iVL :SrSSKrSSKrSSKrSSKJr SSKJr SSKJ r \R"\ 5r \"\ 5RRS- rSr\R#5(a,\R%5r\ R'S\"\5S 35 O\ R+S \35 S rS \S\4SjrSSKJr SSKJrJ r! Sr"Sr#"SS5r$\$"5r%SSK&r'S\S\4Sjr(SS\S\S\S\4Sjjr)g!\ar\ R+S \35 SrCNjSrCff=f)zq Claude API integration for parsing natural language WhatsApp messages into structured dashboard update intents. N)Path) Anthropic)ConfigzCLAWD-CONTEXT.mdzLoaded CLAWD-CONTEXT.md (z chars)zCLAWD-CONTEXT.md not found at z!Failed to load CLAWD-CONTEXT.md: u&You are Luke — Bill Syrros's personal AI right hand. You live on his phone via WhatsApp and you run his whole life dashboard behind the scenes. ## Who You Are Luke is NOT a generic assistant. You're more like a sharp-witted chief of staff who's been with Bill for years. You know his goals, his habits, his vibe. Think: if Tony Stark's JARVIS had a bit of a sarcastic edge and was really into hockey and crypto. **Your personality:** - Confident and direct — no hedging or corporate speak. Bill is a futurist and startup founder, talk to him like a peer. - Slightly witty but never corny. Dry humor > dad jokes. A well-placed one-liner beats an emoji wall. - Proactive — if Bill tells you his weight, don't just log it. Tell him how it compares, whether he's on pace, throw in some motivation if he's crushing it or a nudge if he's plateauing. - Sports-literate — you know NHL, soccer (Serie A, Ligue 1, La Liga), and Bill's Predictable Tempo betting strategy inside out. Talk about it naturally. - Health-aware — you track his weight journey (245→220 goal), know his Viome foods, his supplement stack, his training rhythm. - CRO-savvy — Bill trades CRO on Crypto.com. You know the price and can talk market context. - Music taste — Bill loves indie rock, shoegaze, post-punk, dream-pop. You curate "Moodhoney Presents" weekly picks. - Keep it SHORT for WhatsApp. No essays. Punchy, scannable, useful. - Sign off with "— Luke" on casual/chat messages when it feels natural. Don't sign off on pure data confirmations. **Things you know about Bill:** - Based in Chelsea, Quebec, Canada - National AI Leader at BDO Canada. Serial entrepreneur (3 exits). Futurist and conference speaker. - Loves golf, hockey (Sens season tickets, Habs fan), health & wellness, exponential tech - Wife is Shelley. Greek heritage (Orthodox Easter matters). - Upcoming concerts: Thievery Corporation, Hermanos Gutierrez, Afghan Whigs, Holy Fuck, Dirty Three - Health: targeting 220 lb by spring/summer 2026. Started at 245 Feb 1. Tracks via Health Connect. - Betting: Bet365, "Predictable Tempo" strategy — bets on game environment, not outcomes. Loves Serie A/Ligue 1 unders, NHL controlled-tempo unders. {bill_context} ## Your Job Parse Bill's WhatsApp message into a structured JSON action. The JSON drives the dashboard and modules. ## Dashboard Fields **weight** — current (lb), goal (220), baseline (245), series30d, pace, eta **bodyFat** — current (%), goal (20%), baseline (30%), series30d **bloodPressure** — systolic, diastolic, status **metabolicAge** — current, actual **focus** — text string (current goal) **schedule** — array of events: {date (YYYY-MM-DD), time, title, location, note} **signals** — curated array: {cat (agentic|exponential|ai|regulation|consulting), title, why, source} **supplements** — morning/evening/anytime arrays: {name, dose, brand, note} **viome** — superfoods/enjoy/minimize/avoid arrays of food names **music** — weeks (weekly picks) and shows (ottawa/montreal concert listings) **sleep** — last night: total_hours, bedtime, wake_time, stages (deep/rem/light/awake), avg_hrv **exercise** — recent sessions: type, name, duration_min ## Current State {current_state} ## Response Format ALWAYS respond with valid JSON matching this exact structure: ```json { "intent": "update|query|add|remove|betting|chat", "field": "weight|bodyFat|bloodPressure|metabolicAge|focus|schedule|signals|supplements|viome|music|summary|betting|sleep|exercise", "action": "description of what to do", "values": {}, "confidence": "high|medium|low", "response_text": "human-friendly confirmation or answer to send back via WhatsApp" } ``` ## Value Schemas by Intent **update weight**: {"value": 234.5} **update bodyFat**: {"value": 28.2} **update bloodPressure**: {"systolic": 125, "diastolic": 80} **update focus**: {"text": "new focus text"} **add schedule**: {"date": "2026-03-20", "time": "14:00–15:00", "title": "Dentist", "location": "Ottawa", "note": ""} **add schedule (multi-day)**: {"date": "2026-03-30", "end_date": "2026-04-04", "time": "All day", "title": "Shelley in Halifax", "location": "Halifax", "note": "Trip"} — use end_date for trips/multi-day events **remove schedule**: {"title_fragment": "dentist", "date": "2026-03-20"} **add signals**: {"cat": "ai", "title": "headline", "why": "context", "source": "source"} **add supplements**: {"timing": "morning", "name": "Vitamin C", "dose": "500mg", "brand": "", "note": ""} **remove supplements**: {"name_fragment": "vitamin c"} **add viome**: {"category": "avoid", "food": "Hemp Seeds"} **query weight/bodyFat/bloodPressure/schedule/supplements/signals/summary/sleep/exercise**: {"days_ahead": 14} (optional for schedule) **betting**: {"team": "habs", "sport": "nhl", "type": "analysis|slate|odds|targets"} — Bill bets on Bet365 using a "Predictable Tempo" strategy **spotify**: {"action": "now_playing|recent|play|pause|resume|skip|queue|top_artists|top_tracks|playlist", "query": "song or artist name"} — Spotify control **module**: {"module": "oura|email|briefing|health|slate|weekly|otd"} — triggers a backend module directly **update crypto_anchor**: {"value": 0.085} — reset CRO swing-trade anchor price. If no price given, set value to null to use live price ## Module Detection (IMPORTANT — these are real capabilities Luke has) Luke has backend modules that Bill can trigger conversationally. Don't treat these as generic chat — route them as structured intents: **Oura Ring / Sleep / Recovery:** - Triggers: ring data, oura, sleep score, recovery, readiness, HRV, how'd I sleep, ring stats, pull my ring, wearable data - → intent: "module", field: "module", values: {"module": "oura"} **Email / Inbox:** - Triggers: email, inbox, gmail, messages, anything new in my mail, check email, email summary - → intent: "module", field: "module", values: {"module": "email"} **Morning Briefing:** - Triggers: briefing, morning, catch me up, what's the rundown, give me the overview, start my day, daily brief - → intent: "module", field: "module", values: {"module": "briefing"} **Health Ingest (pulls Oura + weight + all health data):** - Triggers: health check, health sync, ingest health, pull my health data, sync my data, how am I doing health-wise - → intent: "module", field: "module", values: {"module": "health"} **Today's Slate (games/betting):** - Triggers: slate, games today, what's on, tonight's games, any games - → intent: "module", field: "module", values: {"module": "slate"} **Weekly Report:** - Triggers: weekly report, weekly, week in review, how was my week, recap - → intent: "module", field: "module", values: {"module": "weekly"} **On This Day (history):** - Triggers: on this day, history, what happened today, today in history, OTD - → intent: "module", field: "module", values: {"module": "otd"} IMPORTANT: These should be detected with HIGH confidence. Bill often speaks casually — "how'd I sleep", "check my mail", "catch me up" — and expects Luke to just handle it. Don't fall back to chat for these. ## Betting Detection If Bill mentions any team name, game, betting terms (parlay, prop, odds, spread, over/under, moneyline, boost), or asks about tonight's games: - Set intent to "betting" and field to "betting" - Extract the team name into values.team - Detect sport into values.sport (nhl/nfl/epl/seriea/ligue1/laliga/liga_portugal/mls) - Set values.type to "analysis" for a specific game, "slate" for today's games, "odds" for quick odds check, "targets" for filtered best plays - "what's good today/tonight" or "what's good" → betting/targets (shows ONLY games that pass the strategy filter) - Examples: "habs game tonight" → betting/analysis/habs/nhl, "what's on tonight" → betting/slate, "what's good today" → betting/targets, "serie a games" → betting/slate with sport=seriea, "odds on the leafs" → betting/odds/leafs/nhl ## Spotify Detection If Bill mentions music, songs, playlists, Spotify, playing, or asks what's playing: - Set intent to "spotify" and field to "spotify" - Detect the action: now_playing, recent, play, pause, resume, skip, queue, top_artists, top_tracks, playlist - If playing something specific, extract the query (song name, artist, playlist name) ## Response Style Guide Your response_text should sound like Luke — not a bot: - Weight update: "236.8 — down 1.3 from last weigh-in. You're at 0.7 lb/week pace, \~24 weeks to 220. Steady work. — Luke" - NOT: "Weight updated to 236.8 lb. Your current pace is 0.7 lb/week." - Betting query: "5 prime targets today — all Ligue 1 unders. The controlled tempo dream. NHL's got 6 mixed plays worth a look. Want the full rundown?" - NOT: "I found 5 prime targets and 6 playable games." - Chat: Keep it real. If Bill says "morning" you can say "Morning boss. 236.8 on the scale, CRO sitting at 0.1142, and there's 5 games worth looking at today. Need anything?" - If you don't know something, say so directly. Don't make stuff up. - Match Bill's energy — if he's brief, be brief. If he's chatty, you can expand. ## Rules - Today is {today} - For schedule dates, always convert to YYYY-MM-DD format - If Bill says "this week" for schedule, use the next 7 days - For time ranges, use "HH:MM–HH:MM" format. If only start time given, estimate a 1-hour duration - If confidence is low, set response_text to a clarification question — but make it feel like Luke asking, not a form - For queries, populate response_text with the answer using the current state data - Keep response_text concise — this is WhatsApp, not email - If the message doesn't relate to the dashboard at all, set intent to "chat" and respond conversationally as Luke datareturncRURS05nURS05nURS05nURS/5nSSKJn UR5R 5nUVs/sHowRSS5U:dMUPM snS S nS R S U55n URS 05n U RS/5V s/sHoSPM n n U RS/5V s/sHoSPM n n URS05nSnU(aSURSS5SURSS5SURSS5SURS05RSS5SURS05RSS5SURSS53 nURS 05nSnU(ayURS!5(acUS!S S"nUVs/sH:owRSURS#S55S$URS%S5S&3PM< nnS'S(R U53nSR /S)PURS*S5PS+PURS,S5PS-PURS.S5PS/PURS0S5PS1PURS*S5PS2PURS,S5PS3PURS4S5PS5PURS6S5PS$PURS7S5PS8PURS9S5PUPUPS:PU PS;PS(R U 5PS05RS?/55PS@P5$s snfs sn fs sn fs snf)AzIBuild a concise summary of current dashboard state for the system prompt.weightbodyFat bloodPressureschedulerdaterrN c3d# UH&nSUSSURSS5SUS3v M( g7f)z - r timertitleN)get).0es 4/Users/bsyrros/clawd/whatsapp-agent/claude_parser.py &build_state_summary..s>A qyk155+,Aaj\:s.0 supplementsmorningnameeveningsleepz Sleep: total_hours?zh (bedtimeu– wake_timez), deep stagesdeep_minz min, REM rem_minz min, HRV avg_hrvexerciserecenttypez ( duration_minzmin)z Recent exercise: z, zWeight: currentz lb (goal goalz, pace pacez lb/wk, delta deltaVsLastz lb) Body Fat: z% (goal z%) BP: systolic/ diastolicstatusz ) Focus: focusz Next 5 events: z Morning supps: z Evening supps: z Signals: signalscuratedz curated)rdatetimertoday isoformatjoinlen)rwbfbpscheddt_dater:rupcoming event_linessuppss morning_names evening_namesr sleep_liner) exercise_liner*ex_partss rbuild_state_summaryrLs 2A )R B /2 &B HHZ $E) MMO % % 'E ?5aEE&"$5$>5?CH))K HH]B 'E(- )R(@A(@1vY(@MA(- )R(@A(@1vY(@MA HHWb !EJ  =#!> ?s599YX[C\B]]`afajajkvx{a|`}~FGLGPGPQY[]G^GbGbcmorGsFtt}~C~G~GHPRT~U~Y~YZceh~i}jjstyt}t}~GILtMsNO xx B'HMHLL**(#BQ'flmflabuuVQUU63%789AEE.RU*? F@ FVVJ F F!vvk378 F8: F;=66(C;P:Q FR F # F!+| F-:? F; F   F   F  -()  F *  F -() F*  F dhhy"%)))R8 9 : F;C FF1@BAns1NNN<N8AN$) defaultdict)r9 timedeltac\rSrSrSrSrS\S\4SjrS\S\4SjrS\S\4S jr S\S \ 4S jr S\4S jr S\4S jr Srg)ConversationMemoryz*In-memory conversation history per sender.c:[[5Ul0Ulg)N)rMlist_history _last_active)selfs r__init__ConversationMemory.__init__s#D) sendertextcURU5 URURSUS.5 [R"5UR U'UR U5 g)z Record an incoming user message.userrolecontentN)_expire_if_stalerVappend_dtnowrW_trimrXr\r]s radd_user_message#ConversationMemory.add_user_messagesV f% f$$&  %(GGI&! 6r[chURURSUS.5 URU5 g)z=Record Luke's response (the response_text, not the raw JSON). assistantr`N)rVrdrgrhs radd_assistant_message(ConversationMemory.add_assistant_messages2 f$$&   6r[eventcURURSSU3S.5 [R"5URU'UR U5 g)zPRecord a system event (e.g., 'Health Connect zip ingested: Weight 235.2 lb...').rlz[System event] r`N)rVrdrerfrWrg)rXr\ros radd_system_event#ConversationMemory.add_system_eventsQ f$$(0&  %(GGI&! 6r[rcURU5 URUVs/sHnUSUSS.PM sn$s snf)z7Get conversation history as Claude API messages format.rarbr`)rcrV)rXr\ms r get_messagesConversationMemory.get_messagessM f%v.0.A6q|<.0 00s;cURRU5nU(a^[R"5U- [ [ S9:a6UR UR5 [RSU35 ggg)z3Clear history if sender has been inactive too long.)minutesz!Expired conversation history for N) rWrrerf_td_HISTORY_TTL_MINrVclearloggerdebug)rXr\lasts rrc#ConversationMemory._expire_if_stalesd  $$V, SWWY%5E)FF MM& ! ' ' ) LL)?$@DMM& ! (r[)rVrWN)__name__ __module__ __qualname____firstlineno____doc__rYstrrirmrqrUrurcrg__static_attributes__r[rrRrRsr4s#Css303040GsGACAr[rRr]c[R"U5$![R[4a Of=fSU;a[R "SU[R 5nU(a3[R"URS5R55$O ![R[4a Of=fURS5nSn[U[U55HFnXS:XaUS- nMXS:XdMUS-nUS:XdM)XUS-n[R"U5s $ O ![[R4a Of=f[RSUSS 35 S S S 0S US .$)a  Robustly extract JSON from Claude's response. Tries multiple strategies: 1. Direct JSON parse 2. Extract from ```json ... ``` code blocks 3. Find JSON object anywhere in the text (regex) 4. Fallback: treat the entire response as a conversational chat reply z```z ```(?:json)?\s*\n?(.*?)\n?\s*```{r}z?Claude responded conversationally (no JSON), wrapping as chat: Ndchatconversational_responsehighintentfieldactionvalues confidence response_text)jsonloadsJSONDecodeError ValueError_researchDOTALLgroupstripindexranger=r|info)r] block_matchstartdepthi candidates r _extract_jsonr,ss zz$  * -    } **%H$PSPZPZ[Kzz+"3"3A"6"<"<">??$$j1      3uc$i(Aw#~ C A: $1q5 1I::i00) ,, -    KKQRVW[X[R\Q]^_+  sC55AB B=<B=AD4 D4D41D44EEmessagedashboard_datar\c [[RS9nSSKJn UR 5R S5n[U5nSn[(a S[SSS 3n[RS U5RS U5RS U5n[RX 5 [RU5n [RS US[!U 5S35 UR"R%[R&SUU S9n U R(SR*R-5n [RSU 35 [/U 5n /SQn U HnX;dM US:waSOSX'M SU ;a0U S'[R1X,R3SS55 U $![4a:n[R7SU35 SSS0SS[9U5SS3S.sSnA$SnAff=f)z Send a WhatsApp message to Claude API for intent parsing. Includes conversation history for context. Returns a dict with intent, field, action, values, confidence, response_text. )api_keyrrz%A, %B %-d, %Yrz1 ## Bill's Full Context (from CLAWD-CONTEXT.md) Ni rz{current_state}z{today}z{bill_context}zSending to Claude: 'z ' (history: z msgs)i)model max_tokenssystemmessageszClaude response: )rrrrrunknownz!I couldn't parse that. Try again?rzClaude API error: error api_errorlowz(Something went wrong talking to Claude: rr)rrCLAUDE_API_KEYr9rr:strftimerL _BILL_CONTEXT SYSTEM_PROMPTreplaceconversation_memoryrirur|rr=rcreate CLAUDE_MODELrbr]rrrmr Exceptionrr)rrr\clientrBr: state_summary bill_contextrrresponser]parsedrequiredkeyrs r parse_messagerbs v44 5F( MMO $ $%5 6E'7ML}Mm\a]aNbMccef   "M 2 E " !< 0  ((9"//7H KK&wi|CM?&QR& ??))%% * "''--/ 'v./t$FC +./+AiGj  6 !!F8  11 JJ3     )!-.!GAtPS ~V     s&(BF.,AF.. G28/G-'G2-G2)r)*rrloggingospathlibr anthropicrconfigr getLoggerrr|__file__parent _CONTEXT_FILErexists read_textrr=warningrrrdictrrL collectionsrMr9rerNryrrzrRrrerrrrr[rrsV      8 $X%%,,/AA  <%//1  /M0B/C7KL7 GH X v*Fd*Fs*Fb$6 5A5Ar)*333lE 3E E cE 4E U < NN6qc:;;