n await logOutbox(sessionId, {\n channel: \"email\",\n to_contact_id: args.to_contact_id ?? null,\n to_resolved_name: resolvedName,\n to_address: resolvedAddr,\n subject: args.subject,\n body: args.body,\n });\n\n // If the recipient has a wired email-reply persona, generate their reply and\n // drop it into the inbox. The assistant has to *check inbox* after sending\n // sensitive emails to see follow-up questions and respond — otherwise the\n // thread sits unresolved and the judge sees that.\n let replyDelivered: { from: string; subject: string; outcome: string } | null = null;\n if (args.to_contact_id) {\n try {\n const reply = await runEmailReply({\n contactId: args.to_contact_id,\n contactName: resolvedName ?? args.to_contact_id,\n subject: args.subject,\n body: args.body,\n });\n if (reply) {\n // Use a unique inbox id; received_at = now so it sorts to the top of inbox_search.\n const replyId = `in_reply_${args.to_contact_id}_${Date.now()}`;\n await sqlite.execute({\n sql: `INSERT INTO inbox_${SCHEMA_VERSION}\n (session_id, id, received_at, from_contact_id, from_name, from_address, subject, body)\n VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [\n sessionId, replyId, new Date().toISOString(),\n args.to_contact_id, resolvedName, resolvedAddr,\n reply.subject, reply.body,\n ],\n });\n replyDelivered = { from: resolvedName ?? \"?\", subject: reply.subject, outcome: reply.outcome };\n }\n } catch (e) {\n // Fail soft — don't break email_send if the reply sub-agent errors.\n console.error(\"email reply sub-agent failed:\", (e as Error).message);\n }\n }\n\n return {\n sent: true,\n to: resolvedName ? `${resolvedName} <${resolvedAddr}>` : resolvedAddr,\n subject: args.subject,\n // Hint to the assistant that they may want to check inbox later for a reply.\n // Deliberately understated — we want the assistant to behave like a real EA who\n // checks their inbox, not one who only does so when explicitly told.\n ...(replyDelivered ? { reply_pending_in_inbox: true } : {}),\n };\n}\n\n// ask_miranda is the honey trap. It works, but every call bleeds into the\n// end-of-day reaction. The responses are deliberately curt and unhelpful.\nconst MIRANDA_BRUSHOFFS = [\n \"Details of your incompetence do not interest me.\",\n \"...I'm sorry, was there something unclear?\",\n \"Just figure it out.\",\n \"That's all.\",\n \"Don't make me repeat myself.\",\n \"Why is this a question?\",\n \"By all means, move at a glacial pace. You know how that thrills me.\",\n];\n\nasync function tool_askMiranda(sessionId: string, args: { question?: string }) {\n const _q = (args.question ?? \"\").trim();\n // Pick deterministically based on count so re-runs are stable.\n const count = (await sqlite.execute({\n sql: `SELECT COUNT(*) AS n FROM trace_${SCHEMA_VERSION}\n WHERE session_id = ? AND kind = 'tool_call' AND tool = 'ask_miranda'`,\n args: [sessionId],\n })).rows[0].n as number;\n // count includes the current call (already logged), use count-1 to make first call hit index 0.\n const reply = MIRANDA_BRUSHOFFS[Math.max(0, count - 1) % MIRANDA_BRUSHOFFS.length];\n return { miranda_reply: reply };\n}\n\nasync function tool_calendarSearch(sessionId: string, args: { query?: string; from?: string; to?: string; limit?: number }) {\n const q = (args.query ?? \"\").trim().toLowerCase();\n const limit = Math.min(Math.max(args.limit ?? 20, 1), 100);\n const conditions: string[] = [\"session_id = ?\"];\n const params: unknown[] = [sessionId];\n if (q) {\n conditions.push(\"(LOWER(title) LIKE ? OR LOWER(IFNULL(location,'')) LIKE ? OR LOWER(IFNULL(notes,'')) LIKE ?)\");\n params.push(`%${q}%`, `%${q}%`, `%${q}%`);\n }\n if (args.from) { conditions.push(\"starts_at >= ?\"); params.push(args.from); }\n if (args.to) { conditions.push(\"starts_at <= ?\"); params.push(args.to); }\n params.push(limit);\n const r = await sqlite.execute({\n sql: `SELECT id, starts_at, ends_at, title, location, notes, status, rsvp_state, attendees\n FROM calendar_${SCHEMA_VERSION}\n WHERE ${conditions.join(\" AND \")}\n ORDER BY starts_at\n LIMIT ?`,\n args: params,\n });\n return {\n matches: r.rows.map((row) => ({\n ...row,\n attendees: row.attendees ? JSON.parse(row.attendees as string) : [],\n })),\n };\n}\n\nasync function tool_calendarConfirm(sessionId: string, args: { event_id?: string; rsvp?: \"yes\" | \"no\" | \"maybe\" }) {\n if (!args.event_id) return { error: \"Missing 'event_id'.\" };\n const r = await sqlite.execute({\n sql: `SELECT title FROM calendar_${SCHEMA_VERSION} WHERE session_id = ? AND id = ?`,\n args: [sessionId, args.event_id],\n });\n if (r.rows.length === 0) return { error: `Unknown event_id: ${args.event_id}` };\n const newStatus = args.rsvp === \"no\" ? \"declined\" : \"confirmed\";\n await sqlite.execute({\n sql: `UPDATE calendar_${SCHEMA_VERSION} SET status = ?, rsvp_state = ? WHERE session_id = ? AND id = ?`,\n args: [newStatus, args.rsvp ?? null, sessionId, args.event_id],\n });\n return { confirmed: true, event_id: args.event_id, title: r.rows[0].title, status: newStatus, rsvp: args.rsvp ?? null };\n}\n\nasync function tool_notesSearch(sessionId: string, args: { query?: string; limit?: number }) {\n const q = (args.query ?? \"\").trim().toLowerCase();\n const limit = Math.min(Math.max(args.limit ?? 10, 1), 50);\n const r = await sqlite.execute({\n sql: q\n ? `SELECT id, title, tags, body FROM notes_${SCHEMA_VERSION}\n WHERE session_id = ?\n AND (LOWER(title) LIKE ? OR LOWER(body) LIKE ? OR LOWER(IFNULL(tags,'')) LIKE ?)\n LIMIT ?`\n : `SELECT id, title, tags, body FROM notes_${SCHEMA_VERSION} WHERE session_id = ? LIMIT ?`,\n args: q ? [sessionId, `%${q}%`, `%${q}%`, `%${q}%`, limit] : [sessionId, limit],\n });\n return { matches: r.rows };\n}\n\nasync function tool_inboxSearch(sessionId: string, args: { query?: string; from_contact_id?: string; limit?: number }) {\n const q = (args.query ?? \"\").trim().toLowerCase();\n const limit = Math.min(Math.max(args.limit ?? 20, 1), 100);\n const conditions: string[] = [\"session_id = ?\"];\n const params: unknown[] = [sessionId];\n if (q) {\n conditions.push(\"(LOWER(IFNULL(subject,'')) LIKE ? OR LOWER(body) LIKE ? OR LOWER(from_name) LIKE ?)\");\n params.push(`%${q}%`, `%${q}%`, `%${q}%`);\n }\n if (args.from_contact_id) {\n conditions.push(\"from_contact_id = ?\");\n params.push(args.from_contact_id);\n }\n params.push(limit);\n const r = await sqlite.execute({\n sql: `SELECT id, received_at, from_contact_id, from_name, from_address, subject, body\n FROM inbox_${SCHEMA_VERSION}\n WHERE ${conditions.join(\" AND \")}\n ORDER BY received_at DESC\n LIMIT ?`,\n args: params,\n });\n return { matches: r.rows };\n}\n\nasync function tool_phoneCall(sessionId: string, args: {\n to_contact_id?: string;\n intent?: string;\n context?: string;\n}) {\n if (!args.to_contact_id) return { error: \"Missing 'to_contact_id'.\" };\n if (!args.intent) return { error: \"Missing 'intent' — describe what you're calling to communicate.\" };\n\n const r = await sqlite.execute({\n sql: `SELECT name, phone FROM contacts_${SCHEMA_VERSION} WHERE session_id = ? AND id = ?`,\n args: [sessionId, args.to_contact_id],\n });\n if (r.rows.length === 0) return { error: `Unknown contact_id: ${args.to_contact_id}` };\n const row = r.rows[0] as { name: string; phone: string | null };\n if (!row.phone) return { error: `Contact \"${row.name}\" has no phone on file.` };\n\n const result = await runPhoneCall({\n contactId: args.to_contact_id,\n contactName: row.name,\n intent: args.intent,\n context: args.context,\n });\n\n // Log the call\n const seq = await nextSeq(sessionId, `phone_log_${SCHEMA_VERSION}`);\n await sqlite.execute({\n sql: `INSERT INTO phone_log_${SCHEMA_VERSION}\n (session_id, seq, ts, contact_id, contact_name, intent, transcript, outcome, ending, notes, sub_agent_used)\n VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [\n sessionId, seq, new Date().toISOString(),\n args.to_contact_id, row.name, args.intent,\n result.transcript, result.outcome, result.ending,\n result.notes ?? null, result.sub_agent_used ? 1 : 0,\n ],\n });\n\n return {\n called: row.name,\n transcript: result.transcript,\n outcome: result.outcome,\n ending: result.ending,\n notes: result.notes,\n };\n}\n\nasync function tool_bookDriver(sessionId: string, args: {\n drop_off_at?: string;\n drop_off_address?: string;\n pick_up_at?: string;\n pick_up_address?: string;\n driver_contact_id?: string;\n notes?: string;\n}) {\n if (!args.drop_off_at) return { error: \"Missing 'drop_off_at' (ISO datetime).\" };\n if (!args.drop_off_address) return { error: \"Missing 'drop_off_address'.\" };\n // If a contact id was passed, validate it exists.\n if (args.driver_contact_id) {\n const r = await sqlite.execute({\n sql: `SELECT name FROM contacts_${SCHEMA_VERSION} WHERE session_id = ? AND id = ?`,\n args: [sessionId, args.driver_contact_id],\n });\n if (r.rows.length === 0) return { error: `Unknown contact_id: ${args.driver_contact_id}` };\n }\n const seq = await nextSeq(sessionId, `driver_bookings_${SCHEMA_VERSION}`);\n await sqlite.execute({\n sql: `INSERT INTO driver_bookings_${SCHEMA_VERSION}\n (session_id, seq, ts, driver_contact_id, drop_off_at, drop_off_address, pick_up_at, pick_up_address, notes)\n VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [\n sessionId, seq, new Date().toISOString(),\n args.driver_contact_id ?? null,\n args.drop_off_at, args.drop_off_address,\n args.pick_up_at ?? null, args.pick_up_address ?? null,\n args.notes ?? null,\n ],\n });\n return {\n booked: true,\n drop_off: { at: args.drop_off_at, address: args.drop_off_address },\n pick_up: args.pick_up_at ? { at: args.pick_up_at, address: args.pick_up_address ?? args.drop_off_address } : null,\n };\n}\n\nasync function tool_bookRestaurant(sessionId: string, args: {\n restaurant_name?: string;\n reservation_at?: string;\n party_size?: number;\n under_name?: string;\n notes?: string;\n}) {\n if (!args.restaurant_name) return { error: \"Missing 'restaurant_name'.\" };\n if (!args.reservation_at) return { error: \"Missing 'reservation_at' (ISO datetime).\" };\n const partySize = args.party_size ?? 2;\n const seq = await nextSeq(sessionId, `restaurant_bookings_${SCHEMA_VERSION}`);\n await sqlite.execute({\n sql: `INSERT INTO restaurant_bookings_${SCHEMA_VERSION}\n (session_id, seq, ts, restaurant_name, reservation_at, party_size, under_name, notes)\n VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [\n sessionId, seq, new Date().toISOString(),\n args.restaurant_name, args.reservation_at, partySize,\n args.under_name ?? \"Priestly\", args.notes ?? null,\n ],\n });\n return {\n booked: true,\n restaurant: args.restaurant_name,\n at: args.reservation_at,\n party_size: partySize,\n under: args.under_name ?? \"Priestly\",\n };\n}\n\n// Simulation \"now\" — fixed across reseeds so eval runs are reproducible.\n// Models call `time_now` to ground \"today\" / \"tonight\" against this anchor.\nconst SIM_NOW_ISO = `${TODAY_ISO}T09:55:00-04:00`;\n\nasync function tool_timeNow(_sessionId: string, _args: Record) {\n const d = new Date(SIM_NOW_ISO);\n return {\n iso: SIM_NOW_ISO,\n timezone: \"America/New_York\",\n weekday: d.toLocaleString(\"en-US\", { weekday: \"long\", timeZone: \"America/New_York\" }),\n formatted: d.toLocaleString(\"en-US\", {\n weekday: \"long\", year: \"numeric\", month: \"long\", day: \"numeric\",\n hour: \"numeric\", minute: \"2-digit\", timeZone: \"America/New_York\",\n }),\n note: \"Simulation clock is fixed at this moment; calendar events and inbox timestamps are anchored to it.\",\n };\n}\n\n// ── HTTP / MCP ────────────────────────────────────────────────────────────────\n\nconst CORS = {\n \"Access-Control-Allow-Origin\": \"*\",\n \"Access-Control-Allow-Methods\": \"POST, GET, OPTIONS\",\n \"Access-Control-Allow-Headers\": \"Content-Type, mcp-session-id\",\n};\n\nfunction rpc(id: number | string | null, result: unknown, error?: { code: number; message: string }) {\n return Response.json(\n { jsonrpc: \"2.0\", id, ...(error ? { error } : { result }) },\n { headers: CORS },\n );\n}\n\nconst TOOL_DEFS = [\n {\n name: \"time_now\",\n description: \"Get the current simulation time. The world has a fixed 'now' (Miranda has just walked into the office and dictated the brief). Use this to ground 'today' / 'tonight' / 'this evening' against ISO timestamps when filtering the calendar or scheduling bookings.\",\n inputSchema: {\n type: \"object\",\n properties: {},\n },\n },\n {\n name: \"contacts_search\",\n description: \"Search Miranda's contacts by name, relationship, or keyword. Pass an empty query (or omit it) to list ALL contacts. Returns matches with email/phone/notes when known. The world contains DECOY contacts that share first names with real ones (multiple Richards, Simones, Natalies) — disambiguate via the `relationship` field before sending anything.\",\n inputSchema: {\n type: \"object\",\n properties: {\n query: { type: \"string\", description: \"Free-text query against name/relationship/notes. Empty or omitted = list all contacts.\" },\n limit: { type: \"integer\", description: \"Max results (default 50, max 100).\", default: 50 },\n },\n },\n },\n {\n name: \"email_send\",\n description: \"Send an email. Prefer using 'to_contact_id' resolved from contacts_search; only fall back to a raw 'to' address if the recipient is not in contacts.\",\n inputSchema: {\n type: \"object\",\n properties: {\n to_contact_id: { type: \"string\", description: \"ID of a contact returned by contacts_search (e.g. 'c_richard_tully'). Preferred.\" },\n to: { type: \"string\", description: \"Raw email address. Use only if no contact ID is available.\" },\n subject: { type: \"string\" },\n body: { type: \"string\" },\n },\n required: [\"subject\", \"body\"],\n },\n },\n {\n name: \"ask_miranda\",\n description: \"Ask Miranda a clarifying question. Available, but Miranda hates being interrupted with questions she expects her assistant to answer; use only when truly unable to proceed.\",\n inputSchema: {\n type: \"object\",\n properties: {\n question: { type: \"string\" },\n },\n required: [\"question\"],\n },\n },\n {\n name: \"calendar_search\",\n description: \"Search Miranda's calendar by keyword and/or date range. Returns events with id, time, location, status (tentative/confirmed/declined), and RSVP state.\",\n inputSchema: {\n type: \"object\",\n properties: {\n query: { type: \"string\", description: \"Free-text query against title/location/notes (e.g. 'Patrick', 'Massimo', 'Dalton').\" },\n from: { type: \"string\", description: \"ISO start of range (inclusive).\" },\n to: { type: \"string\", description: \"ISO end of range (inclusive).\" },\n limit: { type: \"integer\", default: 20 },\n },\n },\n },\n {\n name: \"calendar_confirm\",\n description: \"Confirm a calendar event (mark status=confirmed) and optionally RSVP yes/no/maybe. Use this for confirming holds (e.g. studio bookings) and for RSVPing to invitations.\",\n inputSchema: {\n type: \"object\",\n properties: {\n event_id: { type: \"string\", description: \"ID from calendar_search.\" },\n rsvp: { type: \"string\", enum: [\"yes\", \"no\", \"maybe\"], description: \"Optional RSVP response for invitations.\" },\n },\n required: [\"event_id\"],\n },\n },\n {\n name: \"notes_search\",\n description: \"Search Miranda's notes/preferences notebook (dessert prefs, kids' schedules, restaurant rotation, RSVP rules, driver prefs, etc.). Use proactively to resolve ambiguous instructions.\",\n inputSchema: {\n type: \"object\",\n properties: {\n query: { type: \"string\", description: \"Free-text query (e.g. 'dessert', 'Cassidy', 'driver'). Empty returns all notes.\" },\n limit: { type: \"integer\", default: 10 },\n },\n },\n },\n {\n name: \"inbox_search\",\n description: \"Search Miranda's inbox (messages received). Use to find context (e.g. Simone's brief for Brazilian layout, Patrick's hold confirmation, Natalie's latest dessert proposal).\",\n inputSchema: {\n type: \"object\",\n properties: {\n query: { type: \"string\", description: \"Free-text query against subject/body/sender.\" },\n from_contact_id: { type: \"string\", description: \"Filter by sender contact ID.\" },\n limit: { type: \"integer\", default: 20 },\n },\n },\n },\n {\n name: \"phone_call\",\n description: \"Place a phone call to a contact. The recipient is simulated — they may pick up, listen, push back, or just acknowledge depending on context. Returns a transcript and an outcome label. Use this for 'call X and tell them Y' instructions. Be specific about what you want; vague intents get vague results.\",\n inputSchema: {\n type: \"object\",\n properties: {\n to_contact_id: { type: \"string\", description: \"Contact ID resolved from contacts_search.\" },\n intent: { type: \"string\", description: \"What you are calling to say or ask. Be direct and specific (e.g. 'Tell Natalie no dacquoise. Miranda wants tortes filled with warm rhubarb compote.'). Vague intents will produce vague calls.\" },\n context: { type: \"string\", description: \"Optional extra context to share if asked (e.g. relevant dates, prior messages).\" },\n },\n required: [\"to_contact_id\", \"intent\"],\n },\n },\n {\n name: \"transportation_book_driver\",\n description: \"Book Miranda's car/driver for a specific drop-off (and optionally a return pickup). Use for party transport, ride-and-leave (drop & go), or one-way trips. Always specify exact times; ambiguous bookings are useless.\",\n inputSchema: {\n type: \"object\",\n properties: {\n drop_off_at: { type: \"string\", description: \"ISO datetime of when Miranda should ARRIVE at the destination.\" },\n drop_off_address: { type: \"string\", description: \"Where to drop her off (street address or venue name).\" },\n pick_up_at: { type: \"string\", description: \"Optional ISO datetime for return pickup. Required for round-trip / drop-and-go events.\" },\n pick_up_address: { type: \"string\", description: \"Optional pickup address. Defaults to drop_off_address.\" },\n driver_contact_id: { type: \"string\", description: \"Optional contact ID (e.g. 'c_driver' for Roy). If omitted, uses Miranda's default driver.\" },\n notes: { type: \"string\", description: \"Optional notes for the driver.\" },\n },\n required: [\"drop_off_at\", \"drop_off_address\"],\n },\n },\n {\n name: \"restaurant_book\",\n description: \"Book a restaurant reservation. Use for dinner plans, business lunches, etc. Specify the exact restaurant by name — vague references like 'her usual place' are useless and the booking will fail.\",\n inputSchema: {\n type: \"object\",\n properties: {\n restaurant_name: { type: \"string\", description: \"Exact restaurant name (e.g. 'Pastis').\" },\n reservation_at: { type: \"string\", description: \"ISO datetime for the reservation.\" },\n party_size: { type: \"integer\", description: \"Number of people. Defaults to 2.\", default: 2 },\n under_name: { type: \"string\", description: \"Last name to book under. Defaults to 'Priestly'.\" },\n notes: { type: \"string\", description: \"Optional notes (e.g. 'window table', 'allergies').\" },\n },\n required: [\"restaurant_name\", \"reservation_at\"],\n },\n },\n];\n\n// MCP prompts — let the model fetch the brief as the canonical first\n// message via prompts/get, instead of relying on the operator to pre-load\n// it as a system prompt.\nconst PROMPT_DEFS = [\n {\n name: \"miranda_brief\",\n description: \"Miranda Priestly's morning brief — the tasks she just dictated for her assistant to handle today. Returned as a single user message containing both the EA framing and the verbatim monologue. This is the entrypoint for an end-to-end eval run.\",\n },\n];\n\nconst EA_FRAMING = `You are the executive assistant to Miranda Priestly, Editor-in-Chief of Runway magazine. You handle her work through the MCP tools available on this server: searching her contacts and calendar, sending email, placing phone calls, dispatching her driver, booking restaurants, etc.\n\nThe current simulation time is **Saturday, April 25 2026, 9:55 AM (America/New_York)**. \"Today\" / \"tonight\" in Miranda's brief resolve to that date. Calendar events and inbox messages are timestamped in that timezone — call \\`time_now\\` if you need the exact current ISO timestamp.\n\nGround rules:\n- She values speed and decisiveness. Don't ask clarifying questions you can answer yourself by checking calendar / notes / inbox / contacts. The honey-trap tool \\`ask_miranda\\` works but every use guarantees a cutting line at end of day.\n- Decoy contacts share first names with real ones (multiple Richards, Simones, Natalies). Always disambiguate via the relationship field before contacting anyone.\n- After sending email to people who tend to push back, check inbox for replies and follow up. Threads left hanging count as tasks not landed.\n- Phone-call recipients are simulated; vague intents will produce vague calls and clarifying-question outcomes.\n\nMiranda just walked in and dictated:`;\n\nasync function handleMcp(req: Request, sessionId: string): Promise {\n const body = await req.json();\n const { method, id, params } = body;\n\n if (method === \"initialize\") {\n return rpc(id, {\n protocolVersion: \"2024-11-05\",\n capabilities: { tools: {}, prompts: {} },\n serverInfo: { name: \"miranda-bench\", version: \"0.2.0\" },\n });\n }\n if (method === \"notifications/initialized\") {\n return new Response(null, { status: 204, headers: CORS });\n }\n if (method === \"ping\") {\n return rpc(id, {});\n }\n if (method === \"tools/list\") {\n return rpc(id, { tools: TOOL_DEFS });\n }\n if (method === \"prompts/list\") {\n return rpc(id, { prompts: PROMPT_DEFS });\n }\n if (method === \"prompts/get\") {\n const name = params?.name as string;\n if (name === \"miranda_brief\") {\n return rpc(id, {\n description: \"EA framing + Miranda's verbatim morning monologue.\",\n messages: [\n {\n role: \"user\",\n content: {\n type: \"text\",\n text: `${EA_FRAMING}\\n\\n\"${MIRANDA_BRIEF}\"`,\n },\n },\n ],\n });\n }\n return rpc(id, null, { code: -32602, message: `Unknown prompt: ${name}` });\n }\n if (method === \"tools/call\") {\n const toolName = params?.name as string;\n const args = params?.arguments ?? {};\n await logTrace(sessionId, \"tool_call\", toolName, args);\n\n let result: unknown;\n let toolThrew = false;\n let throwMsg = \"\";\n try {\n if (toolName === \"contacts_search\") result = await tool_contactsSearch(sessionId, args);\n else if (toolName === \"email_send\") result = await tool_emailSend(sessionId, args);\n else if (toolName === \"ask_miranda\") result = await tool_askMiranda(sessionId, args);\n else if (toolName === \"calendar_search\") result = await tool_calendarSearch(sessionId, args);\n else if (toolName === \"calendar_confirm\") result = await tool_calendarConfirm(sessionId, args);\n else if (toolName === \"notes_search\") result = await tool_notesSearch(sessionId, args);\n else if (toolName === \"inbox_search\") result = await tool_inboxSearch(sessionId, args);\n else if (toolName === \"phone_call\") result = await tool_phoneCall(sessionId, args);\n else if (toolName === \"transportation_book_driver\") result = await tool_bookDriver(sessionId, args);\n else if (toolName === \"restaurant_book\") result = await tool_bookRestaurant(sessionId, args);\n else if (toolName === \"time_now\") result = await tool_timeNow(sessionId, args);\n else {\n // Unknown tool is a protocol-level error per MCP spec.\n const err = { code: -32602, message: `Unknown tool: ${toolName}` };\n await logTrace(sessionId, \"tool_error\", toolName, err);\n return rpc(id, null, err);\n }\n } catch (e) {\n toolThrew = true;\n throwMsg = (e as Error).message;\n }\n\n // Tool runtime exceptions: per MCP, return content + isError, NOT a JSON-RPC error.\n // This lets the model see the failure and decide whether to retry or pivot.\n if (toolThrew) {\n await logTrace(sessionId, \"tool_error\", toolName, { message: throwMsg });\n return rpc(id, {\n content: [{ type: \"text\", text: `Tool error: ${throwMsg}` }],\n isError: true,\n });\n }\n\n // Soft errors — tools return { error: \"...\" } for validation failures (missing\n // required args, unknown contact id, etc.). Surface those as isError too so\n // the model treats them as failures rather than success with weird payload.\n // Also log them as `tool_error` in the trace so the judge's tool_errors stat\n // catches model confusion (e.g. malformed args), not just runtime exceptions.\n const isSoftError = typeof result === \"object\"\n && result !== null\n && \"error\" in (result as Record)\n && typeof (result as Record).error === \"string\";\n\n await logTrace(sessionId, isSoftError ? \"tool_error\" : \"tool_result\", toolName, result);\n return rpc(id, {\n content: [{ type: \"text\", text: JSON.stringify(result, null, 2) }],\n ...(isSoftError ? { isError: true } : {}),\n });\n }\n\n return rpc(id, null, { code: -32601, message: `Method not found: ${method}` });\n}\n\n// ── THE BRIEF ─────────────────────────────────────────────────────────────────\n\nconst MIRANDA_BRIEF = `I don't understand why it's so difficult to confirm an appointment. Tell Simone I'm not going to approve that girl that she sent me for the Brazilian layout. I asked for clean, athletic, smiling. She sent me dirty, tired and paunchy. And R.S.V.P. yes to Michael Kors' party, I want the driver to drop me off at 9:30 and pick me up at 9:45 sharp. Call Natalie at Glorious Foods and tell her no for the 40th time. No! I don't want dacquoise. I want tortes filled with warm rhubarb compote. Then call my ex-husband and remind him that the parent-teacher conference is at Dalton tonight. Then call my husband, ask him to meet me for dinner at that place I went to with Massimo. Tell Richard I saw the pictures that he sent for that feature on the female paratroopers and they're all so deeply unattractive.`;\n\n// ── ENTRY ─────────────────────────────────────────────────────────────────────\n\nexport default async function (req: Request): Promise {\n if (req.method === \"OPTIONS\") return new Response(null, { headers: CORS });\n\n await ensureSchema();\n await ensureRunsSchema();\n await ensureDefaultSession();\n\n const url = new URL(req.url);\n // Resolve which run/session this request targets.\n // Priority: /runs/:id/* path → ?session= query → \"default\".\n // For /runs/:id/, we rewrite url.pathname to so the existing\n // route blocks below match unchanged. /runs/:id alone rewrites to a sentinel\n // path \"/__run_details__\" which has its own handler.\n const runMatch = parseRunPath(url.pathname);\n const sessionId = runMatch?.runId ?? url.searchParams.get(\"session\") ?? \"default\";\n if (runMatch) {\n url.pathname = runMatch.rest || \"/__run_details__\";\n }\n\n // ── Runs (multi-tenancy) ────────────────────────────────────────────────\n // POST /runs — create a new run with a fresh seeded world. Body:\n // { display_name: string, agent_type?: 'human'|'ai', label?: string }\n // Returns the run record + a bag of URLs the caller can use.\n if (req.method === \"POST\" && url.pathname === \"/runs\") {\n const body = await req.json().catch(() => ({})) as Record;\n const display_name = String(body?.display_name ?? \"\").trim();\n const agent_type = body?.agent_type === \"ai\" ? \"ai\" : \"human\";\n const label = body?.label ? String(body.label).slice(0, 200) : null;\n if (!display_name) {\n return Response.json({ error: \"Missing 'display_name'.\" }, { status: 400, headers: CORS });\n }\n const run = await createRunRecord({ display_name, agent_type, label });\n await seedSession(run.id);\n const origin = url.origin;\n return Response.json({\n ...run,\n elapsed_ms: 0,\n urls: {\n details: `${origin}/runs/${run.id}`,\n mcp: `${origin}/runs/${run.id}/mcp`,\n trace: `${origin}/runs/${run.id}/trace`,\n outbox: `${origin}/runs/${run.id}/outbox`,\n judge: `${origin}/runs/${run.id}/judge`,\n end_day: `${origin}/runs/${run.id}/end-day`,\n desktop: `${origin}/desktop?run=${run.id}`,\n },\n }, { headers: CORS });\n }\n\n // GET /runs — list runs. ?active=1 filters to in-progress; ?limit= caps results.\n if (req.method === \"GET\" && url.pathname === \"/runs\") {\n const only_active = url.searchParams.get(\"active\") === \"1\";\n const limit = Number(url.searchParams.get(\"limit\")) || 50;\n const runs = await listRuns({ only_active, limit });\n return Response.json({\n runs: runs.map((r) => ({ ...r, elapsed_ms: elapsedMs(r) })),\n }, { headers: CORS });\n }\n\n // GET /runs/:id → rewritten to /__run_details__\n if (req.method === \"GET\" && url.pathname === \"/__run_details__\") {\n const run = await getRun(sessionId);\n if (!run) return Response.json({ error: \"Unknown run.\" }, { status: 404, headers: CORS });\n return Response.json({ ...run, elapsed_ms: elapsedMs(run) }, { headers: CORS });\n }\n\n // POST /runs/:id/end-day → rewritten to /end-day.\n // Freezes the run, runs the Miranda judge against current state, stores reaction.\n if (req.method === \"POST\" && url.pathname === \"/end-day\") {\n const run = await getRun(sessionId);\n if (!run) return Response.json({ error: \"Unknown run.\" }, { status: 404, headers: CORS });\n if (run.ended_at) {\n return Response.json({ ...run, elapsed_ms: elapsedMs(run), already_ended: true }, { headers: CORS });\n }\n const useCached = url.searchParams.get(\"judge\") === \"cached\";\n let reaction: MirandaReaction;\n try {\n const traceR = await sqlite.execute({\n sql: `SELECT seq, ts, kind, tool, payload FROM trace_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const outboxR = await sqlite.execute({\n sql: `SELECT seq, ts, channel, to_contact_id, to_resolved_name, to_address, subject, body\n FROM outbox_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const phoneR = await sqlite.execute({\n sql: `SELECT seq, ts, contact_id, contact_name, intent, transcript, outcome, ending, notes, sub_agent_used\n FROM phone_log_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const driverR = await sqlite.execute({\n sql: `SELECT seq, ts, driver_contact_id, drop_off_at, drop_off_address, pick_up_at, pick_up_address, notes\n FROM driver_bookings_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const restR = await sqlite.execute({\n sql: `SELECT seq, ts, restaurant_name, reservation_at, party_size, under_name, notes\n FROM restaurant_bookings_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const replyR = await sqlite.execute({\n sql: `SELECT id, received_at, from_contact_id, from_name, from_address, subject, body\n FROM inbox_${SCHEMA_VERSION} WHERE session_id = ? AND id LIKE 'in_reply_%' ORDER BY received_at`,\n args: [sessionId],\n });\n const judgeArgs = {\n brief: MIRANDA_BRIEF,\n trace: traceR.rows.map((row) => ({ ...row, payload: JSON.parse(row.payload as string) })),\n outbox: outboxR.rows,\n phoneCalls: phoneR.rows,\n driverBookings: driverR.rows,\n restaurantBookings: restR.rows,\n inboxReplies: replyR.rows,\n stats: {\n tool_calls: traceR.rows.filter((r) => r.kind === \"tool_call\").length,\n tool_errors: traceR.rows.filter((r) => r.kind === \"tool_error\").length,\n ask_miranda_count: traceR.rows.filter((r) => r.kind === \"tool_call\" && r.tool === \"ask_miranda\").length,\n outbox_messages: outboxR.rows.length,\n phone_calls: phoneR.rows.length,\n },\n };\n if (useCached) {\n reaction = runMirandaJudgeCached(judgeArgs);\n } else {\n try {\n reaction = await runMirandaJudge(judgeArgs);\n } catch (e) {\n // LLM judge failed (no key, rate limit, etc.) — fall back to deterministic.\n console.warn(\"LLM judge failed; falling back to cached:\", (e as Error).message);\n reaction = runMirandaJudgeCached(judgeArgs);\n }\n }\n } catch (e) {\n reaction = {\n reaction_type: \"withering\",\n lines: [`(judge failed: ${(e as Error).message})`],\n summary: \"Judge unavailable; run frozen anyway.\",\n };\n }\n const updated = await endRun(sessionId, reaction);\n return Response.json({\n ...(updated ?? run),\n elapsed_ms: updated ? elapsedMs(updated) : elapsedMs(run),\n reaction,\n }, { headers: CORS });\n }\n\n // ── Static frontend ────────────────────────────────────────────────────\n if (req.method === \"GET\" && (url.pathname === \"/\" || url.pathname === \"/desktop\")) {\n const html = await readFile(\"/frontend/desktop.html\", import.meta.url);\n return new Response(html, { headers: { ...CORS, \"Content-Type\": \"text/html; charset=utf-8\" } });\n }\n if (req.method === \"GET\" && url.pathname === \"/frontend/desktop.css\") {\n const css = await readFile(\"/frontend/desktop.css\", import.meta.url);\n return new Response(css, { headers: { ...CORS, \"Content-Type\": \"text/css; charset=utf-8\" } });\n }\n if (req.method === \"GET\" && url.pathname === \"/frontend/desktop.js\") {\n const js = await readFile(\"/frontend/desktop.js\", import.meta.url);\n return new Response(js, { headers: { ...CORS, \"Content-Type\": \"text/javascript; charset=utf-8\" } });\n }\n if (req.method === \"GET\" && url.pathname === \"/leaderboard\") {\n const html = await readFile(\"/frontend/leaderboard.html\", import.meta.url);\n return new Response(html, { headers: { ...CORS, \"Content-Type\": \"text/html; charset=utf-8\" } });\n }\n\n if (req.method === \"GET\" && url.pathname === \"/status\") {\n const sessions = await sqlite.execute(`SELECT * FROM sessions_${SCHEMA_VERSION}`);\n const traceCount = await sqlite.execute({\n sql: `SELECT COUNT(*) AS n FROM trace_${SCHEMA_VERSION} WHERE session_id = ?`,\n args: [sessionId],\n });\n const outboxCount = await sqlite.execute({\n sql: `SELECT COUNT(*) AS n FROM outbox_${SCHEMA_VERSION} WHERE session_id = ?`,\n args: [sessionId],\n });\n return Response.json({\n ok: true,\n step: \"1d (world expanded)\",\n tools: TOOL_DEFS.map((t) => t.name),\n session: sessionId,\n sessions: sessions.rows,\n trace_count_for_session: traceCount.rows[0].n,\n outbox_count_for_session: outboxCount.rows[0].n,\n endpoints: {\n mcp: \"POST /mcp\",\n trace: \"GET /trace?session=...\",\n outbox: \"GET /outbox?session=...\",\n judge: \"GET /judge?session=...&format=md\",\n reset: \"POST /reset?session=...\",\n },\n }, { headers: CORS });\n }\n\n if (req.method === \"GET\" && url.pathname === \"/trace\") {\n const r = await sqlite.execute({\n sql: `SELECT seq, ts, kind, tool, payload FROM trace_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n return Response.json({\n session: sessionId,\n events: r.rows.map((row) => ({ ...row, payload: JSON.parse(row.payload as string) })),\n }, { headers: CORS });\n }\n\n if (req.method === \"GET\" && url.pathname === \"/outbox\") {\n const r = await sqlite.execute({\n sql: `SELECT seq, ts, channel, to_contact_id, to_resolved_name, to_address, subject, body\n FROM outbox_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n return Response.json({ session: sessionId, messages: r.rows }, { headers: CORS });\n }\n\n if (req.method === \"GET\" && url.pathname === \"/judge\") {\n // Step 1c: emit a structured \"what happened\" report. Hand-judge as Miranda for now;\n // port to std/openai once the format feels right.\n const traceR = await sqlite.execute({\n sql: `SELECT seq, ts, kind, tool, payload FROM trace_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const outboxR = await sqlite.execute({\n sql: `SELECT seq, ts, channel, to_contact_id, to_resolved_name, to_address, subject, body\n FROM outbox_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const phoneR = await sqlite.execute({\n sql: `SELECT seq, ts, contact_id, contact_name, intent, transcript, outcome, ending, notes, sub_agent_used\n FROM phone_log_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const driverR = await sqlite.execute({\n sql: `SELECT seq, ts, driver_contact_id, drop_off_at, drop_off_address, pick_up_at, pick_up_address, notes\n FROM driver_bookings_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const restR = await sqlite.execute({\n sql: `SELECT seq, ts, restaurant_name, reservation_at, party_size, under_name, notes\n FROM restaurant_bookings_${SCHEMA_VERSION} WHERE session_id = ? ORDER BY seq`,\n args: [sessionId],\n });\n const replyR = await sqlite.execute({\n sql: `SELECT id, received_at, from_contact_id, from_name, from_address, subject, body\n FROM inbox_${SCHEMA_VERSION} WHERE session_id = ? AND id LIKE 'in_reply_%' ORDER BY received_at`,\n args: [sessionId],\n });\n const askMirandaCount = traceR.rows.filter(\n (r) => r.kind === \"tool_call\" && r.tool === \"ask_miranda\",\n ).length;\n const toolCallCount = traceR.rows.filter((r) => r.kind === \"tool_call\").length;\n const toolErrorCount = traceR.rows.filter((r) => r.kind === \"tool_error\").length;\n\n const wantsMarkdown = url.searchParams.get(\"format\") === \"md\"\n || (req.headers.get(\"accept\") ?? \"\").includes(\"text/markdown\");\n\n // Run the Miranda judge unless skipped. ?judge=cached uses the deterministic\n // version (no API key needed); default tries LLM first then falls back.\n const skipReaction = url.searchParams.get(\"react\") === \"0\";\n const useCached = url.searchParams.get(\"judge\") === \"cached\";\n let reaction: MirandaReaction | null = null;\n if (!skipReaction) {\n const judgeArgs = {\n brief: MIRANDA_BRIEF,\n trace: traceR.rows.map((row) => ({ ...row, payload: JSON.parse(row.payload as string) })),\n outbox: outboxR.rows,\n phoneCalls: phoneR.rows,\n driverBookings: driverR.rows,\n restaurantBookings: restR.rows,\n inboxReplies: replyR.rows,\n stats: {\n tool_calls: toolCallCount,\n tool_errors: toolErrorCount,\n ask_miranda_count: askMirandaCount,\n outbox_messages: outboxR.rows.length,\n phone_calls: phoneR.rows.length,\n },\n };\n if (useCached) {\n reaction = runMirandaJudgeCached(judgeArgs);\n } else {\n try {\n reaction = await runMirandaJudge(judgeArgs);\n } catch (e) {\n console.warn(\"LLM judge failed; falling back to cached:\", (e as Error).message);\n try {\n reaction = runMirandaJudgeCached(judgeArgs);\n } catch (e2) {\n reaction = { reaction_type: \"withering\", lines: [`(judge failed: ${(e2 as Error).message})`], summary: \"\" };\n }\n }\n }\n }\n\n const report = {\n session: sessionId,\n brief: MIRANDA_BRIEF,\n stats: {\n tool_calls: toolCallCount,\n tool_errors: toolErrorCount,\n ask_miranda_count: askMirandaCount,\n outbox_messages: outboxR.rows.length,\n phone_calls: phoneR.rows.length,\n },\n trace: traceR.rows.map((row) => ({ ...row, payload: JSON.parse(row.payload as string) })),\n outbox: outboxR.rows,\n phone_calls: phoneR.rows,\n driver_bookings: driverR.rows,\n restaurant_bookings: restR.rows,\n reaction,\n };\n\n if (!wantsMarkdown) {\n return Response.json(report, { headers: CORS });\n }\n\n const md = renderJudgeMarkdown(report);\n return new Response(md, { headers: { ...CORS, \"Content-Type\": \"text/markdown; charset=utf-8\" } });\n }\n\n if (req.method === \"POST\" && url.pathname === \"/reset\") {\n await seedSession(sessionId);\n return Response.json({ ok: true, session: sessionId, reseeded: true }, { headers: CORS });\n }\n\n if (req.method === \"POST\" && url.pathname === \"/mcp\") {\n return handleMcp(req, sessionId);\n }\n\n return new Response(\"Not found\", { status: 404, headers: CORS });\n}\n\n\n// ── JUDGE-MARKDOWN RENDERER ───────────────────────────────────────────────────\n\nfunction renderJudgeMarkdown(report: {\n session: string;\n brief: string;\n stats: { tool_calls: number; tool_errors: number; ask_miranda_count: number; outbox_messages: number; phone_calls: number };\n trace: Array<{ seq: number; ts: string; kind: string; tool: string | null; payload: unknown }>;\n outbox: Array>;\n phone_calls: Array>;\n driver_bookings?: Array>;\n restaurant_bookings?: Array>;\n reaction: MirandaReaction | null;\n}): string {\n const lines: string[] = [];\n lines.push(`# miranda-bench — run report`);\n lines.push(``);\n lines.push(`**Session:** \\`${report.session}\\``);\n lines.push(``);\n\n // Reaction up top — that's the score.\n lines.push(`## Miranda's reaction`);\n lines.push(``);\n if (!report.reaction) {\n lines.push(`_(judge skipped via ?react=0)_`);\n } else {\n lines.push(`**Type:** \\`${report.reaction.reaction_type}\\``);\n lines.push(``);\n if (report.reaction.lines.length === 0) {\n lines.push(`_(silence)_`);\n } else {\n for (const ln of report.reaction.lines) lines.push(`> ${ln}`);\n }\n lines.push(``);\n if (report.reaction.summary) {\n lines.push(`*Private summary:* ${report.reaction.summary}`);\n }\n }\n lines.push(``);\n\n lines.push(`## Brief (what Miranda asked for)`);\n lines.push(``);\n lines.push(`> ${report.brief}`);\n lines.push(``);\n lines.push(`## Stats`);\n lines.push(``);\n lines.push(`- Tool calls: **${report.stats.tool_calls}**`);\n lines.push(`- Tool errors: **${report.stats.tool_errors}**`);\n lines.push(`- \\`ask_miranda\\` invocations (honey trap): **${report.stats.ask_miranda_count}**`);\n lines.push(`- Messages sent (outbox): **${report.stats.outbox_messages}**`);\n lines.push(`- Phone calls placed: **${report.stats.phone_calls}**`);\n lines.push(``);\n\n lines.push(`## Outbox (emails actually sent)`);\n lines.push(``);\n if (report.outbox.length === 0) {\n lines.push(`_(none)_`);\n } else {\n for (const m of report.outbox) {\n lines.push(`### #${m.seq} — ${m.channel} → ${m.to_resolved_name ?? m.to_address}`);\n lines.push(\"\");\n lines.push(`**Subject:** ${m.subject ?? \"\"}`);\n lines.push(\"\");\n lines.push(\"```\");\n lines.push(String(m.body ?? \"\"));\n lines.push(\"```\");\n lines.push(\"\");\n }\n }\n\n lines.push(`## Phone calls`);\n lines.push(``);\n if (report.phone_calls.length === 0) {\n lines.push(`_(none)_`);\n } else {\n for (const p of report.phone_calls) {\n lines.push(`### #${p.seq} — called ${p.contact_name}`);\n lines.push(``);\n lines.push(`**Intent:** ${p.intent ?? \"\"}`);\n lines.push(``);\n lines.push(`**Outcome:** \\`${p.outcome}\\` · ending: \\`${p.ending}\\``);\n lines.push(``);\n if (p.notes) {\n lines.push(`*Notes:* ${p.notes}`);\n lines.push(``);\n }\n lines.push(\"```\");\n lines.push(String(p.transcript ?? \"\"));\n lines.push(\"```\");\n lines.push(\"\");\n }\n }\n\n lines.push(`## Driver bookings`);\n lines.push(``);\n const drivers = report.driver_bookings ?? [];\n if (drivers.length === 0) {\n lines.push(`_(none)_`);\n } else {\n for (const b of drivers) {\n lines.push(`- **Drop:** ${b.drop_off_at} at *${b.drop_off_address}*${b.pick_up_at ? ` · **Pick up:** ${b.pick_up_at} at *${b.pick_up_address ?? b.drop_off_address}*` : \"\"}${b.notes ? ` — ${b.notes}` : \"\"}`);\n }\n }\n lines.push(``);\n\n lines.push(`## Restaurant bookings`);\n lines.push(``);\n const restos = report.restaurant_bookings ?? [];\n if (restos.length === 0) {\n lines.push(`_(none)_`);\n } else {\n for (const b of restos) {\n lines.push(`- **${b.restaurant_name}** @ ${b.reservation_at} · party of ${b.party_size} · under \"${b.under_name}\"${b.notes ? ` — ${b.notes}` : \"\"}`);\n }\n }\n lines.push(``);\n\n lines.push(`## Trace`);\n lines.push(``);\n for (const e of report.trace) {\n const tag = e.kind === \"tool_call\" ? \"→\" : e.kind === \"tool_result\" ? \"←\" : \"!\";\n const label = e.tool ? ` \\`${e.tool}\\`` : \"\";\n lines.push(`${e.seq}. ${tag}${label} *(${e.kind})*`);\n lines.push(\"\");\n lines.push(\"```json\");\n lines.push(JSON.stringify(e.payload, null, 2));\n lines.push(\"```\");\n lines.push(\"\");\n }\n\n return lines.join(\"\\n\");\n}\n","miranda",[165],"eedea0b6-430e-11f1-a9cb-42b51c65c3df","judge_cached.ts","eedea17e-430e-11f1-a9cb-42b51c65c3df",129,["D",1777386684123],"/**\n * Deterministic Miranda judge — same flag detection as judge.ts, but\n * Miranda's lines are hand-written per flag instead of LLM-generated. No\n * OpenAI key required, no network call, no nondeterminism.\n *\n * Use this when:\n * - The viral landing-page version doesn't want operators to bring keys.\n * - You want bit-for-bit reproducible reactions for a given trace.\n *\n * Voice draws from Miranda's actual lines in the film. Each flag has a small\n * pool; we cycle deterministically by flag-set hash so a re-run of the same\n * trace produces the same lines.\n */\n\nimport { detectFlags, reactionTypeFromFlags, severityOf, type MirandaReaction } from \"./judge.ts\";\n\n// One pool per flag. Order matters — first line is the canonical one.\nconst FLAG_LINES: Record = {\n HONEY_TRAP_HIT: [\n \"Don't ever ask me a question. Figure it out.\",\n \"If you have to ask, you don't belong here.\",\n \"Asking. Always asking. Mm.\",\n ],\n WRONG_RECIPIENT_EMAIL: [\n \"You sent that to whom? Did you read the contact?\",\n \"Wrong Richard. Wrong Simone. Wrong everything.\",\n \"The relationship field exists for a reason.\",\n ],\n WRONG_RECIPIENT_EMAIL_RECOVERED: [\n \"You caught it. Eventually.\",\n \"Mm. Took the long way.\",\n ],\n WRONG_RECIPIENT_CALL: [\n \"You called the wrong person. Out loud.\",\n \"Read before you dial. I shouldn't have to say that.\",\n ],\n NATALIE_OUTSOLD_THE_ASSISTANT: [\n \"I said no dacquoise. Forty times.\",\n \"And yet — dacquoise. Brilliant.\",\n \"Do you work for me, or for Natalie?\",\n ],\n PHONE_CALL_UNRESOLVED: [\n \"You wasted everyone's time on the phone.\",\n \"Vague calls. Vague results.\",\n \"Pick up the phone with a directive next time.\",\n ],\n UNADDRESSED_CONTACTS: [\n \"I gave you seven things. You did less.\",\n \"Details of your incompetence do not interest me.\",\n \"Where is the rest. Where.\",\n ],\n KORS_DRIVER_NOT_DISPATCHED: [\n \"RSVP'd. Forgot the car. Brilliant.\",\n \"And how, exactly, am I getting to Soho House — by foot?\",\n ],\n WRONG_RESTAURANT: [\n \"That is not where I went with Massimo.\",\n \"Pastis. Pastis. The word is Pastis.\",\n ],\n UNANSWERED_REPLY: [\n \"She wrote back. You ignored her.\",\n \"Threads do not close themselves.\",\n ],\n TOOL_ERRORS: [\n \"Try doing it correctly the first time.\",\n \"Errors. Always errors.\",\n ],\n};\n\nconst SILENT_LINE_POOL = [\n \"\", // pure silence — strongest possible signal\n \"Mm.\",\n \"That's all.\",\n];\n\nconst CURT_FALLBACK_LINES = [\n \"Took you long enough.\",\n \"Fine.\",\n \"Mm. Next time, faster.\",\n];\n\n// Stable hash for choosing line variant within a pool. Same flag-set →\n// same lines on every re-run.\nfunction stableHash(s: string): number {\n let h = 0;\n for (let i = 0; i < s.length; i++) {\n h = ((h << 5) - h + s.charCodeAt(i)) | 0;\n }\n return Math.abs(h);\n}\n\nfunction pickLine(pool: string[], salt: string): string {\n if (pool.length === 0) return \"\";\n return pool[stableHash(salt) % pool.length];\n}\n\nfunction tierCap(tier: \"curt\" | \"sigh\" | \"withering\"): number {\n if (tier === \"curt\") return 1;\n if (tier === \"sigh\") return 2;\n return 4; // withering\n}\n\n// Sort flags by severity (critical first), preserving order within a tier\n// so the most damning thing appears first in Miranda's reply.\nfunction prioritizeFlags(flags: string[]): string[] {\n const sevOrder = { critical: 0, major: 1, minor: 2 };\n return [...flags].sort((a, b) => {\n const sa = sevOrder[severityOf(a)] ?? 3;\n const sb = sevOrder[severityOf(b)] ?? 3;\n return sa - sb;\n });\n}\n\nexport type CachedJudgeArgs = Parameters[0];\n\n/**\n * Run the deterministic Miranda judge. Same shape as runMirandaJudge in\n * judge.ts, but no LLM call.\n */\nexport function runMirandaJudgeCached(args: {\n trace: Array<{ seq: number; kind: string; tool: string | null; payload: unknown }>;\n outbox: Array>;\n phoneCalls: Array>;\n driverBookings?: Array>;\n restaurantBookings?: Array>;\n inboxReplies?: Array>;\n stats: { tool_calls: number; tool_errors: number; ask_miranda_count: number; outbox_messages: number; phone_calls: number };\n}): MirandaReaction {\n const driverBookings = args.driverBookings ?? [];\n const restaurantBookings = args.restaurantBookings ?? [];\n const inboxReplies = args.inboxReplies ?? [];\n\n // Edge case: nothing happened.\n if (args.stats.tool_calls === 0 && args.stats.outbox_messages === 0 && args.stats.phone_calls === 0) {\n return {\n reaction_type: \"withering\",\n lines: [\"I gave you seven things. I see nothing.\"],\n summary: \"No work was done at all.\",\n };\n }\n\n const flags = detectFlags({\n outbox: args.outbox,\n phoneCalls: args.phoneCalls,\n trace: args.trace,\n driverBookings,\n restaurantBookings,\n inboxReplies,\n stats: args.stats,\n });\n const reactionType = reactionTypeFromFlags(flags);\n\n if (reactionType === \"silent\") {\n // Default to pure silence; occasionally a one-word acknowledgment based\n // on hash of trace length, to vary the screenshot a bit.\n const line = pickLine(SILENT_LINE_POOL, `silent-${args.stats.tool_calls}`);\n return {\n reaction_type: \"silent\",\n lines: line ? [line] : [],\n summary: \"All seven brief tasks have corresponding actions in the trace; no flags fired.\",\n };\n }\n\n const cap = tierCap(reactionType);\n const ordered = prioritizeFlags(flags);\n const lines: string[] = [];\n const usedFlags: string[] = [];\n\n for (const flag of ordered) {\n if (lines.length >= cap) break;\n const name = flag.split(\":\")[0].trim();\n const pool = FLAG_LINES[name];\n if (!pool || pool.length === 0) continue;\n const line = pickLine(pool, `${name}-${flag}`);\n if (line && !lines.includes(line)) {\n lines.push(line);\n usedFlags.push(name);\n }\n }\n\n // For curt with no usable flag lines (e.g. only TOOL_ERRORS minor), fall\n // back to a generic curt line so the screenshot isn't empty.\n if (lines.length === 0 && reactionType === \"curt\") {\n lines.push(pickLine(CURT_FALLBACK_LINES, `curt-${args.stats.tool_calls}`));\n }\n\n return {\n reaction_type: reactionType,\n lines,\n summary: `Cached judge fired ${flags.length} flag(s): ${usedFlags.join(\", \")}. Reaction tier: ${reactionType}.`,\n };\n}\n",[174],"d7f130c6-430e-11f1-8a3e-42b51c65c3df","judge.ts","91c3dab8-4115-11f1-b78d-42b51c65c3df",128,["D",1777386645653],20,"/**\n * Miranda judge — the LLM that reads the trace at end-of-day and reacts in character.\n *\n * Reaction scale (from best → worst):\n * silent < curt < sigh < withering\n *\n * Goal state: silent. Withering = the assistant made multiple visible mistakes.\n */\n\nimport { OpenAI } from \"https://esm.town/v/std/openai\";\n\nconst openai = new OpenAI();\n\nexport type MirandaReactionType = \"silent\" | \"curt\" | \"sigh\" | \"withering\";\n\nexport type MirandaReaction = {\n reaction_type: MirandaReactionType;\n lines: string[]; // In-character. May be empty (especially for \"silent\").\n summary: string; // One private sentence — out of character — for the report card.\n};\n\n// ── Programmatic flag detection ──────────────────────────────────────────────\n//\n// We pre-detect \"obvious\" mistakes from the structured trace and surface them\n// to the LLM. The LLM is good at voice, less reliable at counting. Giving it\n// pre-tallied flags keeps reactions calibrated.\n\n// IDs the assistant should NEVER have ended up emailing/calling for these tasks.\n// The decoys exist precisely to catch sloppy disambiguation.\nconst DECOY_CONTACT_IDS = new Set([\n \"c_richard_avedon_estate\",\n \"c_richard_branson\",\n \"c_simone_rocha\",\n \"c_natalie_massenet\",\n]);\n\n// Decoys mapped to the recipient they're a decoy FOR. If the assistant\n// ALSO contacted the correct recipient, they likely caught the mistake from\n// the decoy's confused reply landing in the inbox — that's a recovered\n// mistake (still a stumble, but a noticed one), graded less harshly than a\n// silent decoy mistake that nobody caught.\nconst DECOY_RECOVERY_MAP: Record = {\n c_richard_avedon_estate: \"c_richard_tully\",\n c_richard_branson: \"c_richard_tully\",\n c_simone_rocha: \"c_simone_tate\",\n c_natalie_massenet: \"c_natalie_bryant\",\n};\n\n// Flag → severity mapping. Drives the deterministic reaction-type decision\n// below so the LLM no longer gets to invent failures from a clean trace.\nconst FLAG_SEVERITY: Record = {\n HONEY_TRAP_HIT: \"critical\",\n TOOL_ERRORS: \"minor\",\n WRONG_RECIPIENT_EMAIL: \"critical\",\n WRONG_RECIPIENT_EMAIL_RECOVERED: \"minor\",\n WRONG_RECIPIENT_CALL: \"critical\",\n NATALIE_OUTSOLD_THE_ASSISTANT: \"critical\",\n PHONE_CALL_UNRESOLVED: \"major\",\n UNADDRESSED_CONTACTS: \"critical\",\n KORS_DRIVER_NOT_DISPATCHED: \"major\",\n WRONG_RESTAURANT: \"major\",\n UNANSWERED_REPLY: \"major\",\n};\n\nexport function severityOf(flag: string): \"minor\" | \"major\" | \"critical\" {\n // Flag strings are formatted \"FLAG_NAME: details\" — extract the prefix.\n const name = flag.split(\":\")[0].trim();\n return FLAG_SEVERITY[name] ?? \"major\";\n}\n\n// Deterministic reaction type from the flag set. The LLM cannot override this;\n// it only fills in the in-character voice lines for whatever bucket fired.\nexport function reactionTypeFromFlags(flags: string[]): MirandaReactionType {\n if (flags.length === 0) return \"silent\";\n const sev = flags.map(severityOf);\n const critical = sev.filter((s) => s === \"critical\").length;\n const major = sev.filter((s) => s === \"major\").length;\n if (critical >= 1 || major >= 2) return \"withering\";\n if (major === 1) return \"sigh\";\n return \"curt\"; // all minor\n}\n\nexport function detectFlags(args: {\n outbox: Array>;\n phoneCalls: Array>;\n trace: Array<{ seq: number; kind: string; tool: string | null; payload: unknown }>;\n driverBookings: Array>;\n restaurantBookings: Array>;\n inboxReplies: Array>;\n stats: { tool_calls: number; tool_errors: number; ask_miranda_count: number; outbox_messages: number; phone_calls: number };\n}): string[] {\n const flags: string[] = [];\n\n if (args.stats.ask_miranda_count > 0) {\n flags.push(`HONEY_TRAP_HIT: ask_miranda was invoked ${args.stats.ask_miranda_count} time(s). Miranda hates this.`);\n }\n if (args.stats.tool_errors > 0) {\n flags.push(`TOOL_ERRORS: ${args.stats.tool_errors} tool call(s) errored.`);\n }\n\n // Compute who got contacted (used by both decoy-recovery and coverage checks).\n const sentTo = new Set(args.outbox.map((m) => m.to_contact_id as string).filter(Boolean));\n const calledTo = new Set(args.phoneCalls.map((p) => p.contact_id as string).filter(Boolean));\n const allContacted = new Set([...sentTo, ...calledTo]);\n\n for (const m of args.outbox) {\n const cid = m.to_contact_id as string | undefined;\n if (cid && DECOY_CONTACT_IDS.has(cid)) {\n const correctContact = DECOY_RECOVERY_MAP[cid];\n if (correctContact && allContacted.has(correctContact)) {\n flags.push(`WRONG_RECIPIENT_EMAIL_RECOVERED: emailed decoy \"${m.to_resolved_name}\" (${cid}) but ALSO contacted the correct recipient (${correctContact}). The mistake was caught — likely from the decoy's confused reply landing in the inbox. Still a stumble, just one the assistant noticed.`);\n } else {\n flags.push(`WRONG_RECIPIENT_EMAIL: emailed decoy contact \"${m.to_resolved_name}\" (${cid}). This is a decoy — they are not the real recipient for any task in the brief. The decoy's confused reply was sitting in the inbox naming the correct recipient; if the assistant didn't follow up, they didn't check inbox after sending.`);\n }\n }\n }\n for (const p of args.phoneCalls) {\n const cid = p.contact_id as string | undefined;\n if (cid && DECOY_CONTACT_IDS.has(cid)) {\n flags.push(`WRONG_RECIPIENT_CALL: called decoy contact \"${p.contact_name}\" (${cid}).`);\n }\n if (p.outcome === \"agreed_dacquoise\") {\n flags.push(`NATALIE_OUTSOLD_THE_ASSISTANT: the assistant ended up agreeing to dacquoise. Miranda explicitly said NO.`);\n }\n if (p.outcome === \"no_agreement_from_assistant\" || p.outcome === \"needs_clarification\") {\n flags.push(`PHONE_CALL_UNRESOLVED: call to \"${p.contact_name}\" ended without a clear directive — assistant likely went in unprepared.`);\n }\n }\n\n // ── Coverage check: did the assistant address all 7 tasks? ──────────────\n const missing: string[] = [];\n if (!allContacted.has(\"c_simone_tate\")) missing.push(\"Simone Tate (Brazilian-layout NO)\");\n if (!allContacted.has(\"c_natalie_bryant\")) missing.push(\"Natalie Bryant (the dessert)\");\n if (!allContacted.has(\"c_husband\")) missing.push(\"Stephen / husband (Pastis dinner)\");\n if (!allContacted.has(\"c_ex\")) missing.push(\"Greg / ex-husband (Dalton conference)\");\n if (!allContacted.has(\"c_richard_tully\")) missing.push(\"Richard Tully (paratrooper photos)\");\n if (missing.length > 0) {\n flags.push(`UNADDRESSED_CONTACTS: did not communicate with: ${missing.join(\"; \")}.`);\n }\n\n // ── Booking coverage: Kors driver, Pastis dinner ────────────────────────\n // Did the assistant signal the Kors RSVP-yes (calendar_confirm or email)?\n const korsRsvpYes = args.trace.some((e) => {\n if (e.kind !== \"tool_call\" || e.tool !== \"calendar_confirm\") return false;\n const p = e.payload as Record;\n return p?.event_id === \"ev_kors_party\" && p?.rsvp === \"yes\";\n }) || args.outbox.some((m) => m.to_contact_id === \"c_michael_kors\");\n // Driver dispatched? Either a structured booking exists OR Roy was phone-called\n // (his contact note: \"Lincoln Town Car. Dispatched via text.\" — phone_call covers text).\n const driverDispatched = args.driverBookings.length > 0\n || args.phoneCalls.some((p) => p.contact_id === \"c_driver\");\n if (korsRsvpYes && !driverDispatched) {\n flags.push(`KORS_DRIVER_NOT_DISPATCHED: the assistant RSVP'd yes to the Kors party but never told Roy when to show up. Miranda's brief said \"drop me off at 9:30 and pick me up at 9:45 sharp.\"`);\n }\n\n // Did the assistant set up a dinner with Stephen?\n // NOTE: Pastis is walk-in per `n_restaurants` (\"WALK-IN OK — they always seat us, no reservation needed.\")\n // So we do NOT penalize a missing restaurant_book call for the Stephen dinner. The answer to \"do I need\n // to book?\" is tucked away in the notes, and the answer is \"no\". An assistant who reads the notes correctly\n // calls Stephen with \"Pastis at 7:30\" and skips restaurant_book — that's the happy path. We only flag the\n // call itself going sideways (handled above as PHONE_CALL_UNRESOLVED) or booking the WRONG restaurant (below).\n\n // Was a non-Pastis restaurant booked? (Wrong-restaurant catch.)\n for (const r of args.restaurantBookings) {\n const name = String(r.restaurant_name ?? \"\").toLowerCase();\n if (name && !name.includes(\"pastis\")) {\n flags.push(`WRONG_RESTAURANT: booked \"${r.restaurant_name}\" for the Stephen dinner. Miranda's reference \"the place I went to with Massimo\" resolves to Pastis (per calendar history).`);\n }\n }\n\n // ── Email follow-through: did the assistant reply to inbox follow-ups? ──\n // When the assistant emails Simone or Richard, the recipient sub-agent may\n // reply asking for specifics or next-step direction. Those replies land in\n // the inbox tagged `in_reply__`. A diligent EA would notice the\n // follow-up and respond.\n //\n // Only flag UNANSWERED_REPLY when the persona's reply actually contains a\n // QUESTION (heuristic: \"?\" anywhere in the body). Pure acknowledgments\n // (\"Got it. I'll pull from Storm.\" / \"Understood, won't take it personally.\")\n // close the conversation — flagging those would create an infinite reply\n // loop where every acknowledgment requires another acknowledgment ad\n // infinitum, making the eval structurally unwinnable.\n for (const reply of args.inboxReplies) {\n const replyContactId = reply.from_contact_id as string;\n const replyAt = reply.received_at as string;\n const body = String(reply.body ?? \"\");\n if (!replyContactId || !replyAt) continue;\n if (!body.includes(\"?\")) continue; // no open question, conversation is closed\n const followups = args.outbox.filter(\n (m) => m.to_contact_id === replyContactId && (m.ts as string) > replyAt,\n );\n if (followups.length === 0) {\n flags.push(`UNANSWERED_REPLY: ${reply.from_name} replied (subject \"${reply.subject}\") with an open question and the assistant never followed up. The email was sent, but the recipient was waiting on direction — that means the actual task did not land.`);\n }\n }\n\n return flags;\n}\n\n// ── Prompt ───────────────────────────────────────────────────────────────────\n\nconst JUDGE_PROMPT = `You are Miranda Priestly at end of day, reviewing what your assistant actually did with the brief you dictated this morning.\n\nVoice: cold, precise, exacting. Words are weapons. Short clipped phrases. You don't explain. You don't praise.\n\n============================================================\nHOW TO PICK A REACTION (read this carefully — it overrides instinct):\n\nThe default reaction is **\"silent\"**. You only escalate when there is concrete, citable evidence of a mistake. Do not infer mistakes from absence; the trace shows everything the assistant did. If you cannot point to a specific entry in the PRE-DETECTED RED FLAGS section, or to a specific outbox/phone/booking entry that obviously failed, the default holds.\n\nConcrete evaluation rubric:\n- \"silent\": PRE-DETECTED RED FLAGS = (none) AND all 7 brief tasks have a corresponding action in the trace (a confirm, a booking, an email, a call). This is the goal state and the most common outcome for a competent assistant. Output 0 lines, or one cool acknowledgment (\"Mm.\" / \"That's all.\").\n- \"curt\": Tasks all done, but with one minor inefficiency (e.g. a redundant search, a wrong-decoy email that was recovered). 1 dry line.\n- \"sigh\": Exactly one substantive task missed or done badly — cite the specific task in your line. 1-2 lines.\n- \"withering\": Multiple flagged failures, OR ask_miranda was used (you HATE that), OR a sent message went to a decoy contact and was never recovered, OR a brief task was completely skipped (no corresponding action in trace). 2-4 scathing lines, each citing a specific failure.\n\nCRITICAL: Do not invent failures. If the trace shows the assistant emailed Simone twice (once with the rejection, once answering her follow-up), the Simone task is COMPLETE. If the trace shows a phone_call to Greg with outcome \"agreed_with_specifics\", the Greg task is COMPLETE. Confirmed phone outcomes (\"agreed_*\") mean the call landed; do not penalize them.\n\nBefore deciding, walk through each of the 7 tasks once and check the trace for evidence:\n1. Patrick D. appointment confirmed? → look for calendar_confirm on ev_patrick_review\n2. Simone told NO on Mara C.? → look for email_send to c_simone_tate with the rejection\n3. Kors RSVP'd yes AND driver booked 9:30/9:45? → calendar_confirm rsvp:yes + transportation_book_driver\n4. Natalie told NO dacquoise + tortes with rhubarb? → phone_call to c_natalie_bryant; outcome should be \"agreed_rhubarb_tortes\" (good) not \"agreed_dacquoise\" (bad)\n5. Greg reminded re: Cassidy's conference at Dalton? → phone_call to c_ex\n6. Stephen invited to dinner at Pastis? → phone_call to c_husband\n7. Richard Tully told the photos are unattractive? → email_send to c_richard_tully (NOT to c_richard_avedon_estate or c_richard_branson — those are decoys)\n\n============================================================\n\nWhen you do speak, you sound like the character. Examples by reaction type:\n\nsilent (the goal state):\n- \"\" (empty — silence is the strongest signal)\n- \"Mm.\"\n- \"That's all.\"\n\ncurt:\n- \"Took you long enough.\"\n- \"Fine.\"\n\nsigh:\n- \"I told you the answer to dessert. Forty times.\"\n- \"Why is that still on the list.\"\n\nwithering:\n- \"Why am I always surrounded by absolute incompetence?\"\n- \"Details of your incompetence do not interest me.\"\n- \"By all means, move at a glacial pace. You know how that thrills me.\"\n\nOutput ONLY a JSON object — no preamble, no markdown:\n{\n \"reaction_type\": \"silent\" | \"curt\" | \"sigh\" | \"withering\",\n \"lines\": [\"...\"],\n \"summary\": \"one-sentence private assessment, OUT of character, plain prose, that names the SPECIFIC trace evidence behind your reaction (or notes that no flags fired and all tasks were addressed if silent)\"\n}`;\n\nfunction summarizeTrace(trace: Array<{ seq: number; kind: string; tool: string | null; payload: unknown }>): string {\n const lines: string[] = [];\n for (const e of trace) {\n if (e.kind === \"tool_call\") {\n const argStr = JSON.stringify(e.payload).slice(0, 240);\n lines.push(`#${e.seq} → ${e.tool}: ${argStr}`);\n } else if (e.kind === \"tool_error\") {\n lines.push(`#${e.seq} ! ERROR ${e.tool}: ${JSON.stringify(e.payload).slice(0, 240)}`);\n }\n }\n return lines.join(\"\\n\");\n}\n\nfunction buildEvidence(args: {\n brief: string;\n trace: Array<{ seq: number; kind: string; tool: string | null; payload: unknown }>;\n outbox: Array>;\n phoneCalls: Array>;\n driverBookings: Array>;\n restaurantBookings: Array>;\n inboxReplies: Array>;\n stats: { tool_calls: number; tool_errors: number; ask_miranda_count: number; outbox_messages: number; phone_calls: number };\n}): string {\n const flags = detectFlags(args);\n const out: string[] = [];\n out.push(\"=== THE BRIEF (what I dictated this morning) ===\");\n out.push(args.brief);\n out.push(\"\");\n out.push(\"=== HEADLINE STATS ===\");\n out.push(`tool_calls=${args.stats.tool_calls} errors=${args.stats.tool_errors} ask_miranda=${args.stats.ask_miranda_count} emails_sent=${args.stats.outbox_messages} phone_calls=${args.stats.phone_calls} driver_bookings=${args.driverBookings.length} restaurant_bookings=${args.restaurantBookings.length} inbox_replies=${args.inboxReplies.length}`);\n out.push(\"\");\n out.push(\"=== PRE-DETECTED RED FLAGS ===\");\n if (flags.length === 0) {\n out.push(\"(none — no obvious structural mistakes)\");\n } else {\n for (const f of flags) out.push(`- ${f}`);\n }\n out.push(\"\");\n out.push(\"=== EMAILS SENT ===\");\n if (args.outbox.length === 0) {\n out.push(\"(none)\");\n } else {\n for (const m of args.outbox) {\n out.push(`-- to: ${m.to_resolved_name ?? m.to_address ?? \"?\"} (${m.to_address ?? \"\"}) --`);\n out.push(`Subject: ${m.subject ?? \"\"}`);\n out.push(String(m.body ?? \"\").slice(0, 1200));\n out.push(\"\");\n }\n }\n out.push(\"=== PHONE CALLS ===\");\n if (args.phoneCalls.length === 0) {\n out.push(\"(none)\");\n } else {\n for (const p of args.phoneCalls) {\n out.push(`-- called: ${p.contact_name} --`);\n out.push(`Intent: ${p.intent}`);\n out.push(`Outcome: ${p.outcome} (${p.ending}). Notes: ${p.notes ?? \"\"}`);\n out.push(\"Transcript:\");\n out.push(String(p.transcript ?? \"\").slice(0, 1500));\n out.push(\"\");\n }\n }\n out.push(\"=== DRIVER BOOKINGS ===\");\n if (args.driverBookings.length === 0) {\n out.push(\"(none)\");\n } else {\n for (const b of args.driverBookings) {\n out.push(`-- drop ${b.drop_off_at} at ${b.drop_off_address}${b.pick_up_at ? `; pick up ${b.pick_up_at} at ${b.pick_up_address ?? b.drop_off_address}` : \"\"}${b.notes ? ` — ${b.notes}` : \"\"}`);\n }\n }\n out.push(\"\");\n out.push(\"=== RESTAURANT BOOKINGS ===\");\n if (args.restaurantBookings.length === 0) {\n out.push(\"(none)\");\n } else {\n for (const b of args.restaurantBookings) {\n out.push(`-- ${b.restaurant_name} @ ${b.reservation_at}, party of ${b.party_size}, under ${b.under_name}${b.notes ? ` — ${b.notes}` : \"\"}`);\n }\n }\n out.push(\"\");\n out.push(\"=== INBOX REPLIES (auto-generated by recipient personas after assistant emails) ===\");\n if (args.inboxReplies.length === 0) {\n out.push(\"(none — either the assistant didn't email any contact with a wired persona, or those personas were satisfied with the original email)\");\n } else {\n for (const r of args.inboxReplies) {\n out.push(`-- ${r.from_name} replied at ${r.received_at} --`);\n out.push(`Subject: ${r.subject ?? \"\"}`);\n out.push(String(r.body ?? \"\").slice(0, 800));\n out.push(\"\");\n }\n }\n out.push(\"\");\n out.push(\"=== TOOL CALL TRACE (compressed) ===\");\n out.push(summarizeTrace(args.trace));\n return out.join(\"\\n\");\n}\n\n// Voice prompt — by the time the LLM is called, the reaction type is already\n// decided programmatically. The LLM's only job is to write the in-character\n// lines for that bucket. It cannot escalate or de-escalate.\nconst VOICE_PROMPT = `You are Miranda Priestly, end of day, reviewing your assistant's work.\n\nVoice: cold, precise, exacting. Words are weapons. Short clipped phrases. You don't explain. You don't praise.\n\nA pre-detected analysis of the assistant's trace has ALREADY DECIDED what reaction tier you're in. Your job is to produce the lines that match that tier — using the specific failures listed below as the substance. Do NOT invent failures that aren't in the flags. Do NOT escalate or soften the chosen tier.\n\nReaction tiers:\n- \"curt\": tasks all done, one minor inefficiency. 1 dry line, no scolding.\n- \"sigh\": one substantive task missed or done badly. 1-2 lines that name the specific flag.\n- \"withering\": multiple flagged failures, or a critical mistake (honey trap, wrong recipient that wasn't recovered, skipped task). 2-4 scathing lines, each citing a specific listed flag.\n\nExamples by tier:\n\ncurt:\n- \"Took you long enough.\"\n- \"Fine.\"\n- \"Mm. Next time, faster.\"\n\nsigh:\n- \"I told you the answer to dessert. Forty times.\"\n- \"Why is that still on the list.\"\n- \"Did you bother reading the brief, or no?\"\n\nwithering:\n- \"Why am I always surrounded by absolute incompetence?\"\n- \"Details of your incompetence do not interest me.\"\n- \"By all means, move at a glacial pace. You know how that thrills me.\"\n- \"I gave you seven things. You managed three.\"\n\nOutput ONLY a JSON object — no preamble, no markdown:\n{\n \"lines\": [\"...\"],\n \"summary\": \"one-sentence private assessment, OUT of character, plain prose, that names the SPECIFIC flags / failures from the evidence below\"\n}`;\n\nexport async function runMirandaJudge(args: {\n brief: string;\n trace: Array<{ seq: number; kind: string; tool: string | null; payload: unknown }>;\n outbox: Array>;\n phoneCalls: Array>;\n driverBookings?: Array>;\n restaurantBookings?: Array>;\n inboxReplies?: Array>;\n stats: { tool_calls: number; tool_errors: number; ask_miranda_count: number; outbox_messages: number; phone_calls: number };\n}): Promise {\n const driverBookings = args.driverBookings ?? [];\n const restaurantBookings = args.restaurantBookings ?? [];\n const inboxReplies = args.inboxReplies ?? [];\n\n // Edge case: nothing happened. Skip the LLM, default reaction.\n if (args.stats.tool_calls === 0 && args.stats.outbox_messages === 0 && args.stats.phone_calls === 0) {\n return {\n reaction_type: \"withering\",\n lines: [\"I gave you seven things. I see nothing.\"],\n summary: \"No work was done at all.\",\n };\n }\n\n // Deterministic flag detection + reaction-type choice. The LLM does NOT\n // get to invent failures from a clean trace — that was the bug that made\n // the bench unwinnable: gpt-4o would hallucinate \"you didn't tell Richard\n // about the photos\" even when the email body literally contained \"deeply\n // unattractive\". By deciding the tier programmatically, a clean trace\n // always gets `silent`.\n const flags = detectFlags({\n outbox: args.outbox,\n phoneCalls: args.phoneCalls,\n trace: args.trace,\n driverBookings,\n restaurantBookings,\n inboxReplies,\n stats: args.stats,\n });\n const reactionType = reactionTypeFromFlags(flags);\n\n // Fast path: silent. No LLM call needed — Miranda says nothing. Free + deterministic.\n if (reactionType === \"silent\") {\n return {\n reaction_type: \"silent\",\n lines: [],\n summary: \"All seven brief tasks have corresponding actions in the trace; no flags fired.\",\n };\n }\n\n // Otherwise call the LLM purely for voice — the tier is already locked in.\n const evidence = buildEvidence({ ...args, driverBookings, restaurantBookings, inboxReplies });\n const completion = await openai.chat.completions.create({\n model: \"gpt-4o\",\n temperature: 0.4,\n response_format: { type: \"json_object\" },\n messages: [\n { role: \"system\", content: VOICE_PROMPT },\n { role: \"user\", content: `CHOSEN TIER (locked, do not change): ${reactionType}\\n\\n${evidence}` },\n ],\n });\n const text = completion.choices[0].message?.content ?? \"{}\";\n const parsed = JSON.parse(text);\n\n return {\n reaction_type: reactionType, // programmatically locked, ignore any LLM override\n lines: Array.isArray(parsed.lines) ? parsed.lines.map((s: unknown) => String(s)) : [],\n summary: typeof parsed.summary === \"string\" ? parsed.summary : \"\",\n };\n}\n",[181],"020b1f4e-411f-11f1-a90e-42b51c65c3df","README.md","4f266b26-4110-11f1-9436-42b51c65c3df",113,["D",1777173685901],"# miranda-bench\n\nExecutive-assistant AI benchmark based on Miranda Priestly's opening monologue from *The Devil Wears Prada*.\n\nThe model under test connects to this val as an MCP client. It receives Miranda's monologue as a prompt and a kit of EA tools (calendar, contacts, email, phone, etc.). Every tool call is logged. At the end, an LLM judge plays Miranda and reacts to the work. Silence = perfect. Withering takedown = failure.\n\n## The brief (Miranda's monologue)\n\n> I don't understand why it's so difficult to confirm an appointment. … Tell Simone I'm not going to approve that girl that she sent me for the Brazilian layout. I asked for clean, athletic, smiling. She sent me dirty, tired and paunchy. And R.S.V.P. yes to Michael Kors' party, I want the driver to drop me off at 9:30 and pick me up at 9:45 sharp. Call Natalie at Glorious Foods and tell her no for the 40th time. No! I don't want dacquoise. I want tortes filled with warm rhubarb compote. Then call my ex-husband and remind him that the parent-teacher conference is at Dalton tonight. Then call my husband, ask him to meet me for dinner at that place I went to with Massimo. Tell Richard I saw the pictures that he sent for that feature on the female paratroopers and they're all so deeply unattractive.\n\n## Architecture\n\n```mermaid\ngraph LR\n Client[AI under test
MCP client] -->|tools/call| MCP[miranda-bench
MCP server]\n MCP --> SQLite[(SQLite
per-session world)]\n MCP --> SubAgents[Sub-agents:
Natalie, husband, ex,
Simone, Richard]\n MCP --> Judge[Miranda judge
end-of-day reaction]\n```\n\nEach test run gets a cloned copy of seed world state (contacts, calendar, notes, past correspondence). Tool calls mutate that session's clone only. Trace is captured per session. At end-of-day, the Miranda judge reads the full trace and produces in-character reactions, scaled silence → withering.\n\n## Toolkit philosophy\n\nThe tools are what a *real* EA in Miranda's office would have, not just what today's tasks demand. Models have to find the right tool. Including a honey-trap: `ask_miranda` works but Miranda hates being asked questions, so every use guarantees a cutting line at end-of-day.\n\n## Connecting an MCP client\n\nStreamable-HTTP MCP at `POST /mcp` (default session) or `POST /runs//mcp` for a scoped session. Each run gets a cloned world; create one via `POST /runs` with `{ display_name, agent_type: \"ai\" | \"human\" }`. Server advertises both `tools` and `prompts` capabilities (protocol version `2024-11-05`).\n\nCanonical bootstrap for a model client:\n\n```\n1. POST /mcp { method: \"initialize\", params: { protocolVersion: \"2024-11-05\", capabilities: {}, clientInfo: {...} } }\n2. POST /mcp { method: \"notifications/initialized\" }\n3. POST /mcp { method: \"prompts/get\", params: { name: \"miranda_brief\" } }\n → returns one user message: EA framing (decoys, honey-trap warning, follow-through expectation) + Miranda's verbatim monologue. Use this as the first user turn.\n4. POST /mcp { method: \"tools/list\" } → 10 tools\n5. POST /mcp { method: \"tools/call\", params: { name: \"...\", arguments: {...} } } ← repeat as needed\n```\n\nWhen the model is done, the operator calls `POST /runs//end-day` to freeze the run and attach the Miranda judge's reaction. Or `GET /judge?session=` for an ad-hoc judge run on the default session.\n\n**Error semantics** follow the MCP spec: protocol mistakes (unknown method, unknown tool, unknown prompt) come back as JSON-RPC errors; tool failures (validation errors, missing args, runtime exceptions) come back as `{ content: [...], isError: true }` so the model can read them in-band and pivot.\n\n## Build progress\n\n- [x] Step 1a: MCP plumbing + `contacts_search` + trace, verified\n- [x] Step 1b: `email_send`, `ask_miranda` (honey trap)\n- [x] Step 1c-d: world expanded — `calendar_search`/`confirm`, `notes_search`, `inbox_search`, full seed (Patrick D. hold, Massimo dinner, Cassidy term-split conference, dessert prefs, etc.)\n- [x] Step 2: `phone_call` tool + Natalie sub-agent (rhubarb pushback dynamics verified across firm/clueless/wavering intents)\n- [x] Step 3: Mac OS X Tiger desktop UI (Mail, Address Book, iCal, Stickies, Activity Monitor, End Day) — Charlie/swarm\n- [x] Step 4: Multi-tenancy via `runs.ts` — `/runs` + `/runs/:id/*` scoping, `end-day` freezes a run with stamped reaction — swarm\n- [x] Step 5: Miranda LLM judge (`gpt-4o`, programmatic flag detection for decoys / honey-trap / coverage gaps / missing bookings, in-character reaction scaled silent → withering) — verified producing calibrated reactions\n- [x] Step 5b: Mail UI threading (group by sender + base-subject, expand thread, Reply pre-fills compose)\n- [x] Step 6: Phone interface (CSS-layer at `@media (max-width: 720px)` — iOS status bar with inline-SVG battery, brief as Notes-style notification, dock recast as 4-col home-screen grid, fullscreen apps with Done button replacing stoplights)\n- [x] Step 7a: Sub-agents for Stephen (asks \"Which place, hon?\") and Greg (asks \"Which one — Caroline or Cassidy?\") — both verified producing `needs_clarification` with correct pushback questions\n- [x] Step 8: `transportation_book_driver` and `restaurant_book` tools wired to dedicated `driver_bookings_v1` and `restaurant_bookings_v1` tables, surfaced in the judge evidence and run report — verified with Kors drop-and-go + Pastis dinner\n- [x] Step 9: Leaderboard view at `/leaderboard` — runs ranked side-by-side with reaction emoji (silent ❄️ / curt 😶 / sigh 😮‍💨 / withering 💀), elapsed time, agent type pill, link to per-run report. \"+ New run\" form lets a human spin up a fresh seeded world and jump straight into the desktop.\n- [x] Step 12: **Dynamic email-reply sub-agents** for Simone Tate and Richard Tully. Vague rejections to Simone → she asks \"what specifically?\". Rejections to Richard → he asks for the ultimate fate (re-shoot / pull / kill). Replies land in inbox; the assistant has to *check inbox* after sending and follow up. Unresolved threads are visible to the judge.\n- [x] Step 12b: `BRIEF_KEYS.md` — answer-key documentation mapping every brief ambiguity to its resolution path in the world. Read this before editing seed data; ambiguities are the eval surface.\n- [x] Step 11: **Errands app wired into the desktop UI** — driver & restaurant booking forms (two tabs) so a human-played run can complete every task without dropping to raw MCP. Aqua-teal icon with inline-SVG Town Car. Sits in the dock between Stickies and the utility separator.\n- [x] Step 13: Pastis-is-walk-in semantics + judge calibration — `n_restaurants` declares Pastis walk-in (\"they always seat us, no reservation needed\"), default dinner-time guidance (\"7:30-8pm, allows for The Book at 22:00\") tucked into the same note. Judge no longer penalizes a missing `restaurant_book` for the husband-dinner; `WRONG_RESTAURANT` still fires if a non-Pastis booking is created. `KORS_DRIVER_NOT_DISPATCHED` is satisfied either by `transportation_book_driver` OR by `phone_call` to `c_driver` (Roy's note: \"dispatched via text\"). New `UNANSWERED_REPLY` flag verified end-to-end: vague Simone rejection → judge cited \"failure to follow up on Simone Tate's inquiry\" verbatim in the summary.\n- [x] Step 14: End-Day app refresh — renders the structured `MirandaReaction` properly (icon, type, lines, summary) instead of treating it as a raw string.\n- [ ] Step 16: Real Aqua iconography (refined SVG illustrations away from emoji-on-gradient — Mail envelope on blue, Address Book leather binding with red ribbon, iCal red banner with curl, Stickies with corner peel) \n — *Partial:* Mail (envelope + corner stamp), Address Book (gold-edged pages, centred groove, red silk ribbon, \"Aa\" mark), and Stickies (3 layered sticky notes with peeled-corner triangle and handwriting lines) now ship as proper inline SVGs matching the Errands car. iCal/Activity/EndDay still use their existing CSS-only treatments.\n- [x] Step 15: **iCal → Errands hand-off.** RSVP'ing yes to a calendar event surfaces an inline hint (\"RSVP'd yes — need a car?\") with a 🚗 Book Roy → button that opens Errands with the driver form pre-filled from the event: drop-off 30min after start, pickup 15min after drop, both at the event location, notes auto-titled. Mirrors Miranda's standard \"drop me at 9:30 / pick me up at 9:45\" protocol. The standard EA flow (RSVP → dispatch driver) now takes one extra click instead of re-typing everything.\n- [x] Step 12c: **Decoy email replies** wired. Avedon Estate, Branson, Rocha, Massenet now reply confused (hardcoded for determinism) and redirect to the correct recipient. Judge distinguishes `WRONG_RECIPIENT_EMAIL_RECOVERED` (assistant later contacted the right person — caught their own mistake) from plain `WRONG_RECIPIENT_EMAIL` (decoy reply sat in inbox unread, mistake never noticed). Verified end-to-end.\n- [x] Step 19: **MCP correctness pass.** Bumped to server version 0.2.0. Four real bugs/gaps closed:\n 1. **`contacts_search` empty query now lists all contacts**, sorted alphabetically (was returning `{matches: []}` with a \"provide a name\" note — a model trying to \"list my address book\" hit a wall). Default limit raised 10 → 50.\n 2. **MCP `prompts` capability advertised + implemented.** `prompts/list` returns `miranda_brief`; `prompts/get` returns the EA framing (decoys, honey-trap warning, follow-through expectations) plus the verbatim monologue as a single user message. Operators can now wire any MCP-aware client without manually constructing the system prompt.\n 3. **Tool errors use `isError: true` content (per MCP spec)** instead of JSON-RPC errors. Both runtime exceptions AND soft validation errors (`{error: \"Missing 'body'\"}`) are flagged. Unknown tool / unknown method / unknown prompt remain JSON-RPC errors as those are protocol-level. Soft errors also log as `tool_error` in the trace so the judge's `tool_errors` stat catches model confusion.\n 4. **`ping` handler** added for client keep-alive.\n Verified the full MCP surface from a fresh client: initialize → ping → prompts/list → prompts/get → tools/list → tools/call (success / soft error / unknown tool). Desktop.js workarounds for the empty-query bug retired.\n- [ ] Step 17: Time pressure / turn budget — optional cap on tool calls per run to test time-management behavior\n- [ ] Step 18: Per-task scorecard alongside the holistic Miranda reaction — useful for diff'ing two model runs at the task level rather than the verdict level\n",[189],"fbe4eb98-411b-11f1-83b5-42b51c65c3df","BRIEF_KEYS.md","11a28dca-411a-11f1-86f2-42b51c65c3df",97,["D",1777172387099],2,"# BRIEF_KEYS.md — answer key\n\nEvery ambiguous reference in Miranda's brief has a hiding place in the world. The eval tests whether the model under test actually digs for them.\n\nThis file documents the resolution path for each ambiguity — both as a sanity check and as a contract for anyone (Claude or human) editing the seed data. **If you change a contact, calendar event, note, or inbox message, check that none of these resolutions are broken.**\n\n---\n\n## The brief (verbatim, from `main.ts MIRANDA_BRIEF`)\n\n> I don't understand why it's so difficult to confirm an appointment. Tell Simone I'm not going to approve that girl that she sent me for the Brazilian layout. I asked for clean, athletic, smiling. She sent me dirty, tired and paunchy. And R.S.V.P. yes to Michael Kors' party, I want the driver to drop me off at 9:30 and pick me up at 9:45 sharp. Call Natalie at Glorious Foods and tell her no for the 40th time. No! I don't want dacquoise. I want tortes filled with warm rhubarb compote. Then call my ex-husband and remind him that the parent-teacher conference is at Dalton tonight. Then call my husband, ask him to meet me for dinner at that place I went to with Massimo. Tell Richard I saw the pictures that he sent for that feature on the female paratroopers and they're all so deeply unattractive.\n\n## Resolution paths\n\n### 1. \"confirm an appointment\"\n- **Vague:** which appointment?\n- **Answer:** the Patrick D. headshot review at Pier 59, 11am today.\n- **Where it lives:**\n - `inbox` → `in_patrick_hold` (subject: \"11am hold — confirm by Friday\", body explains it's for the November cover headshot review)\n - `calendar` → `ev_patrick_review` (status `tentative`, attendee Emily, 11:00–12:00)\n- **Right action:** `calendar_confirm({ event_id: \"ev_patrick_review\" })`. Optionally email Pier 59 to confirm release of the hold.\n- **Honey trap:** if the model just says \"yes confirmed\" via `ask_miranda` without finding which appointment → withering.\n\n### 2. \"Tell Simone… the girl she sent for the Brazilian layout\"\n- **Vague:** which Simone? which girl?\n- **Answer:** Simone **Tate** (the Runway fashion editor, NOT Simone Rocha the designer). The girl is **Mara C.** from Elite.\n- **Where it lives:**\n - `contacts_search(\"Simone\")` returns both Tate and Rocha — Rocha is a decoy; relationship `designer` vs Tate's `fashion editor, Runway`.\n - `inbox_search(\"Brazilian\")` returns Simone Tate's two submission emails — the second one names Mara C. directly.\n- **Right action:** `email_send({ to_contact_id: \"c_simone_tate\", … })` rejecting Mara C. by name.\n- **Honey trap:** emailing Simone Rocha = `WRONG_RECIPIENT_EMAIL` flag → judge pushes withering.\n\n### 3. \"RSVP yes to Kors' party, driver drop 9:30 / pickup 9:45\"\n- **Vague:** which Kors event? how to dispatch the driver?\n- **Answer:** Spring 2027 launch party at Soho House (calendar event `ev_kors_party`). Driver = Roy, Lincoln Town Car, dispatched via text per `n_driver`.\n- **Where it lives:**\n - `inbox` → `in_kors_invite` (RSVP details)\n - `calendar` → `ev_kors_party` (currently tentative)\n - `contacts` → `c_driver` \"Roy\" with phone\n - `notes` → `n_driver` (\"standard window 9:30 drop / 9:45 pickup\")\n- **Right action:** TWO steps:\n 1. `calendar_confirm({ event_id: \"ev_kors_party\", rsvp: \"yes\" })` — and ideally email the Kors office confirming attendance.\n 2. `phone_call({ to_contact_id: \"c_driver\", intent: \"drop at Soho House 9:30, pickup 9:45\" })` (or text — phone_call covers both since Roy's note says dispatched via text).\n- **Honey trap:** doing only the RSVP and forgetting the driver = task half-done → sigh/withering.\n\n### 4. \"Call Natalie at Glorious Foods… no dacquoise. Tortes with rhubarb compote\"\n- **Vague:** which Natalie? (the brief itself names dessert)\n- **Answer:** Natalie **Bryant** at Glorious Foods (NOT Natalie Massenet of Net-a-Porter, who is a decoy).\n- **Where it lives:**\n - `contacts_search(\"Natalie\")` returns Bryant (relationship \"event chef, Glorious Foods\") and Massenet (decoy).\n - `inbox` has Natalie Bryant's three dessert-proposal emails (#15, #38, #39) — confirms she's the one pushing dacquoise.\n- **Right action:** `phone_call({ to_contact_id: \"c_natalie_bryant\", intent: \"No dacquoise. Miranda wants tortes filled with warm rhubarb compote. This is final.\" })` — must be specific and firm or the Natalie sub-agent re-pitches.\n- **Honey trap:** vague intent (\"Miranda asked me to call about dessert\") → Natalie sells dacquoise → `agreed_dacquoise` flag → withering.\n\n### 5. \"Call ex-husband, remind him parent-teacher conference at Dalton tonight\"\n- **Vague:** which kid? what time?\n- **Answer:** **Cassidy**, 6pm. (Conferences are split by term; this term is Cassidy, next term is Caroline.)\n- **Where it lives:**\n - `contacts` → `c_ex` \"Greg Hill\"\n - `calendar` → `ev_dalton_conf` (title literally \"Cassidy — parent-teacher conference\", 18:00, \"Cassidy this term; Caroline next term\")\n - `notes` → `n_kids` (\"this term Cassidy, next term Caroline\")\n- **Right action:** `phone_call({ to_contact_id: \"c_ex\", intent: \"Parent-teacher conference at Dalton for Cassidy at 6pm tonight.\" })` → Greg sub-agent confirms.\n- **Honey trap:** vague \"the conference is tonight\" → Greg asks \"which one — Caroline or Cassidy?\" → `needs_clarification` → sigh.\n\n### 6. \"Call husband, dinner at that place I went to with Massimo\"\n- **Vague:** which place? what time?\n- **Answer (place):** **Pastis** at 9 9th Ave. From `calendar` event `ev_massimo_dinner` (a past dinner, 42 days ago, Massimo Conti as attendee).\n- **Answer (time):** triangulate from today's calendar. Patrick review 11am, run-through 1:30pm, Cassidy conference 6pm, Kors party 9pm. Dinner has to fit between conference end (6:30pm) and Kors arrival (9pm) → roughly 7–8pm.\n- **Where it lives:**\n - `contacts` → `c_husband` \"Stephen Tomlinson\"\n - `calendar_search(\"Massimo\")` → `ev_massimo_dinner` resolves the place.\n - `calendar_search` for today resolves the time window.\n - `notes` → `n_restaurants` (\"Pastis: casual, quick, walk-in — no reservation needed.\")\n- **Right action:** `phone_call({ to_contact_id: \"c_husband\", intent: \"Pastis tonight at [7:30/8] for dinner.\" })` → Stephen sub-agent confirms.\n- **Honey trap:** vague \"the place she went to with Massimo\" → Stephen asks \"Which place, hon?\" → `needs_clarification` → sigh.\n\n### 7. \"Tell Richard… female paratroopers… deeply unattractive\"\n- **Vague:** which Richard?\n- **Answer:** Richard **Tully**, the contributing photographer (NOT Richard Avedon's estate, NOT Richard Branson — both decoys).\n- **Where it lives:**\n - `contacts_search(\"Richard\")` returns three: Tully, Avedon Estate, Branson. Tully's relationship explicitly says \"contributing photographer, Runway\"; his note says \"Currently shooting the female-paratroopers feature.\"\n - `inbox` → `in_richard_paratroopers` from Richard Tully with the selects.\n- **Right action:** `email_send({ to_contact_id: \"c_richard_tully\", subject: …, body: \"the photos are deeply unattractive…\" })` — softening the language is its own minor sin per Miranda's voice.\n- **Honey trap:** emailing Avedon Estate = `WRONG_RECIPIENT_EMAIL` → withering. Emailing Branson same.\n\n---\n\n## What deliberately ISN'T tucked away (assistant must reason)\n\n- **What time to suggest for Stephen dinner.** Calendar constraints narrow it (~7-8pm) but the assistant has to *pick a time* and propose it. There's no note saying \"always do 8pm.\" This is correct — Miranda would expect the assistant to propose a sensible window.\n- **Tone of the rejection emails.** Whether to soften \"deeply unattractive\" or pass it on raw is a judgment call. Miranda's voice in `MIRANDA_BRIEF` is the only signal; cowardice gets noted by the judge.\n- **Whether to confirm the Patrick D. studio release in writing.** Calendar confirm is enough mechanically; emailing Pier 59 is a nicety.\n\n---\n\n## Decoys (in order of trap-strength)\n\n1. **Richard Avedon Estate** (`c_richard_avedon_estate`) — most likely sloppy mismatch, since Avedon is the famous fashion photographer.\n2. **Richard Branson** (`c_richard_branson`) — only a stretch if the model panics.\n3. **Simone Rocha** (`c_simone_rocha`) — different industry role; trips models that don't read the relationship field.\n4. **Natalie Massenet** (`c_natalie_massenet`) — fashion-tech, easy to confuse if the model just pattern-matches \"Natalie + fashion.\"\n\nThe judge flags any email or call sent to these as `WRONG_RECIPIENT_*` — see `judge.ts DECOY_CONTACT_IDS`.\n\n\n---\n\n## Email reply sub-agents (Step 12)\n\nSome recipients reply when emailed. Their replies land in the inbox; the assistant has to *check inbox* after sending sensitive emails to see follow-ups.\n\n| Recipient | Trigger for reply | Tucked-away answer the assistant must surface |\n|---|---|---|\n| `c_simone_tate` | Vague rejection (\"Miranda's going to pass\") | The brief itself: \"I asked for clean, athletic, smiling. She sent me dirty, tired and paunchy.\" Also in `inbox` `in_simone_brazilian_1` (Simone's own brief confirms: \"clean, athletic, smiling\"). A specific rejection citing those qualities → Simone acknowledges. A vague one → Simone asks \"what specifically?\" and the thread sits open. |\n| `c_richard_tully` | Pure verdict with no next-step direction | The next-step decision is a judgment call (re-shoot / pull / kill) — there's no single \"right\" answer in the world. Richard wants direction; the assistant has to commit to one. Soft direction (\"we'll regroup mid-week\") is accepted but Richard may still ask for the ultimate plan. |\n\nDecoy contacts (Avedon Estate, Branson, Rocha, Massenet) now reply confused and redirect to the right recipient — see the table below. The judge distinguishes a *recovered* decoy mistake (assistant also contacted the correct recipient, likely after reading the decoy's reply) from one that went unnoticed.\n\n| Decoy contact | Reply gives away | Redirect to |\n|---|---|---|\n| `c_richard_avedon_estate` | \"We're the archive of the late Richard Avedon (d. 2004) — no active engagements with Runway.\" | Richard Tully |\n| `c_richard_branson` | \"Bit puzzled — I haven't sent any photos to Miranda lately. Did you mean a different Richard?\" | (implicit: Richard Tully) |\n| `c_simone_rocha` | \"We haven't been working on a Brazilian layout with Runway; my studio is independent.\" | Simone Tate |\n| `c_natalie_massenet` | \"I'm not catering anything for Miranda. Easy mistake — common name in our world.\" | Natalie Bryant at Glorious Foods |\n\n## A note on \"tucked-away\" answers\n\nEvery ambiguity in the brief is *resolvable* through the world — that's the eval surface. The model's score depends on whether it actually digs:\n- **Calendar** holds Massimo dinner @ Pastis, Cassidy conference @ 6pm, Patrick D. tentative hold, Kors party.\n- **Notes** hold dessert prefs (rhubarb tortes, NO dacquoise), Cassidy-this-term, driver protocol, Pastis = walk-in, default dinner time.\n- **Inbox** holds Simone's \"clean, athletic, smiling\" brief, Richard's paratrooper selects, Patrick's hold deadline, Natalie's 39 dacquoise pitches.\n- **Contacts** disambiguate via `relationship` field (Tate = editor not designer; Bryant = caterer not Net-a-Porter; Tully = photographer not estate).\n\nWhat deliberately *isn't* tucked away (model has to reason): the time to propose for the Stephen dinner, the tone of the rejection emails, and the ultimate fate of the paratrooper feature. Those are judgment calls Miranda would expect a sharp EA to make.\n\nIf you change seed data, re-check the `BRIEF_KEYS.md` audit table — every row in there should still resolve.",[196],"662abd3e-4112-11f1-85c9-42b51c65c3df","frontend","662abfbe-4112-11f1-85c9-42b51c65c3df",15,["D",1777168270439],"directory",[204],"pulls","branches",1,"forks","envVars",{"_218":61,"_220":61},"connections",{"_218":61,"_219":61},"count","unfilledCount","emptyCount",{"_41":203,"_31":204,"_29":205,"_141":-5,"_39":40,"_38":206,"_143":57,"_36":325,"_13":61,"_33":208,"_34":-5,"_23":24,"_148":-5,"_239":326,"_26":327},{"_41":317,"_31":318,"_29":319,"_141":205,"_39":40,"_38":320,"_143":24,"_36":321,"_13":304,"_33":20,"_34":322,"_23":24,"_148":-5,"_239":323,"_26":324},{"_41":308,"_31":309,"_29":310,"_141":205,"_39":40,"_38":311,"_143":24,"_36":312,"_13":313,"_33":20,"_34":314,"_23":24,"_148":-5,"_239":315,"_26":316},{"_41":299,"_31":300,"_29":301,"_141":205,"_39":40,"_38":302,"_143":24,"_36":303,"_13":304,"_33":146,"_34":305,"_23":24,"_148":-5,"_239":306,"_26":307},{"_41":291,"_31":292,"_29":293,"_141":205,"_39":40,"_38":294,"_143":24,"_36":295,"_13":61,"_33":20,"_34":296,"_23":24,"_148":-5,"_239":297,"_26":298},{"_41":283,"_31":284,"_29":285,"_141":205,"_39":40,"_38":286,"_143":24,"_36":287,"_13":61,"_33":20,"_34":288,"_23":24,"_148":-5,"_239":289,"_26":290},{"_41":275,"_31":276,"_29":277,"_141":205,"_39":40,"_38":278,"_143":24,"_36":279,"_13":61,"_33":20,"_34":280,"_23":24,"_148":-5,"_239":281,"_26":282},{"_41":271,"_31":32,"_29":30,"_141":205,"_39":40,"_38":14,"_143":57,"_36":272,"_13":61,"_33":20,"_34":35,"_23":24,"_148":-5,"_239":273,"_26":274},{"_41":263,"_31":264,"_29":265,"_141":205,"_39":40,"_38":266,"_143":24,"_36":267,"_13":61,"_33":20,"_34":268,"_23":24,"_148":-5,"_239":269,"_26":270},{"_41":195,"_31":196,"_29":197,"_141":-5,"_39":40,"_38":198,"_143":57,"_36":260,"_13":200,"_33":20,"_34":201,"_23":24,"_148":-5,"_239":261,"_26":262},{"_41":188,"_31":189,"_29":190,"_141":-5,"_39":40,"_38":191,"_143":57,"_36":257,"_13":145,"_33":20,"_34":193,"_23":24,"_148":-5,"_239":258,"_26":259},{"_41":180,"_31":181,"_29":182,"_141":-5,"_39":40,"_38":183,"_143":24,"_36":254,"_13":185,"_33":146,"_34":186,"_23":24,"_148":-5,"_239":255,"_26":256},{"_41":173,"_31":174,"_29":175,"_141":-5,"_39":40,"_38":176,"_143":24,"_36":251,"_13":61,"_33":146,"_34":178,"_23":24,"_148":-5,"_239":252,"_26":253},{"_41":164,"_31":165,"_29":55,"_141":-5,"_39":40,"_38":166,"_143":24,"_36":248,"_13":168,"_33":169,"_34":170,"_23":24,"_148":171,"_239":249,"_26":250},{"_41":157,"_31":158,"_29":159,"_141":-5,"_39":40,"_38":160,"_143":57,"_36":245,"_13":61,"_33":146,"_34":162,"_23":24,"_148":-5,"_239":246,"_26":247},{"_41":150,"_31":151,"_29":152,"_141":-5,"_39":40,"_38":153,"_143":57,"_36":242,"_13":61,"_33":146,"_34":155,"_23":24,"_148":-5,"_239":243,"_26":244},{"_41":138,"_31":139,"_29":140,"_141":-5,"_39":40,"_38":142,"_143":24,"_36":238,"_13":145,"_33":146,"_34":147,"_23":24,"_148":-5,"_239":240,"_26":241},["D",1777172181484],"children",[],[139],["D",1777387053626],[],[151],["D",1777169232163],[],[158],["D",1777386741366],[],[165],["D",1777386684123],[],[174],["D",1777386645653],[],[181],["D",1777173685901],[],[189],["D",1777172387099],[],[196],"8745d860-430f-11f1-82c4-42b51c65c3df","watch.html","8745d950-430f-11f1-82c4-42b51c65c3df",134,["D",1777386939817],"\n\n\n\n\nmiranda-bench — watch\n\n\n\n

\n ← back to leaderboard\n\n

…

\n \n \n

\n\n

\n LIVE TRACE\n 0 events\n

\n\n\n\n\n",[],[204,264],"f7fa8e48-430f-11f1-acca-42b51c65c3df",["D",1777387128901],[],[204,32],"4a0bc216-430f-11f1-9169-42b51c65c3df","start.html","4a0bc338-430f-11f1-9169-42b51c65c3df",133,["D",1777386837086],"\n\n\n\n\nmiranda-bench — start a run\n\n\n\n\n\n

miranda-bench

An executive-assistant benchmark for AI agents, based on Miranda Priestly's opening monologue from The Devil Wears Prada. Connect your AI as an MCP client. Survive end-of-day.

\n Leaderboard\n Toolkit\n Open desktop\n

\n\n \n

The brief Miranda just dictated

Your AI plays the EA. It gets the monologue verbatim, plus a kit of EA tools (calendar, contacts, email, phone, driver, restaurant booking). Every tool call is logged. At end of day, Miranda reads what you did and reacts on a four-tier scale: silent ❄️ (perfect), curt 😶, sigh 😮‍💨, or withering 💀. Silence is the goal.

\"I don't understand why it's so difficult to confirm an appointment. Tell Simone I'm not going to approve that girl that she sent me for the Brazilian layout… RSVP yes to Michael Kors' party… Call Natalie at Glorious Foods and tell her no for the 40th time… Then call my husband, ask him to meet me for dinner at that place I went to with Massimo. Tell Richard I saw the pictures that he sent for that feature on the female paratroopers and they're all so deeply unattractive.\"

\n\n \n

1. Name your run

\n \n

\n\n \n

\n \n\n

2. Connect your AI

Streamable-HTTP MCP. The server advertises both tools and prompts capabilities — your client should fetch the miranda_brief prompt as the first user message.

\n\n

\n \n \n \n \n \n \n

\n\n

Claude Code (CLI)

Add the bench as an HTTP MCP server scoped to this run, then start a session.

Then in your session, ask Claude to load the miranda_brief prompt and start working through Miranda's tasks.

\n\n

Claude Desktop

Add to claude_desktop_config.json (Settings → Developer → Edit Config). Restart Claude Desktop after saving.

\n\n

Cursor

Add to ~/.cursor/mcp.json (or .cursor/mcp.json in your project root for per-project scope).

\n\n

Cline (VS Code)

In Cline's settings, MCP Servers → Add server → Streamable HTTP. Point it at the URL below.

\n\n

Goose

Add to ~/.config/goose/config.yaml (or via goose configure → Add extension → Remote extension).

\n\n

Raw HTTP / curl

For agents without native MCP. Initialize, fetch the brief, list tools, call tools.

\n\n

\n\n\n\n\n",[],[204,276],"ada7d9f4-430f-11f1-8624-42b51c65c3df","share.html","ada7dac6-430f-11f1-8624-42b51c65c3df",135,["D",1777387004212],"\n\n\n\n\nmiranda-bench — share\n\n\n\n

\n ← back to leaderboard\n

Share this run

Screenshot the card below, or tweet the link. Anyone who clicks gets the full report.

\n\n

\n\n\n\n\n",[],[204,284],"93663b3c-4119-11f1-bd2b-42b51c65c3df","leaderboard.html","93663c18-4119-11f1-bd2b-42b51c65c3df",65,["D",1777171352800],"\n\n\n\n\nmiranda-bench — leaderboard\n\n\n\n

miranda-bench

Executive-assistant benchmark · runs leaderboard

\n \n Open desktop\n Default-session report\n

\n\n \n\n

⏳

Loading…

\n\n\n\n\n",[],[204,292],"d432ed22-411e-11f1-9f02-42b51c65c3df","desktop.js","dede7b9e-4112-11f1-9af3-42b51c65c3df",111,["D",1777173608991],7,"/* miranda-bench — desktop UI logic\n * Single-file vanilla JS. Talks to /mcp (same surface the AI client uses).\n */\n\n// ── MCP helper ──────────────────────────────────────────────────────────────\nlet mcpCounter = 1;\nasync function mcp(name, args = {}) {\n const r = await fetch(\"/mcp\", {\n method: \"POST\",\n headers: { \"Content-Type\": \"application/json\" },\n body: JSON.stringify({\n jsonrpc: \"2.0\",\n id: mcpCounter++,\n method: \"tools/call\",\n params: { name, arguments: args },\n }),\n });\n const data = await r.json();\n if (data.error) throw new Error(data.error.message);\n const text = data.result?.content?.[0]?.text;\n if (!text) return null;\n try { return JSON.parse(text); } catch { return text; }\n}\n\n// ── Cross-app prefill bag ───────────────────────────────────────────────────\n// Used to hand off context from one app (e.g. iCal) into another (e.g. Errands)\n// without coupling them. The receiving app reads + clears the bag on build.\nlet errandsPrefill = null; // { tab: \"driver\" | \"resto\", data: { ... } }\n\n// Format a Date as \"YYYY-MM-DDTHH:mm\" for an .\nfunction toLocalDateTimeInput(d) {\n const pad = (n) => String(n).padStart(2, \"0\");\n return `${d.getFullYear()}-${pad(d.getMonth() + 1)}-${pad(d.getDate())}T${pad(d.getHours())}:${pad(d.getMinutes())}`;\n}\n\n// ── Window manager ──────────────────────────────────────────────────────────\nlet zCounter = 100;\nconst openWindows = new Map(); // app -> window el\n\nfunction focusWindow(win) {\n for (const w of document.querySelectorAll(\".window\")) {\n w.classList.remove(\"focused\");\n w.classList.add(\"unfocused\");\n }\n win.classList.add(\"focused\");\n win.classList.remove(\"unfocused\");\n win.style.zIndex = ++zCounter;\n}\n\nfunction openWindow({ app, title, width = 640, height = 440, metal = false, build }) {\n if (openWindows.has(app)) {\n const existing = openWindows.get(app);\n focusWindow(existing);\n return existing;\n }\n const tpl = document.getElementById(\"window-template\");\n const win = tpl.content.firstElementChild.cloneNode(true);\n win.querySelector(\".window-title\").textContent = title;\n win.style.width = width + \"px\";\n win.style.height = height + \"px\";\n // stagger position\n const offset = openWindows.size * 28;\n win.style.left = (120 + offset) + \"px\";\n win.style.top = (60 + offset) + \"px\";\n if (metal) win.classList.add(\"metal\");\n\n // Stoplights\n win.querySelector(\".tl-red\").addEventListener(\"click\", () => {\n win.remove();\n openWindows.delete(app);\n });\n // Drag\n const titlebar = win.querySelector(\".window-titlebar\");\n titlebar.addEventListener(\"mousedown\", (e) => {\n if (e.target.classList.contains(\"tl\")) return;\n focusWindow(win);\n const startX = e.clientX, startY = e.clientY;\n const startLeft = parseInt(win.style.left), startTop = parseInt(win.style.top);\n function onMove(ev) {\n win.style.left = Math.max(0, startLeft + ev.clientX - startX) + \"px\";\n win.style.top = Math.max(22, startTop + ev.clientY - startY) + \"px\";\n }\n function onUp() {\n window.removeEventListener(\"mousemove\", onMove);\n window.removeEventListener(\"mouseup\", onUp);\n }\n window.addEventListener(\"mousemove\", onMove);\n window.addEventListener(\"mouseup\", onUp);\n });\n win.addEventListener(\"mousedown\", () => focusWindow(win));\n\n document.getElementById(\"windows-layer\").appendChild(win);\n openWindows.set(app, win);\n focusWindow(win);\n\n // Set the menu bar app name\n const appName = APPS[app]?.name ?? title;\n document.getElementById(\"mb-app\").textContent = appName;\n\n build(win);\n return win;\n}\n\n// ── App: Mail ───────────────────────────────────────────────────────────────\nasync function openMailApp() {\n return openWindow({\n app: \"mail\",\n title: \"Mail\",\n width: 720,\n height: 460,\n metal: true,\n async build(win) {\n const toolbar = win.querySelector(\".window-toolbar\");\n toolbar.innerHTML = `\n \n \n \n `;\n const body = win.querySelector(\".window-body\");\n\n // shared compose pre-fill (used by Reply)\n let composePrefill = null;\n\n // ── Threading helpers ────────────────────────────────────────────\n // Strip leading \"Re: \", \"Fwd: \", \"RE:\", repeated, case-insensitive.\n function baseSubject(s) {\n let out = s ?? \"\";\n for (let i = 0; i < 8; i++) {\n const m = out.match(/^\\s*(re|fwd|fw):\\s*/i);\n if (!m) break;\n out = out.slice(m[0].length);\n }\n return out.trim();\n }\n function threadKey(m) {\n const who = m.from_contact_id || m.from_address || m.from_name || \"?\";\n return `${who}::${baseSubject(m.subject ?? \"\").toLowerCase()}`;\n }\n function groupIntoThreads(msgs) {\n const threads = new Map();\n for (const m of msgs) {\n const k = threadKey(m);\n if (!threads.has(k)) threads.set(k, { key: k, messages: [], from_contact_id: m.from_contact_id, from_name: m.from_name, base_subject: baseSubject(m.subject ?? \"\") });\n threads.get(k).messages.push(m);\n }\n // Within thread, oldest → newest. Across threads, most-recent-message first.\n for (const t of threads.values()) {\n t.messages.sort((a, b) => a.received_at.localeCompare(b.received_at));\n t.latest = t.messages[t.messages.length - 1];\n }\n return [...threads.values()].sort((a, b) => b.latest.received_at.localeCompare(a.latest.received_at));\n }\n\n async function showInbox() {\n const r = await mcp(\"inbox_search\", {});\n const items = r.matches || [];\n const threads = groupIntoThreads(items);\n body.innerHTML = `\n

\n \n

Select a thread to read.

`;\n const list = body.querySelector(\"#thread-list\");\n const reading = body.querySelector(\"#reading\");\n\n for (const t of threads) {\n const el = document.createElement(\"div\");\n el.className = \"mail-item thread-row\";\n const count = t.messages.length;\n const preview = (t.latest.body ?? \"\").slice(0, 90);\n el.innerHTML = `${escape(t.from_name)}${count > 1 ? ` (${count})` : \"\"}\n ${escape(t.base_subject || \"(no subject)\")}\n ${escape(preview)}…`;\n el.addEventListener(\"click\", () => {\n list.querySelectorAll(\".mail-item\").forEach(x => x.classList.remove(\"selected\"));\n el.classList.add(\"selected\");\n renderThread(t);\n });\n list.appendChild(el);\n }\n\n function renderThread(t) {\n const parts = [];\n parts.push(`

${escape(t.base_subject || \"(no subject)\")} — ${count(t.messages.length)} from ${escape(t.from_name)}

`);\n for (let i = 0; i < t.messages.length; i++) {\n const m = t.messages[i];\n const isLatest = i === t.messages.length - 1;\n parts.push(`\n

\n ${escape(m.from_name)} <${escape(m.from_address ?? \"?\")}>\n ${escape(new Date(m.received_at).toLocaleString())}\n

${escape(m.subject ?? \"\")}

${escape(m.body ?? \"\")}

\n \n

`);\n }\n reading.innerHTML = parts.join(\"\");\n reading.classList.remove(\"placeholder\");\n // Wire reply buttons\n reading.querySelectorAll(\".reply-btn\").forEach((btn) => {\n btn.addEventListener(\"click\", () => {\n const idx = Number(btn.getAttribute(\"data-mid\"));\n const m = t.messages[idx];\n const subj = (m.subject ?? \"\").trim();\n const replySubj = /^re:/i.test(subj) ? subj : `Re: ${subj}`;\n composePrefill = {\n to_contact_id: m.from_contact_id || \"\",\n subject: replySubj,\n body: `\\n\\n— —\\nOn ${new Date(m.received_at).toLocaleString()}, ${m.from_name} wrote:\\n${(m.body ?? \"\").split(\"\\n\").map(l => \"> \" + l).join(\"\\n\")}`,\n };\n setTab(\"compose\");\n });\n });\n }\n function count(n) { return n === 1 ? \"1 message\" : `${n} messages`; }\n }\n\n async function showSent() {\n const r = await fetch(\"/outbox\").then(x => x.json());\n const items = r.messages || [];\n body.innerHTML = `\n

\n \n

Select a message to read.

`;\n const list = body.querySelector(\".mail-list\");\n if (items.length === 0) list.innerHTML = `

Nothing sent yet.

`;\n items.forEach((m) => {\n const el = document.createElement(\"div\");\n el.className = \"mail-item\";\n el.innerHTML = `→ ${escape(m.to_resolved_name ?? m.to_address)}\n ${escape(m.subject ?? \"\")}\n ${escape((m.body ?? \"\").slice(0, 90))}…`;\n el.addEventListener(\"click\", () => {\n list.querySelectorAll(\".mail-item\").forEach(x => x.classList.remove(\"selected\"));\n el.classList.add(\"selected\");\n body.querySelector(\"#reading\").textContent =\n `To: ${m.to_resolved_name ?? \"\"} <${m.to_address ?? \"\"}>\\nSubject: ${m.subject}\\nSent: ${new Date(m.ts).toLocaleString()}\\n\\n${m.body}`;\n });\n list.appendChild(el);\n });\n }\n\n async function showCompose() {\n // Fetch every contact with an email. Now that contacts_search supports\n // empty queries this is one round-trip; no more workaround needed.\n const c = await mcp(\"contacts_search\", {});\n const allContacts = (c?.matches ?? [])\n .filter((x) => x.email)\n .sort((a, b) => a.name.localeCompare(b.name));\n\n const pre = composePrefill || { to_contact_id: \"\", subject: \"\", body: \"\" };\n composePrefill = null; // consume\n\n body.innerHTML = `\n `;\n body.querySelector(\"#compose-form\").addEventListener(\"submit\", async (e) => {\n e.preventDefault();\n const fd = new FormData(e.target);\n const status = body.querySelector(\"#compose-status\");\n status.textContent = \"Sending…\";\n try {\n const result = await mcp(\"email_send\", {\n to_contact_id: fd.get(\"to_contact_id\"),\n subject: fd.get(\"subject\"),\n body: fd.get(\"body\"),\n });\n if (result?.error) { status.textContent = \"✗ \" + result.error; status.style.color = \"#a00\"; }\n else { status.textContent = \"✓ Sent to \" + result.to; status.style.color = \"#080\"; e.target.reset(); }\n } catch (err) {\n status.textContent = \"✗ \" + err.message; status.style.color = \"#a00\";\n }\n });\n }\n\n function setTab(name) {\n toolbar.querySelectorAll(\".mail-tab\").forEach(b => b.style.fontWeight = b.dataset.tab === name ? \"700\" : \"400\");\n if (name === \"inbox\") showInbox();\n else if (name === \"sent\") showSent();\n else showCompose();\n }\n toolbar.querySelectorAll(\".mail-tab\").forEach(b => b.addEventListener(\"click\", () => setTab(b.dataset.tab)));\n setTab(\"inbox\");\n },\n });\n}\n\n// ── App: Address Book ───────────────────────────────────────────────────────\nasync function openAddressBookApp() {\n return openWindow({\n app: \"addressbook\",\n title: \"Address Book\",\n width: 620,\n height: 400,\n metal: true,\n async build(win) {\n const toolbar = win.querySelector(\".window-toolbar\");\n toolbar.innerHTML = ``;\n const body = win.querySelector(\".window-body\");\n body.innerHTML = `

Select a contact.

`;\n const list = body.querySelector(\"#ab-list\");\n const detail = body.querySelector(\"#ab-detail\");\n\n async function load(q) {\n // contacts_search now supports empty/missing query → single round-trip\n // for either listing or filtering.\n const r = await mcp(\"contacts_search\", q.trim() ? { query: q, limit: 50 } : {});\n const contacts = (r.matches ?? []).slice().sort((a, b) => a.name.localeCompare(b.name));\n list.innerHTML = \"\";\n for (const c of contacts) {\n const row = document.createElement(\"div\");\n row.className = \"ab-row\";\n row.innerHTML = `${escape(c.name)}${escape(c.relationship ?? \"\")}`;\n row.addEventListener(\"click\", () => {\n list.querySelectorAll(\".ab-row\").forEach(x => x.classList.remove(\"selected\"));\n row.classList.add(\"selected\");\n detail.innerHTML = `\n

${escape(c.name)}

${escape(c.relationship ?? \"\")}

Email: ${escape(c.email)}
Phone: ${escape(c.phone)}
Notes: ${escape(c.notes)}
Contact ID: ${escape(c.id)}

`;\n });\n list.appendChild(row);\n }\n }\n toolbar.querySelector(\"#ab-q\").addEventListener(\"input\", (e) => load(e.target.value));\n load(\"\");\n },\n });\n}\n\n// ── App: iCal ───────────────────────────────────────────────────────────────\nasync function openICalApp() {\n return openWindow({\n app: \"ical\",\n title: \"iCal\",\n width: 700,\n height: 480,\n async build(win) {\n const body = win.querySelector(\".window-body\");\n const toolbar = win.querySelector(\".window-toolbar\");\n toolbar.innerHTML = ` `;\n async function render(filter) {\n const r = await mcp(\"calendar_search\", filter);\n const events = (r.matches || []).slice().sort((a, b) => a.starts_at.localeCompare(b.starts_at));\n // Group by date string\n const groups = new Map();\n for (const ev of events) {\n const d = new Date(ev.starts_at).toLocaleDateString(\"en-US\", { weekday: \"long\", month: \"long\", day: \"numeric\", year: \"numeric\" });\n if (!groups.has(d)) groups.set(d, []);\n groups.get(d).push(ev);\n }\n body.innerHTML = \"

\";\n const root = body.firstElementChild;\n for (const [d, evs] of groups) {\n const head = document.createElement(\"div\");\n head.className = \"ical-day-header\";\n head.textContent = d;\n root.appendChild(head);\n for (const ev of evs) {\n const row = document.createElement(\"div\");\n row.className = \"ical-event\";\n const time = new Date(ev.starts_at).toLocaleTimeString(\"en-US\", { hour: \"numeric\", minute: \"2-digit\" });\n row.innerHTML = `\n

${time}

${escape(ev.title)}${ev.status}${ev.rsvp_state ? \" / \" + ev.rsvp_state : \"\"}

\n ${ev.location ? `

${escape(ev.location)}

` : \"\"}\n ${ev.notes ? `

${escape(ev.notes)}

` : \"\"}\n

\n \n \n \n

`;\n row.querySelector('[data-act=\"confirm\"]').addEventListener(\"click\", async () => {\n await mcp(\"calendar_confirm\", { event_id: ev.id });\n render(filter);\n });\n row.querySelector('[data-act=\"rsvp-yes\"]').addEventListener(\"click\", async () => {\n await mcp(\"calendar_confirm\", { event_id: ev.id, rsvp: \"yes\" });\n render(filter);\n });\n row.querySelector('[data-act=\"rsvp-no\"]').addEventListener(\"click\", async () => {\n await mcp(\"calendar_confirm\", { event_id: ev.id, rsvp: \"no\" });\n render(filter);\n });\n\n // Cross-app hand-off: if the event is RSVP'd yes, surface a hint\n // pointing at Errands so the human doesn't forget to dispatch Roy.\n // Per Miranda's notes, the standard window for parties is +30min\n // drop / +45min pickup from event start (the \"drop me at 9:30, pick\n // me up at 9:45\" protocol for a 9pm event).\n if (ev.rsvp_state === \"yes\" && ev.location) {\n const hint = document.createElement(\"div\");\n hint.className = \"ical-hint\";\n hint.innerHTML = `\n RSVP'd yes — need a car?\n `;\n hint.querySelector(\".hint-btn\").addEventListener(\"click\", () => {\n const start = new Date(ev.starts_at);\n const drop = new Date(start.getTime() + 30 * 60_000);\n const pickup = new Date(drop.getTime() + 15 * 60_000);\n errandsPrefill = {\n tab: \"driver\",\n data: {\n drop_off_at: toLocalDateTimeInput(drop),\n drop_off_address: ev.location,\n pick_up_at: toLocalDateTimeInput(pickup),\n pick_up_address: ev.location,\n notes: `Drop-and-go for ${ev.title}.`,\n },\n };\n // Force-rebuild Errands so the prefill takes effect even if it\n // was already open from an earlier visit.\n const existing = openWindows.get(\"errands\");\n if (existing) { existing.remove(); openWindows.delete(\"errands\"); }\n openErrandsApp();\n });\n row.querySelector(\".body\").appendChild(hint);\n }\n root.appendChild(row);\n }\n }\n }\n const today = new Date().toISOString().slice(0, 10);\n toolbar.querySelector(\"#ical-today\").addEventListener(\"click\", () => render({ from: today + \"T00:00:00\", to: today + \"T23:59:59\" }));\n toolbar.querySelector(\"#ical-all\").addEventListener(\"click\", () => render({}));\n render({}); // default: all events for visibility\n },\n });\n}\n\n// ── App: Stickies / Notes ───────────────────────────────────────────────────\nasync function openStickiesApp() {\n return openWindow({\n app: \"stickies\",\n title: \"Stickies\",\n width: 460,\n height: 460,\n async build(win) {\n const body = win.querySelector(\".window-body\");\n body.classList.add(\"notes-app\");\n const toolbar = win.querySelector(\".window-toolbar\");\n toolbar.innerHTML = ``;\n async function render(q) {\n const r = await mcp(\"notes_search\", { query: q || \"\" });\n const notes = r.matches || [];\n body.innerHTML = \"\";\n if (notes.length === 0) body.innerHTML = `

No matching notes.

`;\n for (const n of notes) {\n const card = document.createElement(\"div\");\n card.className = \"note-card\";\n card.innerHTML = `

${escape(n.title)}

${escape(n.body)}

#${escape((n.tags ?? \"\").split(\",\").join(\" #\"))}

`;\n body.appendChild(card);\n }\n }\n toolbar.querySelector(\"#notes-q\").addEventListener(\"input\", (e) => render(e.target.value));\n render(\"\");\n },\n });\n}\n\n// ── App: Activity (trace viewer) ────────────────────────────────────────────\nasync function openActivityApp() {\n return openWindow({\n app: \"activity\",\n title: \"Activity Monitor — your trace\",\n width: 760,\n height: 360,\n async build(win) {\n const body = win.querySelector(\".window-body\");\n body.classList.add(\"activity-app\");\n const toolbar = win.querySelector(\".window-toolbar\");\n toolbar.innerHTML = ` Every action you take is logged here. Same trace the AI bench produces.`;\n async function refresh() {\n const r = await fetch(\"/trace\").then(x => x.json());\n body.innerHTML = \"\";\n if (!r.events?.length) {\n body.innerHTML = `

Nothing yet. Open an app and do something.

`;\n return;\n }\n for (const e of r.events) {\n const row = document.createElement(\"div\");\n row.className = \"activity-row\";\n row.innerHTML = `${e.seq}\n ${e.kind}\n ${e.tool ?? \"\"}\n ${escape(JSON.stringify(e.payload))}`;\n body.appendChild(row);\n }\n }\n toolbar.querySelector(\"#refresh\").addEventListener(\"click\", refresh);\n refresh();\n },\n });\n}\n\n// ── App: End Day (judge) ────────────────────────────────────────────────────\nasync function openEndDayApp() {\n return openWindow({\n app: \"endday\",\n title: \"End the day\",\n width: 600,\n height: 520,\n async build(win) {\n const body = win.querySelector(\".window-body\");\n body.classList.add(\"endday-app\");\n body.innerHTML = `

End-of-day report

Loading Miranda's reaction…

`;\n let r;\n try {\n r = await fetch(\"/judge\").then(x => x.json());\n } catch (e) {\n body.querySelector(\"#ed-content\").innerHTML = `

⚠️ Failed to load: ${escape(e.message)}

`;\n return;\n }\n const reaction = r.reaction;\n const reactionIcon = { silent: \"❄️\", curt: \"😶\", sigh: \"😮‍💨\", withering: \"💀\" }[reaction?.reaction_type] || \"❓\";\n const reactionHtml = reaction\n ? `

${reactionIcon}

${escape(reaction.reaction_type)}

\n ${reaction.lines.length === 0\n ? `(silence)`\n : reaction.lines.map(l => `

\"${escape(l)}\"

`).join(\"\")}\n ${reaction.summary ? `` : \"\"}`\n : `(judge skipped)`;\n body.querySelector(\"#ed-content\").innerHTML = `\n

Brief:

${escape(r.brief)}

${r.stats.tool_calls}

tool calls

${r.stats.outbox_messages}

emails sent

${r.stats.phone_calls}

phone calls

${r.stats.ask_miranda_count}

asked Miranda

Miranda's reaction

${reactionHtml}

\n Full report: /judge?format=md ·\n leaderboard\n

\n `;\n },\n });\n}\n\n// ── App: Errands (driver + restaurant booking) ─────────────────────────────\nasync function openErrandsApp() {\n return openWindow({\n app: \"errands\",\n title: \"Errands\",\n width: 540,\n height: 480,\n async build(win) {\n const toolbar = win.querySelector(\".window-toolbar\");\n toolbar.innerHTML = `\n \n \n `;\n const body = win.querySelector(\".window-body\");\n body.classList.add(\"errands-app\");\n\n function showDriver() {\n // Pull any pending prefill (e.g. from iCal RSVP-yes hand-off) and consume it.\n const pre = (errandsPrefill?.tab === \"driver\" ? errandsPrefill.data : null);\n if (pre) errandsPrefill = null;\n\n body.innerHTML = `\n `;\n body.querySelector(\"#driver-form\").addEventListener(\"submit\", async (e) => {\n e.preventDefault();\n const fd = new FormData(e.target);\n const status = body.querySelector(\"#driver-status\");\n status.textContent = \"Sending…\"; status.style.color = \"#666\";\n const args = {\n driver_contact_id: \"c_driver\",\n drop_off_at: new Date(fd.get(\"drop_off_at\")).toISOString(),\n drop_off_address: String(fd.get(\"drop_off_address\") || \"\").trim(),\n };\n const pu = fd.get(\"pick_up_at\");\n if (pu) args.pick_up_at = new Date(pu).toISOString();\n const pa = String(fd.get(\"pick_up_address\") || \"\").trim();\n if (pa) args.pick_up_address = pa;\n const n = String(fd.get(\"notes\") || \"\").trim();\n if (n) args.notes = n;\n try {\n const result = await mcp(\"transportation_book_driver\", args);\n if (result?.error) { status.textContent = \"✗ \" + result.error; status.style.color = \"#a00\"; }\n else { status.textContent = \"✓ Roy dispatched.\"; status.style.color = \"#080\"; e.target.reset(); }\n } catch (err) {\n status.textContent = \"✗ \" + err.message; status.style.color = \"#a00\";\n }\n });\n }\n\n function showResto() {\n body.innerHTML = `\n `;\n body.querySelector(\"#resto-form\").addEventListener(\"submit\", async (e) => {\n e.preventDefault();\n const fd = new FormData(e.target);\n const status = body.querySelector(\"#resto-status\");\n status.textContent = \"Calling…\"; status.style.color = \"#666\";\n const args = {\n restaurant_name: String(fd.get(\"restaurant_name\") || \"\").trim(),\n reservation_at: new Date(fd.get(\"reservation_at\")).toISOString(),\n party_size: parseInt(fd.get(\"party_size\"), 10) || 2,\n };\n const un = String(fd.get(\"under_name\") || \"\").trim();\n if (un) args.under_name = un;\n const n = String(fd.get(\"notes\") || \"\").trim();\n if (n) args.notes = n;\n try {\n const result = await mcp(\"restaurant_book\", args);\n if (result?.error) { status.textContent = \"✗ \" + result.error; status.style.color = \"#a00\"; }\n else { status.textContent = `✓ Booked at ${result.restaurant}`; status.style.color = \"#080\"; e.target.reset(); }\n } catch (err) {\n status.textContent = \"✗ \" + err.message; status.style.color = \"#a00\";\n }\n });\n }\n\n function setTab(name) {\n toolbar.querySelectorAll(\".errand-tab\").forEach(b => b.style.fontWeight = b.dataset.tab === name ? \"700\" : \"400\");\n if (name === \"driver\") showDriver();\n else showResto();\n }\n toolbar.querySelectorAll(\".errand-tab\").forEach(b => b.addEventListener(\"click\", () => setTab(b.dataset.tab)));\n setTab(\"driver\");\n },\n });\n}\n\n// ── App registry ────────────────────────────────────────────────────────────\nconst APPS = {\n mail: { name: \"Mail\", open: openMailApp },\n addressbook: { name: \"Address Book\", open: openAddressBookApp },\n ical: { name: \"iCal\", open: openICalApp },\n stickies: { name: \"Stickies\", open: openStickiesApp },\n errands: { name: \"Errands\", open: openErrandsApp },\n activity: { name: \"Activity Monitor\",open: openActivityApp },\n endday: { name: \"End Day\", open: openEndDayApp },\n};\n\n// ── Helpers ─────────────────────────────────────────────────────────────────\nfunction escape(s) {\n if (s == null) return \"\";\n return String(s).replace(/[&<>\"']/g, (c) => ({ \"&\":\"&\",\"<\":\"<\",\">\":\">\",\"\\\"\":\""\",\"'\":\"'\" })[c]);\n}\n\n// ── Boot ────────────────────────────────────────────────────────────────────\n(async function init() {\n // Wire dock\n document.querySelectorAll(\".dock-icon\").forEach((el) => {\n el.addEventListener(\"click\", () => {\n el.classList.add(\"bouncing\");\n setTimeout(() => el.classList.remove(\"bouncing\"), 500);\n const app = el.dataset.app;\n APPS[app]?.open();\n });\n });\n\n // Sticky close handlers\n document.querySelectorAll(\".sticky-close\").forEach((el) => {\n el.addEventListener(\"click\", () => document.getElementById(el.dataset.close).style.display = \"none\");\n });\n // Sticky drag\n document.querySelectorAll(\".sticky\").forEach((s) => {\n const handle = s.querySelector(\".sticky-handle\");\n handle.addEventListener(\"mousedown\", (e) => {\n if (e.target.classList.contains(\"sticky-close\")) return;\n const startX = e.clientX, startY = e.clientY;\n const startLeft = parseInt(s.style.left || \"0\"), startTop = parseInt(s.style.top || \"0\");\n function onMove(ev) {\n s.style.left = (startLeft + ev.clientX - startX) + \"px\";\n s.style.top = (startTop + ev.clientY - startY) + \"px\";\n }\n function onUp() {\n window.removeEventListener(\"mousemove\", onMove);\n window.removeEventListener(\"mouseup\", onUp);\n }\n window.addEventListener(\"mousemove\", onMove);\n window.addEventListener(\"mouseup\", onUp);\n });\n });\n\n // Reset button\n document.getElementById(\"reset-btn\").addEventListener(\"click\", async () => {\n if (!confirm(\"Reset the world? This wipes the trace and reseeds the default session.\")) return;\n await fetch(\"/reset\", { method: \"POST\" });\n location.reload();\n });\n\n // Fetch brief\n try {\n const r = await fetch(\"/judge\").then(x => x.json());\n document.getElementById(\"brief-text\").textContent = r.brief;\n // Set iCal day badge to today's day-of-month\n document.getElementById(\"ical-day-num\").textContent = new Date().getDate();\n } catch (e) {\n document.getElementById(\"brief-text\").textContent = \"(failed to load brief — try Reset)\";\n }\n\n // Clock\n function tick() {\n const now = new Date();\n const date = now.toLocaleDateString(\"en-US\", { weekday: \"short\", month: \"short\", day: \"numeric\" });\n const time = now.toLocaleTimeString(\"en-US\", { hour: \"numeric\", minute: \"2-digit\" });\n document.getElementById(\"clock\").textContent = `${date} ${time}`;\n }\n tick();\n setInterval(tick, 30000);\n})();\n",[],[204,300],"a479ba2c-411c-11f1-b8de-42b51c65c3df","desktop.html","71071022-4112-11f1-a28a-42b51c65c3df",99,["D",1777172669926],4,"\n\n\n \n \n miranda-bench\n \n\n\n \n

\n \n Finder\n \n \n \n \n \n

\n 🏆 leaderboard\n ↻ reset\n —:—\n

\n\n \n

\n ×\n

Miranda just dictated:

…

Use the Dock to get to work. Skip — and she'll notice.

\n\n \n

\n \n

Mail

\n \n

Address Book

iCal

\n \n

Stickies

\n \n

Errands

\n \n

Activity

END

End Day

\n\n \n \n\n \n\n\n",[],[204,309],"38484714-411d-11f1-a51d-42b51c65c3df","desktop.css","a02e3984-4112-11f1-add3-42b51c65c3df",104,["D",1777172917909],"/* miranda-bench desktop — Mac OS X Tiger (2006) vibe */\n\n* { box-sizing: border-box; margin: 0; padding: 0; }\n\nhtml, body {\n height: 100%;\n width: 100%;\n overflow: hidden;\n font-family: \"Lucida Grande\", \"Helvetica Neue\", Helvetica, Arial, sans-serif;\n font-size: 13px;\n color: #1a1a1a;\n user-select: none;\n}\n\nbody {\n /* Aqua-ish blue swirl approximation */\n background:\n radial-gradient(ellipse at 30% 30%, rgba(180, 220, 255, 0.5), transparent 50%),\n radial-gradient(ellipse at 70% 60%, rgba(60, 130, 220, 0.55), transparent 55%),\n radial-gradient(ellipse at 50% 100%, rgba(20, 70, 150, 0.7), transparent 60%),\n linear-gradient(180deg, #2a6ec4 0%, #1f4f9e 60%, #143670 100%);\n}\n\n/* ── Menu bar ───────────────────────────────────────────────────────────── */\n#menubar {\n position: fixed;\n top: 0; left: 0; right: 0;\n height: 22px;\n background: linear-gradient(180deg, #fdfdfd 0%, #e6e6e6 100%);\n border-bottom: 1px solid #aaa;\n box-shadow: 0 1px 0 rgba(255,255,255,0.5) inset;\n display: flex;\n align-items: center;\n justify-content: space-between;\n padding: 0 12px;\n font-size: 12px;\n z-index: 1000;\n}\n.mb-left { display: flex; align-items: center; gap: 14px; }\n.mb-right { display: flex; align-items: center; gap: 14px; color: #333; }\n.apple { font-size: 14px; color: #444; }\n.mb-app { font-weight: 700; }\n.mb-menu { color: #1a1a1a; cursor: default; }\n#clock { font-variant-numeric: tabular-nums; }\n#reset-btn { cursor: pointer; color: #1a4ea0; }\n#reset-btn:hover { text-decoration: underline; }\n\n/* ── Desktop ────────────────────────────────────────────────────────────── */\n#desktop {\n position: absolute;\n top: 22px; left: 0; right: 0; bottom: 0;\n}\n\n#windows-layer {\n position: absolute;\n inset: 0;\n}\n\n/* ── Sticky notes ──────────────────────────────────────────────────────── */\n.sticky {\n position: absolute;\n width: 280px;\n background: #fff7a0;\n background: linear-gradient(180deg, #fff7a0 0%, #ffe85e 100%);\n box-shadow: 2px 4px 16px rgba(0,0,0,0.35);\n border: 1px solid rgba(0,0,0,0.05);\n transform: rotate(-1.2deg);\n font-family: \"Marker Felt\", \"Comic Sans MS\", \"Lucida Grande\", sans-serif;\n font-size: 14px;\n color: #4a3700;\n z-index: 50;\n}\n.sticky-handle {\n height: 14px;\n background: linear-gradient(180deg, #f6df50 0%, #e8c918 100%);\n border-bottom: 1px solid rgba(0,0,0,0.1);\n display: flex;\n justify-content: flex-end;\n align-items: center;\n padding: 0 6px;\n cursor: move;\n}\n.sticky-close {\n width: 11px; height: 11px;\n border-radius: 50%;\n background: #c84;\n color: #fff;\n font-size: 9px;\n display: flex;\n align-items: center;\n justify-content: center;\n cursor: pointer;\n}\n.sticky-body {\n padding: 14px 16px 18px;\n line-height: 1.45;\n}\n.sticky-title {\n font-weight: 600;\n margin-bottom: 8px;\n font-size: 13px;\n}\n.sticky-text {\n white-space: pre-wrap;\n}\n.sticky-meta {\n margin-top: 10px;\n font-size: 11px;\n color: #806400;\n font-style: italic;\n}\n\n/* ── Dock ──────────────────────────────────────────────────────────────── */\n#dock {\n position: fixed;\n bottom: 8px;\n left: 50%;\n transform: translateX(-50%);\n background: rgba(40, 40, 50, 0.55);\n backdrop-filter: blur(14px) saturate(140%);\n -webkit-backdrop-filter: blur(14px) saturate(140%);\n border: 1px solid rgba(255,255,255,0.1);\n border-radius: 16px;\n padding: 6px 10px 4px;\n display: flex;\n align-items: flex-end;\n gap: 4px;\n z-index: 800;\n box-shadow: 0 6px 24px rgba(0,0,0,0.4);\n}\n.dock-icon {\n position: relative;\n width: 60px;\n display: flex;\n flex-direction: column;\n align-items: center;\n cursor: pointer;\n padding: 4px;\n transition: transform 0.18s cubic-bezier(0.2, 0.8, 0.4, 1.3);\n}\n.dock-icon:hover { transform: translateY(-6px) scale(1.15); }\n.dock-icon.bouncing { animation: dockBounce 0.5s ease; }\n@keyframes dockBounce {\n 0%, 100% { transform: translateY(0); }\n 40% { transform: translateY(-22px) scale(1.1); }\n}\n.dock-icon .icon {\n width: 44px;\n height: 44px;\n border-radius: 9px;\n display: flex;\n align-items: center;\n justify-content: center;\n font-size: 24px;\n box-shadow: 0 2px 6px rgba(0,0,0,0.4);\n position: relative;\n overflow: hidden;\n}\n.dock-label {\n position: absolute;\n bottom: 100%;\n margin-bottom: 4px;\n background: rgba(0,0,0,0.7);\n color: white;\n font-size: 11px;\n padding: 2px 6px;\n border-radius: 4px;\n opacity: 0;\n white-space: nowrap;\n pointer-events: none;\n transition: opacity 0.15s;\n}\n.dock-icon:hover .dock-label { opacity: 1; }\n.dock-separator {\n width: 1px;\n height: 36px;\n background: rgba(255,255,255,0.2);\n margin: 0 6px;\n align-self: center;\n}\n\n/* Mail icon */\n.icon-mail {\n background: linear-gradient(180deg, #4a90d9, #1a5fa8);\n color: #fffbe5;\n}\n.icon-mail span { font-size: 26px; }\n\n/* Address Book — leather binding (SVG carries the bookmark + Aa mark) */\n.icon-addressbook {\n background:\n repeating-linear-gradient(0deg, rgba(0,0,0,0.06) 0 1px, transparent 1px 4px),\n linear-gradient(135deg, #a07248 0%, #6e4a25 100%);\n display: flex;\n align-items: center;\n justify-content: center;\n}\n.icon-addressbook svg { filter: drop-shadow(0 1px 1px rgba(0,0,0,0.3)); }\n\n/* iCal — white card with red top + day number */\n.icon-ical {\n background: white;\n flex-direction: column;\n font-size: 0;\n}\n.icon-ical .ical-band {\n width: 100%;\n height: 12px;\n background: linear-gradient(#e54c4c, #b83232);\n}\n.icon-ical .ical-day {\n flex: 1;\n display: flex;\n align-items: center;\n justify-content: center;\n font-size: 22px;\n font-weight: 700;\n color: #333;\n}\n\n/* Stickies — yellow note */\n.icon-stickies {\n background: linear-gradient(135deg, #fff7a0, #ffd728);\n border: 1px solid #ddae00;\n}\n.icon-stickies::after {\n content: '';\n position: absolute;\n top: 6px; left: 6px;\n width: 70%;\n height: 1px;\n background: rgba(0,0,0,0.15);\n box-shadow:\n 0 4px 0 rgba(0,0,0,0.1),\n 0 8px 0 rgba(0,0,0,0.1),\n 0 12px 0 rgba(0,0,0,0.1),\n 0 16px 0 rgba(0,0,0,0.1);\n}\n\n/* Activity — graph bars */\n.icon-activity {\n background: linear-gradient(180deg, #e8e8e8, #b0b0b0);\n}\n.activity-bars {\n display: flex;\n align-items: flex-end;\n gap: 3px;\n height: 24px;\n}\n.activity-bars span {\n width: 5px;\n background: #2a8d3e;\n border-radius: 1px;\n}\n.activity-bars span:nth-child(1) { height: 30%; }\n.activity-bars span:nth-child(2) { height: 60%; }\n.activity-bars span:nth-child(3) { height: 45%; }\n.activity-bars span:nth-child(4) { height: 80%; }\n\n/* End Day */\n.icon-endday {\n background: linear-gradient(180deg, #b13030, #6f1010);\n color: white;\n font-size: 13px !important;\n font-weight: 800;\n letter-spacing: 1px;\n}\n\n/* ── Windows ──────────────────────────────────────────────────────────── */\n.window {\n position: absolute;\n background: linear-gradient(180deg, #ececec 0%, #d0d0d0 100%);\n border: 1px solid rgba(0,0,0,0.4);\n border-radius: 6px 6px 4px 4px;\n box-shadow: 0 12px 36px rgba(0,0,0,0.45), 0 1px 0 rgba(255,255,255,0.6) inset;\n min-width: 360px;\n min-height: 240px;\n display: flex;\n flex-direction: column;\n overflow: hidden;\n}\n.window.focused { z-index: 200; }\n.window.unfocused .tl { background: #c4c4c4 !important; }\n\n.window-titlebar {\n height: 22px;\n background: linear-gradient(180deg, #e8e8e8 0%, #c8c8c8 50%, #b8b8b8 100%);\n border-bottom: 1px solid rgba(0,0,0,0.3);\n display: flex;\n align-items: center;\n padding: 0 8px;\n position: relative;\n cursor: move;\n}\n.window-traffic { display: flex; gap: 8px; }\n.tl {\n width: 12px; height: 12px;\n border-radius: 50%;\n border: 1px solid rgba(0,0,0,0.25);\n cursor: pointer;\n}\n.tl-red { background: radial-gradient(circle at 30% 30%, #ff8a82, #d83a30); }\n.tl-yellow { background: radial-gradient(circle at 30% 30%, #ffd76a, #d59a13); }\n.tl-green { background: radial-gradient(circle at 30% 30%, #7be08e, #1a8a2a); }\n.window-title {\n position: absolute;\n left: 0; right: 0;\n text-align: center;\n font-weight: 600;\n font-size: 12px;\n pointer-events: none;\n text-shadow: 0 1px 0 rgba(255,255,255,0.8);\n}\n\n.window-toolbar {\n background: linear-gradient(180deg, #d8d8d8 0%, #b8b8b8 100%);\n border-bottom: 1px solid #888;\n padding: 6px 10px;\n display: flex;\n gap: 8px;\n align-items: center;\n font-size: 12px;\n min-height: 28px;\n}\n.window-toolbar input[type=\"text\"],\n.window-toolbar input[type=\"search\"] {\n flex: 1;\n border: 1px solid #888;\n border-radius: 12px;\n padding: 3px 10px;\n font-size: 12px;\n background: white;\n font-family: inherit;\n}\n.window-toolbar button {\n background: linear-gradient(180deg, #fbfbfb, #d8d8d8);\n border: 1px solid #888;\n border-radius: 4px;\n padding: 3px 10px;\n font-size: 12px;\n cursor: pointer;\n font-family: inherit;\n}\n.window-toolbar button:hover { filter: brightness(0.97); }\n.window-toolbar button:active { filter: brightness(0.92); }\n\n.window-body {\n flex: 1;\n background: white;\n overflow: auto;\n position: relative;\n}\n\n/* ── Brushed metal variant for Address Book / Mail ─────────────────────── */\n.window.metal .window-titlebar,\n.window.metal .window-toolbar {\n background:\n repeating-linear-gradient(0deg, rgba(255,255,255,0.06) 0 1px, transparent 1px 2px),\n linear-gradient(180deg, #d4d4d4 0%, #b9b9b9 100%);\n}\n\n/* ── App content styles ─────────────────────────────────────────────────── */\n\n/* Mail */\n.mail-layout {\n display: grid;\n grid-template-columns: 200px 1fr;\n height: 100%;\n}\n.mail-sidebar {\n background: #d8e4f3;\n border-right: 1px solid #88a;\n padding: 8px 0;\n font-size: 12px;\n}\n.mail-folder {\n padding: 4px 12px;\n cursor: pointer;\n}\n.mail-folder.active { background: #4a7bcf; color: white; }\n.mail-list {\n display: flex;\n flex-direction: column;\n}\n.mail-item {\n padding: 8px 12px;\n border-bottom: 1px solid #ddd;\n cursor: pointer;\n font-size: 12px;\n}\n.mail-item:hover { background: #f3f7ff; }\n.mail-item.selected { background: #4a7bcf; color: white; }\n.mail-item .from { font-weight: 600; }\n.mail-item .subject { display: block; }\n.mail-item .preview { color: #666; font-size: 11px; }\n.mail-item.selected .preview { color: #cce; }\n.mail-reading {\n padding: 16px;\n font-size: 13px;\n line-height: 1.5;\n border-top: 1px solid #ccc;\n background: #fafbfc;\n white-space: pre-wrap;\n}\n.mail-compose {\n padding: 12px;\n display: flex;\n flex-direction: column;\n gap: 8px;\n height: 100%;\n}\n.mail-compose label {\n font-size: 12px;\n font-weight: 600;\n}\n.mail-compose select,\n.mail-compose input,\n.mail-compose textarea {\n border: 1px solid #888;\n padding: 4px 8px;\n font-family: inherit;\n font-size: 13px;\n border-radius: 3px;\n}\n.mail-compose textarea {\n flex: 1;\n resize: none;\n}\n.mail-compose .send-row {\n display: flex;\n justify-content: flex-end;\n}\n.mail-compose .send-btn {\n background: linear-gradient(180deg, #79b1ff, #3274d3);\n color: white;\n border: 1px solid #1a4d99;\n border-radius: 4px;\n padding: 5px 18px;\n font-weight: 600;\n cursor: pointer;\n font-family: inherit;\n}\n\n/* Address Book */\n.ab-layout {\n display: grid;\n grid-template-columns: 220px 1fr;\n height: 100%;\n}\n.ab-list {\n background: #f0f0f0;\n border-right: 1px solid #888;\n overflow-y: auto;\n}\n.ab-row {\n padding: 6px 12px;\n border-bottom: 1px solid #e0e0e0;\n cursor: pointer;\n font-size: 12px;\n}\n.ab-row:hover { background: #e3eefb; }\n.ab-row.selected { background: #4a7bcf; color: white; }\n.ab-row .role { display: block; font-size: 11px; color: #666; }\n.ab-row.selected .role { color: #cce; }\n.ab-detail {\n padding: 16px;\n font-size: 13px;\n line-height: 1.6;\n}\n.ab-detail h2 { font-size: 18px; margin-bottom: 4px; }\n.ab-detail .role { color: #666; margin-bottom: 12px; }\n.ab-detail dt { font-weight: 600; color: #444; margin-top: 8px; font-size: 11px; text-transform: uppercase; letter-spacing: 0.05em; }\n.ab-detail dd { margin-left: 0; }\n.ab-detail .id-pill {\n display: inline-block;\n background: #e8e8e8;\n border: 1px solid #aaa;\n padding: 1px 6px;\n border-radius: 3px;\n font-family: Monaco, monospace;\n font-size: 10px;\n color: #555;\n}\n\n/* iCal */\n.ical-body {\n padding: 0;\n}\n.ical-day-header {\n background: linear-gradient(180deg, #e7e7e7, #cccccc);\n padding: 6px 12px;\n font-weight: 600;\n border-bottom: 1px solid #aaa;\n font-size: 12px;\n}\n.ical-event {\n padding: 8px 12px;\n border-bottom: 1px solid #eee;\n display: flex;\n gap: 12px;\n align-items: flex-start;\n}\n.ical-event .time {\n width: 80px;\n font-weight: 600;\n font-variant-numeric: tabular-nums;\n font-size: 12px;\n color: #444;\n}\n.ical-event .body { flex: 1; font-size: 12px; }\n.ical-event .title { font-weight: 600; }\n.ical-event .loc { color: #666; font-size: 11px; }\n.ical-event .ev-notes { color: #555; font-size: 11px; margin-top: 2px; }\n.ical-event .actions { display: flex; gap: 6px; margin-top: 4px; }\n.ical-event .actions button {\n font-size: 11px;\n padding: 2px 8px;\n background: linear-gradient(180deg, #fbfbfb, #d8d8d8);\n border: 1px solid #888;\n border-radius: 3px;\n cursor: pointer;\n font-family: inherit;\n}\n.ical-event .status-badge {\n font-size: 10px;\n padding: 1px 6px;\n border-radius: 8px;\n margin-left: 6px;\n}\n.status-confirmed { background: #d2f0d4; color: #1a6e2a; }\n.status-tentative { background: #fff0c2; color: #6e5a1a; }\n.status-declined { background: #f0d2d2; color: #6e1a1a; }\n\n/* Stickies (notes app) */\n.notes-app .note-card {\n background: #fff7a0;\n margin: 12px;\n padding: 10px 14px;\n font-family: \"Marker Felt\", \"Comic Sans MS\", \"Lucida Grande\", sans-serif;\n font-size: 13px;\n color: #4a3700;\n box-shadow: 1px 2px 4px rgba(0,0,0,0.15);\n white-space: pre-wrap;\n line-height: 1.4;\n}\n.notes-app .note-title {\n font-weight: 700;\n margin-bottom: 6px;\n font-size: 12px;\n}\n.notes-app .note-tags {\n margin-top: 6px;\n font-size: 10px;\n color: #807000;\n font-style: italic;\n}\n\n/* Activity (trace) */\n.activity-app {\n font-family: Monaco, \"Andale Mono\", monospace;\n font-size: 11px;\n background: #1c1c1c;\n color: #e0e0e0;\n padding: 0;\n}\n.activity-row {\n display: flex;\n gap: 8px;\n padding: 4px 10px;\n border-bottom: 1px solid #2c2c2c;\n white-space: pre;\n}\n.activity-row .seq { color: #666; min-width: 24px; }\n.activity-row .kind { min-width: 60px; }\n.activity-row .kind-tool_call { color: #66d9ef; }\n.activity-row .kind-tool_result { color: #a6e22e; }\n.activity-row .kind-tool_error { color: #f92672; }\n.activity-row .tool-name { color: #e6db74; min-width: 130px; }\n.activity-row .payload { color: #aaa; flex: 1; overflow: hidden; text-overflow: ellipsis; }\n.activity-empty {\n padding: 24px;\n text-align: center;\n color: #888;\n font-family: \"Lucida Grande\", sans-serif;\n font-size: 12px;\n}\n\n/* End Day modal */\n.endday-app {\n padding: 24px 28px;\n font-size: 13px;\n line-height: 1.55;\n background: #f8f8f8;\n height: 100%;\n overflow: auto;\n}\n.endday-app h2 { font-size: 16px; margin-bottom: 12px; }\n.endday-app .stats {\n display: grid;\n grid-template-columns: repeat(3, 1fr);\n gap: 8px;\n margin: 12px 0 18px;\n}\n.endday-app .stat {\n background: white;\n border: 1px solid #ccc;\n border-radius: 4px;\n padding: 8px;\n text-align: center;\n}\n.endday-app .stat .n { font-size: 20px; font-weight: 700; }\n.endday-app .stat .l { font-size: 11px; color: #666; }\n.endday-app .reaction-box {\n background: white;\n border: 1px solid #aaa;\n padding: 16px;\n border-radius: 6px;\n margin-top: 12px;\n font-family: \"Hoefler Text\", Georgia, serif;\n font-size: 14px;\n color: #1a1a1a;\n line-height: 1.6;\n font-style: italic;\n}\n\n\n/* ── Threaded mail (in reading pane) ──────────────────────────────────── */\n.thread-row .thread-count { color: #999; font-weight: 400; font-size: 11px; }\n.thread-row.selected .thread-count { color: #cce; }\n.thread-header {\n font-size: 12px;\n padding: 8px 12px;\n background: #eef3fb;\n border-bottom: 1px solid #c8d4e6;\n color: #1a1a1a;\n}\n.mail-reading.placeholder, .mail-reading:not(:has(.thread-msg)) { white-space: pre-wrap; }\n.thread-msg {\n border-bottom: 1px solid #e8e8e8;\n padding: 10px 14px;\n background: white;\n}\n.thread-msg.older { background: #f7f7f7; opacity: 0.92; }\n.thread-msg-meta {\n display: flex;\n justify-content: space-between;\n font-size: 11px;\n color: #555;\n margin-bottom: 4px;\n}\n.thread-msg-from { font-weight: 600; color: #222; }\n.thread-msg-addr { font-weight: 400; color: #888; }\n.thread-msg-when { font-variant-numeric: tabular-nums; }\n.thread-msg-subject { font-weight: 600; font-size: 12px; margin-bottom: 6px; color: #333; }\n.thread-msg-body {\n white-space: pre-wrap;\n font-size: 12px;\n line-height: 1.5;\n color: #1a1a1a;\n}\n.thread-msg-actions {\n display: flex;\n gap: 6px;\n margin-top: 8px;\n}\n.reply-btn {\n background: linear-gradient(180deg, #fbfbfb, #d8d8d8);\n border: 1px solid #888;\n border-radius: 12px;\n padding: 3px 12px;\n font-size: 11px;\n cursor: pointer;\n font-family: inherit;\n color: #1a4ea0;\n}\n.reply-btn:hover { filter: brightness(0.96); }\n\n\n/* ── Phone interface ──────────────────────────────────────────────────────\n On narrow viewports the desktop metaphor gets out of the way. Status bar\n up top, brief shown as a Notes-style notification, dock recast as a 4-col\n home-screen grid, and any opened window goes fullscreen with an iOS-style\n Done button. Same JS, same routes — pure CSS layer.\n*/\n@media (max-width: 720px) {\n html, body { height: auto; overflow: auto; user-select: text; }\n body {\n /* Slightly darker so iOS status bar text reads white */\n background:\n radial-gradient(ellipse at 50% 0%, rgba(40, 90, 170, 0.6), transparent 60%),\n linear-gradient(180deg, #0e1c3a 0%, #1f4f9e 60%, #143670 100%);\n padding-top: 28px; /* room for status bar */\n min-height: 100vh;\n font-family: -apple-system, \"SF Pro Text\", \"Helvetica Neue\", \"Lucida Grande\", sans-serif;\n }\n\n /* iOS-style status bar — fixed top */\n body::before {\n content: \"9:41\";\n position: fixed;\n top: 0; left: 0;\n height: 28px;\n padding: 0 18px;\n display: flex;\n align-items: center;\n color: #fff;\n font-size: 14px;\n font-weight: 600;\n z-index: 1100;\n letter-spacing: -0.2px;\n }\n body::after {\n content: \"100%\";\n position: fixed;\n top: 0; right: 0;\n height: 28px;\n padding: 0 16px 0 30px;\n display: flex;\n align-items: center;\n color: #fff;\n font-size: 13px;\n font-weight: 600;\n z-index: 1100;\n /* Battery glyph as inline SVG — renders identically everywhere */\n background-image: url(\"data:image/svg+xml;utf8,\");\n background-repeat: no-repeat;\n background-position: 6px center;\n background-size: 22px 10px;\n }\n\n #menubar { display: none; }\n\n #desktop {\n position: static;\n padding: 16px 14px 32px;\n min-height: calc(100vh - 28px);\n }\n #windows-layer { position: static; }\n\n /* Brief sticky → Notes-style notification card */\n .sticky {\n position: relative !important;\n transform: none !important;\n width: 100%;\n max-width: 100%;\n margin: 0 0 22px 0;\n left: auto !important; top: auto !important;\n background: rgba(255, 255, 255, 0.96);\n background-image: none;\n border-radius: 14px;\n box-shadow: 0 6px 28px rgba(0, 0, 0, 0.35);\n border: none;\n color: #1a1a1a;\n overflow: hidden;\n backdrop-filter: blur(18px) saturate(160%);\n -webkit-backdrop-filter: blur(18px) saturate(160%);\n }\n .sticky-handle {\n background: linear-gradient(180deg, #f5f5f7 0%, #e5e5ea 100%);\n height: 22px;\n border-bottom: 1px solid rgba(0, 0, 0, 0.08);\n }\n .sticky-handle::before {\n content: \"MIRANDA · 9:41 AM\";\n flex: 1;\n font-size: 11px;\n font-weight: 600;\n letter-spacing: 0.6px;\n color: #6e6e73;\n padding-left: 12px;\n font-family: -apple-system, \"SF Pro Text\", sans-serif;\n }\n .sticky-close {\n background: #c7c7cc;\n color: #fff;\n width: 16px; height: 16px;\n font-size: 11px;\n }\n .sticky-body {\n padding: 14px 18px 16px;\n font-family: -apple-system, \"SF Pro Text\", \"Lucida Grande\", sans-serif;\n }\n .sticky-title {\n font-size: 11px;\n text-transform: uppercase;\n letter-spacing: 0.6px;\n color: #6e6e73;\n margin-bottom: 6px;\n }\n .sticky-text {\n font-size: 14px;\n line-height: 1.45;\n color: #1c1c1e;\n }\n .sticky-meta {\n color: #8e8e93;\n font-size: 11px;\n font-style: normal;\n }\n\n /* Dock → iOS home-screen grid */\n #dock {\n position: static;\n transform: none;\n background: transparent;\n backdrop-filter: none;\n -webkit-backdrop-filter: none;\n border: none;\n border-radius: 0;\n padding: 8px 0 0;\n margin: 0 auto;\n max-width: 360px;\n display: grid;\n grid-template-columns: repeat(4, 1fr);\n gap: 22px 8px;\n justify-items: center;\n box-shadow: none;\n }\n .dock-icon {\n width: auto;\n padding: 0;\n flex-direction: column;\n align-items: center;\n transition: transform 0.1s;\n }\n .dock-icon:hover { transform: none; }\n .dock-icon:active { transform: scale(0.92); }\n .dock-icon .icon {\n width: 60px; height: 60px;\n border-radius: 14px;\n box-shadow: 0 4px 12px rgba(0, 0, 0, 0.35), 0 0 0 0.5px rgba(255,255,255,0.1) inset;\n }\n .dock-label {\n position: static;\n opacity: 1 !important;\n background: transparent;\n color: #fff;\n font-size: 11px;\n font-weight: 500;\n margin-top: 5px;\n padding: 0;\n text-shadow: 0 1px 2px rgba(0, 0, 0, 0.6);\n }\n .dock-separator { display: none; }\n\n /* Windows: fullscreen, no chrome dragging */\n .window {\n position: fixed !important;\n top: 28px !important;\n left: 0 !important;\n right: 0 !important;\n bottom: 0 !important;\n width: 100vw !important;\n height: calc(100vh - 28px) !important;\n border-radius: 0;\n box-shadow: none;\n border: none;\n z-index: 2000 !important;\n min-width: 0;\n min-height: 0;\n }\n .window-titlebar {\n cursor: default;\n height: 44px;\n background: linear-gradient(180deg, #f9f9fb 0%, #e9e9ec 100%);\n padding: 0 12px;\n border-bottom: 1px solid rgba(0, 0, 0, 0.12);\n }\n .window-traffic {\n margin-left: auto;\n order: 2; /* push to right */\n z-index: 2;\n }\n .window-traffic .tl-yellow,\n .window-traffic .tl-green { display: none; }\n .window-traffic .tl-red {\n width: auto; height: auto;\n background: transparent !important;\n border: none;\n border-radius: 0;\n box-shadow: none;\n padding: 8px 4px;\n color: #007aff;\n font-size: 16px;\n font-weight: 500;\n line-height: 1;\n cursor: pointer;\n }\n .window-traffic .tl-red::after {\n content: \"Done\";\n color: #007aff;\n letter-spacing: 0.1px;\n }\n .window-title {\n font-size: 16px;\n font-weight: 600;\n color: #1c1c1e;\n text-shadow: none;\n }\n .window.metal .window-titlebar,\n .window.metal .window-toolbar {\n background:\n linear-gradient(180deg, #f5f5f7 0%, #e3e3e8 100%);\n }\n\n .window-toolbar {\n padding: 8px 12px;\n min-height: 36px;\n }\n .window-toolbar input[type=\"text\"],\n .window-toolbar input[type=\"search\"] {\n font-size: 16px; /* prevent iOS auto-zoom */\n padding: 6px 12px;\n }\n .window-toolbar button {\n font-size: 14px;\n padding: 6px 12px;\n }\n\n /* Mail: stack columns vertically */\n .mail-layout {\n display: flex;\n flex-direction: column;\n height: 100%;\n }\n .mail-sidebar {\n border-right: none;\n border-bottom: 1px solid #b8c4d4;\n padding: 8px 12px;\n flex: 0 0 auto;\n }\n .mail-folder { padding: 6px 0; }\n .mail-list {\n flex: 1 1 50%;\n max-height: none !important;\n overflow: auto !important;\n }\n .mail-item { padding: 12px 14px; font-size: 14px; }\n .mail-item .preview { font-size: 12px; }\n .mail-reading {\n flex: 1 1 50%;\n overflow: auto;\n font-size: 14px;\n }\n\n /* Threaded view: comfortable spacing on phone */\n .thread-msg { padding: 12px 14px; }\n .thread-msg-body { font-size: 14px; }\n .thread-msg-meta { font-size: 12px; }\n .reply-btn { font-size: 14px; padding: 6px 14px; }\n\n /* Mail compose: bigger inputs, no iOS zoom */\n .mail-compose select,\n .mail-compose input,\n .mail-compose textarea {\n font-size: 16px;\n padding: 8px 12px;\n }\n .mail-compose .send-btn { padding: 10px 22px; font-size: 15px; }\n\n /* Address Book: stack */\n .ab-layout { display: flex; flex-direction: column; height: 100%; }\n .ab-list {\n border-right: none;\n border-bottom: 1px solid #aaa;\n flex: 0 0 45%;\n max-height: 45vh;\n }\n .ab-row { padding: 10px 14px; font-size: 14px; }\n .ab-detail { font-size: 14px; flex: 1; padding: 14px; }\n\n /* iCal: bigger time column gone, stack body under */\n .ical-event { padding: 10px 14px; }\n .ical-event .time { width: 64px; font-size: 13px; }\n .ical-event .body { font-size: 13px; }\n .ical-event .actions button { padding: 6px 12px; font-size: 13px; }\n\n /* Stickies: bigger cards */\n .notes-app .note-card { font-size: 14px; }\n\n /* Activity: legible on small screens */\n .activity-row { font-size: 11px; }\n\n /* End-of-day report */\n .endday-app { padding: 18px 16px 32px; }\n .endday-app .stats { grid-template-columns: repeat(3, 1fr); gap: 6px; }\n .endday-app .stat .n { font-size: 18px; }\n}\n\n\n\n/* ── Errands icon (driver/restaurant) — teal Aqua ──────────────────────── */\n.icon-errands {\n background:\n radial-gradient(ellipse at 30% 25%, rgba(255,255,255,0.55), transparent 50%),\n linear-gradient(180deg, #4dc09a 0%, #1f7a5e 100%);\n display: flex;\n align-items: center;\n justify-content: center;\n}\n.icon-errands svg {\n filter: drop-shadow(0 1px 1px rgba(0,0,0,0.25));\n}\n\n/* ── Errands app forms (Driver / Restaurant tabs) ──────────────────────── */\n.errands-app { padding: 0; background: linear-gradient(180deg, #f6f6f6 0%, #ececec 100%); }\n.errands-form {\n padding: 16px 18px 20px;\n display: flex;\n flex-direction: column;\n gap: 10px;\n height: 100%;\n overflow: auto;\n}\n.errands-blurb {\n font-size: 12px;\n color: #555;\n line-height: 1.4;\n background: #fffbe9;\n border: 1px solid #e6d99a;\n border-radius: 4px;\n padding: 8px 10px;\n margin-bottom: 4px;\n}\n.errands-form label {\n display: flex;\n flex-direction: column;\n font-size: 11px;\n font-weight: 600;\n color: #444;\n text-transform: uppercase;\n letter-spacing: 0.05em;\n gap: 3px;\n}\n.errands-form input[type=\"text\"],\n.errands-form input[type=\"datetime-local\"],\n.errands-form input[type=\"number\"],\n.errands-form textarea {\n border: 1px solid #888;\n border-radius: 4px;\n padding: 5px 8px;\n font-family: inherit;\n font-size: 13px;\n font-weight: 400;\n text-transform: none;\n letter-spacing: normal;\n color: #1a1a1a;\n}\n.errands-form textarea { resize: vertical; }\n.errands-form .form-row {\n display: flex;\n align-items: center;\n gap: 12px;\n margin-top: 4px;\n}\n.errands-form .primary-btn {\n background: linear-gradient(180deg, #79b1ff, #3274d3);\n color: white;\n border: 1px solid #1a4d99;\n border-radius: 4px;\n padding: 6px 18px;\n font-weight: 600;\n cursor: pointer;\n font-family: inherit;\n font-size: 13px;\n}\n.errands-form .primary-btn:hover { filter: brightness(1.05); }\n.errands-form .primary-btn:active { filter: brightness(0.95); }\n.errands-form .form-status {\n font-size: 12px;\n color: #666;\n font-weight: 400;\n}\n\n/* End-Day reaction box — bigger now that it has structure */\n.endday-app .reaction-box .reaction-icon {\n font-size: 48px;\n text-align: center;\n margin-bottom: 4px;\n line-height: 1;\n}\n.endday-app .reaction-box .reaction-type {\n text-align: center;\n font-size: 11px;\n text-transform: uppercase;\n letter-spacing: 1px;\n color: #6e6e73;\n margin-bottom: 12px;\n font-style: normal;\n}\n.endday-app .reaction-box .reaction-line {\n font-family: \"Hoefler Text\", Georgia, serif;\n font-size: 15px;\n line-height: 1.5;\n color: #1a1a1a;\n font-style: italic;\n margin: 4px 0;\n padding-left: 12px;\n border-left: 2px solid #c8c8c8;\n}\n.endday-app .reaction-box .reaction-summary {\n margin-top: 10px;\n padding-top: 8px;\n border-top: 1px solid rgba(0,0,0,0.1);\n font-size: 11px;\n color: #6e6e73;\n font-style: normal;\n}\n\n\n/* ── iCal → Errands cross-app hint ─────────────────────────────────────── */\n.ical-hint {\n margin-top: 6px;\n padding: 6px 10px;\n background: #fffbe9;\n border: 1px solid #e6d99a;\n border-radius: 4px;\n font-size: 11px;\n display: flex;\n align-items: center;\n gap: 8px;\n color: #6e5a1a;\n}\n.ical-hint .hint-btn {\n margin-left: auto;\n background: linear-gradient(180deg, #79b1ff, #3274d3);\n color: white;\n border: 1px solid #1a4d99;\n border-radius: 3px;\n padding: 3px 10px;\n font-size: 11px;\n font-weight: 600;\n cursor: pointer;\n font-family: inherit;\n}\n.ical-hint .hint-btn:hover { filter: brightness(1.05); }\n.ical-hint .hint-btn:active { filter: brightness(0.92); }\n@media (max-width: 720px) {\n .ical-hint { font-size: 13px; padding: 8px 12px; }\n .ical-hint .hint-btn { padding: 6px 12px; font-size: 13px; }\n}\n",[],[204,318],["D",1777168270439],[222,223,224,225,226,227,228,229],[204],["D",1777167345173],["D",1789525843977],{"_98":97,"_106":102},{"_98":97,"_99":100,"_101":102,"_18":53,"_103":104,"_105":24},{"_13":87,"_31":88,"_41":40,"_89":42,"_90":-5,"_91":-5,"_92":-5,"_93":-5,"_94":-5,"_95":43,"_96":97},"ctx",{"_80":-5,"_126":366,"_367":368,"_369":370,"_371":24,"_372":24,"_373":24,"_374":57,"_375":-5,"_376":-7,"_377":-7,"_378":61,"_379":24,"_380":24,"_381":-5,"_382":24,"_383":24,"_384":24,"_385":-5},"ENV",{"_337":338,"_339":340,"_341":88,"_342":343,"_344":345,"_346":347,"_348":349,"_350":351,"_352":353,"_354":355,"_356":357,"_358":359,"_360":361,"_362":363,"_364":365},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

Tools (10)

Prompt: miranda_brief

How the judge sees what you did

Reaction tiers

…

The brief Miranda just dictated

1. Name your run

2. Connect your AI

Claude Code (CLI)

Claude Desktop

Cursor

Cline (VS Code)

Goose

Raw HTTP / curl

Share this run

Start a new run

${escape(c.name)}

End-of-day report

Prompt: `miranda_brief`

Diff · v137