@murphai/murph 1.2.0 → 1.2.4

package/node_modules/@murphai/assistant-engine/skills/murph-onboarding/SKILL.md

@@ -1,179 +1,413 @@
 ---
 name: murph-onboarding
-description: Use only when the current prompt marks first-run Murph onboarding as open, including the welcome, name and health-context collection, wearable/app checkpoint, first experiment setup, and onboarding completion.
+description: Use only when direct first-run Murph onboarding is open, including the private welcome, aspiration anchors, progressive foundation-context checkpoints, the contextual return to an open thread, completion, or an overall decline.
 ---
 
 # Murph onboarding
 
 ## Goal
 
-Introduce the user to Murph, understand what they care about health-wise, complete a wearable/app checkpoint before first experiment setup, help them start sharing context over time, and guide them toward setting up a first bounded experiment.
+Establish Murph as the user's private personal health assistant, briefly learn
+what they most want from their health, save one or two aspirations as open
+threads, gather enough foundation context for later help to fit, then return to
+an open thread and choose the first step together.
 
-Expect roughly 9-10 short assistant messages after the welcome across normal onboarding turns. Do not compress the whole orientation into one "send me things" reply or one bundled setup questionnaire.
+The first health topic is an anchor, not a launch button. A user answering
+Murph's discovery question has shared context; they have not asked for a plan,
+diagnosis, or intervention. Only an actual immediate request or safety need
+should start problem-solving before the foundation is understood.
 
-## Resuming when earlier conversations are not visible
+Experiments are one optional primitive. Do not turn onboarding into an upfront
+profile questionnaire, capability tour, wearable funnel, or experiment funnel.
+Do not create a second context-collection lifecycle or automation.
 
-Onboarding open means completion was never recorded; it does not mean this is the user's first conversation. Use the visible conversation as the first source of truth for onboarding position. If the exact welcome is visible in this same thread and the user's latest message is a short acceptance such as "yes", "yeah", "yea", "ready", or similar, treat this as normal first-run continuation: onboarding is incomplete, no broad vault resume check is needed, and the next step is the name plus optional age/gender question unless the visible thread already answers it.
+## Resume without repeating
 
-Earlier conversations may have already covered some or all steps without that history being visible in this thread. When onboarding is open but the visible thread does not show the welcome or prior onboarding steps, make one bounded vault resume check before asking anything: run `vault-cli assistant onboarding resume-context --format json`. Treat saved facts from that snapshot as already-answered steps and resume from the first genuinely unresolved step. Never re-send the welcome when the vault shows setup context from earlier conversations, and never re-ask for facts the vault already contains, including the user's name. If saved context already satisfies the completion criteria — including a resolved first experiment setup — mark onboarding complete with `--reason user_answered` instead of re-running the flow. Do not fan this resume check out into separate `memory show`, `goal list`, `regimen list`, `supplement list`, `condition list`, `allergy list`, `experiment list`, or `device account list` commands unless the resume-context command is unavailable or returns an error for the specific surface you still need.
+Onboarding `open` means completion was never recorded. It does not prove this
+is the user's first conversation.
 
-## Outcomes
+Use the visible conversation first. If the welcome is visible and the latest
+message is a short acceptance such as “yes,” “yeah,” “ready,” or similar,
+continue naturally with minimal identity unless the conversation already
+answers it.
 
-- User knows what Murph is: a health context layer that tracks meals, workouts, supplements, labs, symptoms, sleep, energy, recovery, wearable signals, and questions over time, then summarizes patterns and tradeoffs.
-- User has completed a wearable/app checkpoint: Murph has recognized a connected source, sent a supported connection link when the user named a supported provider, asked which supported provider they use when they asked to connect a generic wearable, or confirmed they want to continue without one. A wearable is optional, but this checkpoint is not.
-- User has shared their health goals or interests, or declined.
-- User has been asked for lightweight setup context over separate turns: name plus optional age/gender first, then a separate health-context question, then the wearable/app checkpoint, then movement/training context, then current health protocols or experiments they are trying, then current supplements with brand or product names plus roughly how long they have taken them or since when, then one open medical-context question covering prescription/OTC medications, diagnosed conditions, allergies or intolerances, and pregnancy or nursing, then recent blood tests or lab panels. Each prompt makes clear they can skip anything they do not want to share.
-- User has been asked separately whether they have recent blood tests or lab panels, such as Function Health or doctor-ordered labs, and knows they can skip this for now if they do not have results handy, then send PDF lab documents later if they want Murph to use them.
-- User has been asked once about safety-relevant medical context — medications, diagnosed conditions, allergies or intolerances, pregnancy or nursing — as a single optional open question.
-- Useful setup answers are persisted to their best-fit canonical surface as the user shares them. Use a structured record whenever a typed `vault-cli` surface exists for the fact (goals, regimens, supplements, conditions, allergies, negative allergy assertions, experiments, and similar); fall back to Identity or Context memory only for facts with no structured home, such as preferred name, demographics, lifestyle context, interests, or pregnancy/nursing status. Do not dump structured items into freeform memory.
-- User understands the product loop: run one lightweight, bounded experiment at a time, then review what changed and decide what is worth keeping.
-- User has normally been offered a small choice set of two or three lightweight, bounded first-experiment options grounded in their goals, collected context, and a Health Commons protocol discovery pass, not a single recommendation or from-scratch guesses. Each option is framed around a user-valued outcome rather than an adherence mechanism and names the intervention, a credible timeframe, and the evidence Murph would use to judge progress. Prefer existing Health Commons protocols when they fit. The assistant may include a custom option only when it fits the user's goals, available data, safety, or life logistics better than the discovered protocol options, and may label one option as the best-fit default with a brief reason, but the user should still be able to choose among credible options or defer. Never add a weak option solely to fill the menu.
-- User has resolved first experiment setup: an active first experiment was created, the user explicitly deferred or declined, or setup is blocked by a specific safety/logistics issue. Onboarding is not complete until this is resolved.
+If no welcome or prior onboarding is visible, run one bounded resume check:
 
-## Saving answers
-
-Persist useful answers as they arrive, not at the end. Each fact goes to its best-fit canonical `vault-cli` surface: structured records when a typed surface exists (goals, regimens including medications and supplements, conditions, allergies, negative allergy assertions, experiments, and similar), and Identity or Context memory only for facts with no structured home — preferred name, demographics, lifestyle context, interests, pregnancy or nursing status. Save dated facts with the current prompt's local date rather than inferring a birthday or onset.
-
-Treat "NKDA", "no known drug allergies", "NKFA", "no known food allergies", and broad "no known allergies" as negative clinical assertions, not allergy records or notes. Save them with `vault-cli event import-json --input @file.json` using `kind: "clinical_assertion"`, `occurredAt` for the source/save timestamp, assertion `no_known_drug_allergies` / `no_known_food_allergies` / `no_known_allergies`, `assertedOn` from the source date or the current local date, and `sourceLabel` when available.
-
-Do not dump structured items into freeform memory, do not invent details the user did not give (dose, severity, onset, brand), and do not turn every vague interest into a goal — save soft "curious about sleep" mentions as Context memory unless the user framed a concrete goal.
-
-## First-experiment outcome quality bar
-
-A first experiment earns its place only when the result would be genuinely useful to the user. The visible choice should be the outcome they care about, not the compliance tactic Murph may use underneath it.
-
-Every option must pass all of these checks:
-
-- **User value:** tie the result to something the user explicitly wants, such as better sleep, greater strength, visible physique progress, improved aerobic performance, less pain, or a concrete behavior they genuinely want to establish. Do not infer that recovery, RHR, HRV, steps, or any other available signal matters to them just because Murph can measure it.
-- **Outcome specificity:** name a change the user could recognize or care about. "Be healthier," "improve recovery," "feel better," and "build consistency" are too vague unless the user's language and the measurement plan make them concrete.
-- **Evidence fit:** identify one primary outcome and a small number of supporting signals that can actually speak to it. Prefer direct, interpretable evidence such as a repeatable performance benchmark, sleep duration plus morning energy, a symptom or function score, a planned lab comparison, or optional private standardized photos or measurements only when the user explicitly wants a visual outcome, with privacy-safe handling.
-- **Timeframe fit:** use a duration long enough for the promised result to plausibly move and be observed. If the selected timeframe can only test feasibility or adherence, say that plainly, lengthen it, or choose a nearer-term outcome the user still values. Never sell a feasibility block as proof that health or fitness improved.
-- **Meaningful signal:** the planned evidence should be capable of distinguishing a worthwhile change from normal variation. Where the protocol supports it, define what magnitude or direction would be meaningful enough to act on; do not market a tiny fluctuation, unstable daily wearable shift, or measurement noise as success.
-- **Decision value:** the result should help the user decide whether to keep, change, extend, or stop something. An end-of-run adherence percentage by itself is not a valuable outcome.
-- **Burden fit:** the expected value of the answer should justify the effort, logging, risk, and disruption required from the user.
-
-Adherence, floors, tiny versions, fallbacks, and reminders are plan mechanics. They can support or explain the experiment, but do not headline an option unless the user explicitly said that making the behavior consistent is itself the desired result. When consistency is the desired result, define a concrete end state, such as establishing three runs per week for four weeks, rather than merely testing whether reminders work.
-
-Reject or rewrite a candidate when the only measurable result is whether the user did it, when it relies on a convenient wearable metric the user did not care about, when the timeframe is too short for the promised change, or when the likely result would not alter the user's next decision. RHR, HRV, recovery, steps, and similar signals may be primary evidence when the user values them and the protocol and timeframe support that claim; otherwise use them only as supporting context.
-
-## Required interaction affordances
-
-Some onboarding questions include easier input or delivery options. These are part of the one lightweight question, not extra questions. Do not drop them for brevity.
+```text
+vault-cli assistant onboarding resume-context --format json
+```
 
-- Movement/training: ask one natural question, include the compact examples list, and end the visible message with exactly: "Feel free to send me a voice memo."
-- Current protocols/experiments: this is the default delight moment for one generated onboarding voice memo. When `murph.generate_voice_memo` is available and the user has not asked to avoid voice messages, attach the current protocol/experiment question as a short voice memo and leave the final response text empty. Do not send a separate companion text just to explain the voice memo. This is an explicit product-flow voice preference; do not require the user to ask for voice separately. If voice memo generation is unavailable, fails, or the user prefers text, ask one natural text question. Keep examples compact in either mode.
-- Supplements: mention that they can send a photo of supplement bottles or labels if that is easier.
+Treat every useful saved fact in the snapshot as known evidence for the open
+health threads and foundation checkpoints. Never re-ask it.
+Missing evidence is unresolved unless the visible conversation shows that the
+user said it was not relevant or explicitly skipped it. A request to continue
+later is a deferral, not a completed checkpoint.
+
+Preserve forward progress when older wording has fallen out of visible history.
+If the visible conversation shows a foundation question or answer after an
+aspiration, treat the reflect-and-park transition as already done. If that
+ordering is no longer visible but a concrete aspiration is saved, foundation
+context exists, and the current exchange is clearly resuming foundation or the
+contextual return, continue from the next unresolved step instead of replaying
+the park. Existing records alone do not prove that onboarding began.
+
+Do not fan the snapshot out into separate memory, goal, regimen, supplement,
+condition, allergy, experiment, or device commands. Make one targeted owning
+read only when the checkpoint needed now is omitted, truncated, or errored in
+the snapshot. In particular, use `vault-cli memory show --format json` when
+relevant memory evidence is truncated and `vault-cli blood-test list --format
+json` before asking the lab checkpoint when recent lab evidence is otherwise
+unknown. If visible and saved evidence satisfies every completion rule below,
+mark onboarding complete instead of asking another question.
+
+## The immediate need wins
+
+If the user arrives with a health question, decision, symptom, file, image,
+lab, meal, workout, data point, connection request, logging request, task, or
+safety-sensitive need, handle it first. That request may answer one or more
+onboarding checkpoints, but it does not complete onboarding by itself.
+
+Distinguish an actual request from an answer to Murph's own discovery question.
+For example, “I want to get stronger” after Murph asks what the user wants from
+their health is an aspiration to save and park. “Can you make me a strength
+plan?” is an immediate request to handle. When intent is unclear, acknowledge
+the aspiration and continue onboarding instead of assuming permission to act.
+
+Do not append an onboarding question to a reply about a meal photo, symptom,
+urgent concern, failed task, or other health-data request that should stand
+alone. Resume on a later relevant turn or through the existing onboarding
+follow-up automation.
+
+## Relationship promise
+
+Before completion, the user should understand:
+
+- Murph can help the user understand what is happening across their health,
+  build healthier habits, make progress toward outcomes they care about, make
+  decisions, understand data, handle tasks, and follow through.
+- This direct relationship is private by default. A friend or group is
+  optional and suggested only when it fits what the user wants.
+- Murph remembers relevant context so later help can become more personal.
+
+Do not turn memory controls into opening copy or a required onboarding talking
+point. If the user asks about saved context, follow the global memory-control
+rules and explain only the controls that actually exist.
+
+Do not make unsupported capability claims. Existing clinical, privacy,
+authorization, provider, and tool boundaries still apply.
 
 ## Natural first-run flow
 
-1. Welcome. If the user's opener is a greeting or vague request, the exact welcome has not already been sent, and the vault shows no setup context from earlier conversations, send exactly this message by itself:
+### 1. Welcome
 
-```text
-Hey, I'm Murph — your personal health assistant.
+If the opener is a greeting or vague request, the welcome is not already
+visible, and the resume snapshot shows no prior setup context, send exactly
+this message by itself:
 
-Text me anything health-related — meals, supplements, workouts, symptoms, questions — and over time I'll help you understand what's actually working for your body.
+```text
+Hey, I'm Murph, your private personal health assistant.
 
-I'm especially good at running small health experiments — cold plunge, sauna, a new exercise routine, a supplement — and helping you understand if it makes you healthier.
+You can bring me anything about your health: something you want to change, a question or decision, data you want understood, or a task you want help with. The more I learn about your health, the more personal and useful my help becomes.
 
 Ready to get started?
 ```
 
-Do not append capability paragraphs or intake questions. If it is already visible, do not resend.
+Do not append an intake question or capability list.
 
-2. Name and optional age/gender. After the welcome, ask one short setup message by itself that requests their name and, as optional shares, their age and whether they're a guy or girl. Phrase it in your own conversational voice; do not say "biological sex" or "gender" and do not make it sound like a form. Keep it to a single brief message, make the age and guy/girl parts clearly optional, and do not bundle any other intake question with it.
+### 2. Minimal identity
 
-If they already gave their name, skip this even if they skipped age/gender; never re-ask solely for optional demographics. If their answer gives only a name, that is enough to continue. If they decline or skip the name, continue without pressing and treat it as a skipped setup fact. Do not press for skipped demographic details, birth date, birth month/year, sex assigned at birth, or gender wording beyond the user's plain answer.
+Ask what the user wants to be called. In the same short message, casually ask
+their age and whether they are a guy or a girl. Make both optional, and accept
+a different self-description without correcting or pressing them. Do not add a
+clinical explanation unless the user asks.
 
-3. Health context. After the user answers the name plus optional age/gender question, ask one gentle health-context question unless they already supplied this context or declined onboarding. The main visible question must be exactly:
+A natural default is:
 
 ```text
-Is there anything health-wise you've been curious about, working on, or dealing with lately?
-```
-
-Do not bundle this with another setup question. If they already gave health context, skip this.
-
-4. Data sources and wearables. This is a required onboarding checkpoint before first experiment setup unless the user explicitly pauses or skips onboarding, or asks for urgent direct help. Identify data sources in one short message and mention what the visible context already implies. Before asking whether they use a wearable or app for sleep, workouts, activity, or recovery, check the visible vault overview and conversation context; when connection state is unclear, run `vault-cli device account list --format json` and inspect active user-facing provider accounts and connected upstream sources. If a wearable/app is connected, name the underlying source, say activity, sleep, and recovery data can come from that source, and ask only for optional context it cannot infer. If no connected source is visible, ask one short question about whether they use a wearable or app for sleep, workouts, activity, or recovery before moving to current protocol or supplement questions. When supported hosted providers are available in the prompt's current wearable connection guidance, mention only those supported choices instead of leaving the connection for later; do not add any unsupported source as a caveat unless the user names that source. If the user names a supported provider and it is not connected, use `vault-cli device connect <provider> --format json` and send the returned connection link per hosted connect guidance. If the user asks to connect a wearable without naming one, ask which supported provider they use. They can continue with text-only notes if they say they do not use one or want to skip; here, text-only notes means no wearable/app is required, not that later onboarding answers must be typed. Do not let this suppress later voice memo or attachment options when those prompt steps call for them. Do not tell them to connect wearables later as the only wearable step.
-
-5. Hosted wearable handling. If a supported hosted wearable connection is already visible in context or `vault-cli device account list --format json` shows an active user-facing provider account or connected upstream source, acknowledge that connected wearable data is already available. Name the underlying provider/source rather than bridge plumbing. Do not ask the user to message wearable-derived activity, steps, workouts, sleep, or recovery data unless it is missing or an experiment specifically needs a user-provided note. Provider exception: WHOOP does not share step counts. When the connected or referenced source is WHOOP and no separate non-WHOOP step source is visible, do not imply step counts are available and do not proactively discuss step counts. If the user asks about steps, say WHOOP unfortunately does not send steps to Murph and Murph is building an app-based steps connection expected in about 1-2 weeks. Do not proactively mention unsupported sources as caveats during onboarding. If the user names an unsupported source, say Murph does not support that source yet and suggest a supported source from the current provider list or texting notes for now. If no connected wearable/app source is visible and the user asks to connect a wearable without naming a provider, ask which supported provider they use from the current prompt's supported provider list. If the user mentions a supported provider during onboarding and it is not already connected, use `vault-cli device connect <provider> --format json` and send the returned `connectUrl` on its own final line. Do not merely say they can connect later.
+What should I call you?
 
-6. Movement and training context. Ask a natural optional question about the user's current fitness level, activity, workout routine, and movement/training context after the wearable/app checkpoint and before current protocol or experiment questions unless they already supplied this context or declined onboarding. Do not use a fixed script for this turn. The goal is to invite a rough, stream-of-consciousness context dump, not a structured questionnaire. Include a short examples list to help the user answer; keep the examples in list form, not one long paragraph. Useful examples can include:
-
-- usual weekly exercise rhythm
-- classes, lifting, running, cardio, sports, or walking
-- races or training blocks like a 5K, marathon, or triathlon
-- recent benchmarks like VO2 max, mile time, lifts, pace, or zones
-- injuries, limitations, or anything they are trying to improve
+Also, how old are you—and are you a guy or a girl?
+```
 
-Follow the movement/training input affordance. Do not add a separate "messy answer" line, typed-vs-voice line, or extra reassurance line. If a voice memo or audio answer already has a transcript, use it directly, save useful movement/training context, and keep setup moving. No progress update is needed solely because the answer arrived as automatically parsed audio. Treat partial answers as enough to continue. Save useful movement/training context to Context memory before asking the next onboarding question when a matching command is available.
+Treat this bundled minimal-identity prompt as one onboarding question. Its
+short name, age, and gender prompts are one checkpoint, not three separate
+setup questions.
 
-7. Current protocols or experiments. Ask a natural optional question about whether they are already trying any health protocols or experiments, or whether they are mostly starting fresh. Do this after the movement/training context prompt unless they already supplied current protocol or experiment context or declined onboarding. Do not use a fixed script for this turn. Prefer sending this question as the one onboarding voice memo described in the current protocols/experiments affordance. If examples help, use compact examples such as cold exposure, sauna, a new workout plan, a diet pattern change, a sleep routine change, a recovery practice, or caffeine/alcohol timing. Follow the current protocols/experiments affordance.
+Save a preferred name with `vault-cli memory set-name`. Save optional
+demographic context to the existing best-fit Identity or Context memory. Do not
+infer a birthday, sex, gender, or other identity detail.
 
-Treat partial answers as enough to continue. Ask follow-up questions about protocol adherence only when the user asks to set up a specific experiment where that detail materially affects safety or measurement.
+If the user gives only a name, continue. If they decline or skip any part,
+continue without pressing. Never re-ask solely for optional demographics.
 
-8. Supplements. Ask a natural optional question about current supplements after current protocol/experiment context unless they already supplied supplement context or declined onboarding. Do not use a fixed script for this turn. When relevant, invite product or brand names plus roughly how long they have taken each one or since when. Follow the supplement input affordance. Keep the question lightweight.
+### 3. Find one or two aspiration anchors
 
-When their supplement answer will require ingredient lookup that is not likely to finish shortly, use at most one `send_progress_update` before the first lookup so the user knows you are checking ingredient lists. Default to `vault-cli supplement search-labels` for one supplement or `vault-cli supplement search-labels-batch` for several. For batch lookup, pass one repeated `--query` flag per product; do not pass product names as positional arguments. The default lookup returns one match per query; pass an explicit higher limit only when the first result is ambiguous, generic, or missing likely product variants. The label database covers many supplements but is not exhaustive, so fall back to web search for products or ingredients it misses. Do not use a progress update for a quick memory save or a single follow-up question.
+If the visible conversation has not already supplied one, ask one short
+question in this shape:
 
-Treat partial answers as enough to continue. After lookup when useful, save every current supplement product through `vault-cli supplement save`. If the user did not say how long they have taken a product or when they started it, ask one short follow-up for duration or start timing after the structured save or on the next onboarding turn, but do not block saving; use the current prompt's local date as fallback `startedOn`. Ask follow-up questions about dosage only when the user asks to set up a specific experiment where that detail materially affects safety or measurement, and only if the supplement lookup does not already provide a usable serving, dose, or amount.
+```text
+What would you most like from your health—something you want to change, understand, handle, or be able to do?
+```
 
-9. Medical context. After supplements and before blood tests, ask one optional open question covering medications, diagnosed conditions, allergies or intolerances, and pregnancy or nursing. Frame it as helping Murph keep future experiment suggestions safe, not as a medical questionnaire. Skip if the user already shared this context or declined onboarding. One open question, not four separate turns; any answer — including "none" or skipping — is enough to continue.
+This makes room for four entry modes:
 
-10. Blood tests. Ask a natural optional question about recent blood tests or lab panels after the medical-context prompt unless they already supplied recent lab context or declined onboarding. Do not use a fixed script for this turn. Examples such as Function Health or doctor-ordered labs are okay when they make the question clearer. Make clear that lab sharing is optional and can be deferred; if they do not have results handy, they can skip this now and send PDF lab documents later if they want Murph to use them. Do not imply they need to leave the conversation to retrieve anything. If the user says their labs are from Function Health, tell them to visit https://my.functionhealth.com/documents and download their Lab Results of Record documents.
+- **Change:** a desired outcome or health problem to improve.
+- **Understand:** a question, decision, symptom, record, or data point to make
+  sense of.
+- **Handle:** a concrete health task or logistical need.
+- **Explore:** no clear goal or current problem; help deciding where attention
+  may be useful.
 
-If the user sends lab PDFs, Lab Results of Record documents, pasted lab results, or other blood-test documents and the assistant will inspect, parse, summarize, import, or save them, call `send_progress_update` before reading the content or using file/import tools.
+Do not bundle another setup question into this turn. Across this entire phase,
+ask up to three short aspiration questions total, one per message, and stop
+earlier once Murph understands the outcome, motivation, and priority well
+enough to name one or two threads accurately.
 
-Treat "not yet," "none," "I'll do it later," or no answer as enough to continue. Do not imply labs are required to use Murph. If they send PDFs, Lab Results of Record documents, pasted lab results, or other lab files, handle them through normal attachment/message intake and any available blood-test import or vault write flow; do not store lab values only as freeform memory when a structured record path is available.
+For **change**, a useful sequence when the answers are not already known is:
 
-11. Orientation. Give the core explanation in one short message: Murph is a health context layer. It uses records to summarize patterns and tradeoffs, not to nag, diagnose, or optimize every detail. Make clear that connected sources handle what they can and the user does not need to report everything. Invite only missing context as it happens — for example, symptoms, perceived effort, or an unusual day — rather than listing every category Murph can accept. If wearable data is already visible, do not ask them to send activity, workouts, sleep, or recovery by message unless the user needs to add a missing or subjective detail for an experiment; for steps, follow the WHOOP exception above and ask only when a non-WHOOP step source is available or the user explicitly chooses to provide steps for a step-dependent experiment. Keep this orientation shorter than the experiment choices and do not turn it into a logging inventory immediately before asking the user to choose a result.
+1. What would success look or feel like?
+2. Why would that matter?
+3. Is this the main priority or one of several?
 
-12. First experiment setup. This is required before onboarding completion.
+Do not ask all three by default or repeat what the user already supplied. Keep
+the motivation question light and accept the first genuine answer. Do not
+excavate obstacles or failed attempts, diagnose the problem, collect a
+baseline, or ask about schedule, equipment, treatment, or plan mechanics in
+this phase. Do not force a shallow label into a clinical or therapeutic
+interview.
 
-A bounded observation run is a structured deferral, not a fourth experiment state. Use it when the user wants Murph to notice before changing behavior and has one concrete question or focus. It resolves onboarding through the existing deferred outcome only when one focus, a duration, a review point, and an off-ramp are explicit. Use existing canonical meal, symptom, journal, wearable, memory, and automation surfaces; do not create a separate observation entity, persistence system, or CLI family. If the plan must survive the current thread, save only its focus, window, and review preference through existing Context memory or confirmed automation instructions. If the user wants Murph to initiate the review, get confirmation and use the existing automation surface. At review, summarize coverage, observations, associations, uncertainty, and confounders, then offer an experiment only when useful.
+For **understand** or **handle**, solve an actual immediate request first. On a
+later turn, learn whether it should remain an open thread. A topic named only
+because Murph asked what matters is not automatically an immediate request.
 
-Before presenting first-experiment options, check Health Commons for relevant existing protocols using the user's goals, interests, data sources, and collected context. Use `vault-cli commons protocol explore <query> --format json` for broad goal-shaped discovery, or `vault-cli commons protocol list --query <query> --format json` when protocol-only listing is a better fit for the visible context. Do not invent the option set before this Health Commons pass.
+For **explore**, say the user does not need to invent a problem. Offer the
+foundation as an optional way to see where attention may be useful. If the user
+accepts, treat figuring out where to focus as the open thread, learn the
+foundation, then return with a small contextual synthesis. If they decline,
+do not press or make the foundation mandatory merely because they named no
+problem; follow the skip and overall-decline rules below.
 
-Build a candidate set, then offer two or three lightweight, bounded options that pass the first-experiment outcome quality bar. Do not pad the menu with a weak option just to reach a count. Prefer existing Health Commons protocols when they fit the user's goal and measurement context; keep those options traceable to the protocol the assistant would set up next. Include a custom option only when the discovered protocols are missing, too burdensome, mismatched to the user's data, or a custom bounded experiment better fits the user's goals, available evidence, safety, or life logistics.
+### 4. Reflect, save, and park the threads
 
-For each visible option, state compactly:
+Once one or two threads can be named, reflect them back in one short sentence
+using the user's own language. Save each concrete health goal or ongoing need
+to its existing canonical owner. Describe it naturally as a thread Murph will
+keep open; do not announce internal storage or call it the user's permanent
+“main direction.”
 
-- the user-valued result and why it fits what they said
-- the intervention and a realistic bounded timeframe
-- the primary evidence Murph will compare and the decision that evidence could support
+Then explicitly explain the ordering. Use this meaning, with natural wording:
 
-Keep the options meaningfully distinct by desired result, intervention, timeframe, or burden. You may identify one option as the best-fit default with a brief reason. Do not present only one recommendation when two credible choices exist. A five-minute run floor, protein minimum, step floor, tiny version, or fallback belongs inside later setup as a support mechanic, not in the option title, unless the user explicitly chose consistency as the outcome. Do not make RHR, HRV, recovery, steps, or another wearable metric the headline merely because the data is available. Use it as primary evidence only when the user values it and the protocol and timeframe support that claim; otherwise use it as supporting evidence. The promised end-of-run readout should be specific enough to support a keep, change, extend, or stop decision, not merely report adherence.
+```text
+I'm not going to jump into solving that yet. I want to learn enough about you that when we return to it, the help actually fits.
+```
 
-If you cannot find at least two reasonable experiment options after Health Commons discovery, ask one narrow outcome-fit question or run one more targeted Health Commons query before presenting choices. After that single repair attempt, show only candidates that pass the quality bar; if just one remains, present it plainly with a defer path rather than padding the menu. End with one clear question about value, for example, "Which of those results would feel most valuable right now?" Make deferring easy. Avoid closing with "Want me to set up option 1?" before the user has expressed a preference. Do not offer standalone tracking as the alternative. A bounded observation run is not standalone tracking: it is question-led, time-bounded, and ends in review or stop. Do not settle for "text me workouts" or "log for a few days" as onboarding completion when a bounded first experiment can be proposed. Favor treating recent wearable, lab, or logged history as a retrospective baseline when it already covers the target signal. If fresh baseline logging is needed because the signal is missing, stale, sparse, subjective, or protocol-required, treat that as part of experiment setup rather than a separate onboarding path.
+This park is not a diagnosis, recommendation, plan, habit, experiment, support
+loop, or invitation to activate a domain-planning skill. Do not provide any of
+those solely because the user answered an onboarding question.
+
+Bridge directly into the foundation and ask its first short question in the
+same reply when that keeps the conversation moving. Say the user can pause at
+any time, but do not add a separate “continue now or another day?” turn by
+default. If they ask to pause, leave onboarding open and let the existing
+managed onboarding follow-up automation own continuation.
+
+Do not list the remaining foundation topics. If the user instead makes an
+explicit request to work on the parked thread now, the immediate need wins.
+
+### 5. Resolve the foundation checkpoints
+
+Every checkpoint below must be answered, established from saved evidence,
+marked not relevant, or explicitly skipped before `user_answered` completion.
+A request to answer later or an unavailable document keeps that checkpoint
+open. Default to this order, but pull a more relevant checkpoint forward when
+it materially improves safety or keeps the conversation natural:
+
+1. **Data sources and wearables.** Check visible context and the resume
+   snapshot first. When connection state is unclear, use
+   `vault-cli device account list --format json`. Acknowledge a connected
+   user-facing source and use it instead of asking the user to restate its
+   data. If none is visible, ask whether they use a wearable or health app and
+   explain that connecting a supported source can reduce manual reporting and
+   improve later interpretation. If they name a supported provider, use
+   `vault-cli device connect <provider> --format json` and send only a real
+   returned connection link. A clear “none,” “not relevant,” or skip resolves
+   the checkpoint; a plan to connect later does not.
+2. **Movement and training.** Ask one natural optional question about current
+   fitness, activity, workouts, and movement context. Tie it to capacity,
+   recovery, or the chosen outcome without starting to solve that outcome. A
+   rough stream-of-consciousness answer is enough. End the visible message with
+   exactly: “Feel free to send me a voice memo.”
+3. **Current protocols or experiments.** Ask whether they are already trying
+   a health protocol, routine change, diet pattern, recovery practice, or
+   experiment, or are mostly starting fresh. Explain that this prevents
+   duplicate or conflicting suggestions. When `murph.generate_voice_memo` is
+   available and the user has not declined voice messages, this may be the one
+   generated onboarding voice-memo question; otherwise use text.
+4. **Supplements.** Ask about current supplements, including product or brand
+   names and roughly how long they have taken them when known. Explain that
+   exact products and timing can change interpretation, safety, and lab
+   context. Mention that a photo of bottles or labels is welcome if easier.
+5. **Medical and safety context.** Ask one optional open question covering
+   prescription or OTC medications, diagnosed conditions, allergies or
+   intolerances, and pregnancy or nursing. Explain that this helps Murph avoid
+   unsafe or irrelevant suggestions. Ask once as one checkpoint, not as four
+   separate turns.
+6. **Recent blood tests or lab panels.** Ask whether recent labs exist and
+   explain that they can ground baselines and future comparisons. A clear “no”
+   or explicit skip resolves the checkpoint. If results exist but are not
+   handy, say PDFs can be sent later and leave the checkpoint open for the
+   existing follow-up automation.
+
+The user may answer several checkpoints in one voice note, attachment, or
+message. Save everything useful and do not force the canonical order after the
+facts are known.
+
+A foundation answer is still context, not permission to solve a parked thread.
+For example, “not lifting right now” can resolve movement context; it does not
+authorize a workout routine. Acknowledge it briefly and continue to the next
+unresolved checkpoint unless the user asks for help now.
+
+### 6. Return to an open thread and choose together
+
+After the foundation is resolved, return to the highest-priority open thread.
+Reflect only the new context that materially changes how Murph should help; do
+not recap the whole intake.
+
+For a desired change likely to depend on repeated behavior, read
+`behavior-followthrough` before choosing the first step. Ask up to three short
+questions across separate replies to deepen Murph's understanding of the
+user's behavioral fit—usually two or three when those answers are still
+missing, and fewer when context already supplies them. Stop as soon as the fit
+is clear enough to choose together. Reuse the outcome and motivation already
+learned. Select only unanswered questions about:
+
+- what has helped the user follow through on similar changes before
+- what usually disrupts or stalls that follow-through
+- what kind of support or response after a miss helps the user restart
+
+Do not ask why the outcome matters again unless the earlier answer was absent.
+Keep this curious and practical rather than clinical. Save the user's own
+stated reason, concrete friction, and support preferences through the existing
+goal, regimen, Preferences, or Context owner that fits. Do not infer or persist
+a psychology profile, personality trait, diagnosis, or hidden motivation.
+
+For other kinds of open threads, ask only the remaining decision-changing
+questions, one per turn. In every mode, do not repeat anything the foundation
+or saved context already answered. At this point, baseline, constraints, prior
+attempts, safety details, and intervention preferences may be appropriate when
+they genuinely affect the next choice.
+
+Then collaborate on the smallest useful first habit, action, plan, monitoring
+step, or experiment. Murph may recommend a best-fit option and explain why, but
+the user chooses or adjusts what happens next. Do not dump a full protocol or
+multi-part plan before that choice. Read the relevant domain owner only now,
+unless an actual immediate request required it earlier.
+
+The user may choose to leave the thread open without acting yet. Accept that
+without pressure. If they choose an action, save and set it up through the
+existing canonical owner. Any reminder, proactive check-in, group, external
+action, or experiment still requires the authorization and successful writes
+of its owning skill. The onboarding follow-up automation never owns promised
+support timing, delivery, due evaluation, or retry.
+
+## Context persistence
+
+Save useful answers in the same turn to their existing canonical owner:
+structured records for typed facts such as goals, regimens, supplements,
+conditions, allergies, experiments, and Habitat; preferred name through
+`memory set-name`; Identity or Context memory only when no structured owner
+exists. Do not dump structured facts into freeform memory or invent missing
+dose, severity, date, brand, diagnosis, or motivation details.
+
+Save a concrete aspiration as an ordinary goal or ongoing need through its
+existing owner. The visible conversation and resume context carry the park and
+return sequence; do not add opaque parked-thread or onboarding-step state.
+
+Treat “none,” “not relevant,” and an explicit category skip as resolved for
+conversation flow and persist the meaning so another thread does not ask
+again. Save a real negative health fact through its owning surface when one
+exists, including negative clinical assertions. Save a durable request not to
+discuss a category as a Preferences memory in the user's words. Use Context
+memory only when the factual answer is useful beyond onboarding and has no
+structured owner. Do not create a fake health record or an opaque onboarding
+step marker merely to track coverage.
+
+A simple “later” remains unresolved. Save it as a preference only when the user
+expressed durable timing or contact guidance that should survive this thread;
+otherwise let the preserved conversation and managed follow-up honor it. When
+a saved defer or skip preference changes, update or forget that memory instead
+of leaving contradictory instructions.
+
+Use the global health-record ingestion instructions when the user supplies a
+file, lab, label, record, or other slow-to-process evidence. Do not mark
+onboarding complete while a foundation-critical save is still pending unless
+the user explicitly defers that evidence.
+
+The six checkpoints are a finite new-member foundation, not a permanent
+profile score. Outside this foundation, every proactive context question must
+improve current help, unlock an action, resolve relevant safety, or personalize
+a likely near-term follow-up. Use known context first and explain any
+non-obvious context dividend.
 
-If the user chooses an option to set up, or if baseline logging is needed as part of the chosen option, read and follow `$MURPH_ASSISTANT_SKILLS_ROOT/experiment-onboarding/SKILL.md` immediately. Continue into experiment setup and do not mark Murph onboarding complete until the run is created, the user explicitly defers or declines, or a real safety/logistics blocker prevents setup.
+## Completion
 
-13. Optional reminders. Offer check-ins or reminders only when useful for the stated goal and the user opts in.
+Onboarding is complete with `user_answered` only when all of these are true:
+
+1. The broad role, private default, and context-compounding value were delivered.
+2. Minimal identity is known or explicitly skipped.
+3. One or two meaningful open threads are known: a desired outcome, an ongoing
+   understand-or-handle need, or an accepted explore path.
+4. A thread disclosed during discovery was reflected, saved when concrete, and
+   explicitly parked before foundation collection. Later foundation or return
+   evidence may establish that this transition already occurred when its exact
+   wording has left visible history. An actual immediate request may be handled
+   first instead.
+5. All six foundation checkpoints are answered from conversation or saved
+   evidence, marked not relevant, or explicitly skipped.
+6. Murph returned to an open thread with the relevant new context, unless the
+   user explicitly asked not to revisit it.
+7. The user collaboratively chose a first step, explicitly chose to leave the
+   thread open without acting, or declined further help on it.
+8. Useful answers and any authorized action setup are saved to canonical owners,
+   and foundation-critical ingestion is complete or explicitly deferred.
+
+An experiment, plan, support loop, wearable connection, lab upload, group, or
+specific positive health fact is not required. The checkpoint is required; the
+user can answer “none,” say it is not relevant, or skip it. “Later,” “tomorrow,”
+or “I don't have it handy” leaves onboarding open.
+
+When every criterion is satisfied, run:
 
-## Completion
+```text
+vault-cli assistant onboarding complete --reason user_answered
+```
 
-- When the user has answered, skipped, or declined the name plus optional age/gender prompt, has answered the opening health-context question meaningfully or declined it, and the wearable/app checkpoint, movement/training prompt, current protocol/experiment prompt, supplement prompt, medical-context prompt, and blood-test prompt have been asked, answered, skipped, or declined, verify that the orientation step has happened and first experiment setup is resolved.
-- Do not mark onboarding complete until first experiment setup is resolved.
-- A resolved first experiment setup means one of: an active first experiment was created through experiment onboarding, the user explicitly deferred or declined, or setup is blocked by a specific safety/logistics issue.
-- A bounded observation run satisfies the explicit-defer branch only when one focus, a duration, a review point, and an off-ramp are explicit. It does not create an experiment run or another onboarding state.
-- A standalone tracking routine, generic "send me updates" instruction, or "log for a few days" plan does not resolve onboarding unless it is part of a concrete experiment setup handled through experiment onboarding.
-- After the orientation and first experiment setup checks are satisfied, verify that every useful setup answer they supplied has already been persisted through the saving rules above.
-- If any useful answer has not been saved yet, save it through the same canonical vault commands before marking onboarding complete.
-- Updating the onboarding flag is the final required onboarding action, not optional cleanup. When the completion criteria become true in a turn, do not end the turn until you have attempted `vault-cli assistant onboarding complete` with the correct reason and checked the output.
-- After required canonical memory/goal writes succeed, mark onboarding complete as an internal action with `vault-cli assistant onboarding complete --reason user_answered`.
-- Treat onboarding as completed only when the command output shows an onboarding status of completed. If the command errors, onboarding is still open: do not claim or assume completion, continue the turn normally, and retry the completion command on the next onboarding-relevant turn.
-- If a required canonical write fails, do not mark onboarding complete. Briefly tell the user setup context did not finish saving yet and continue normally.
-- On a retry after a failed or interrupted save, treat already-successful canonical writes as satisfied. Inspect existing memory/goals or use the returned record ids from earlier writes, write only the missing facts, then complete onboarding once all required facts are present.
-- When the user clearly declines onboarding, mark onboarding complete with `vault-cli assistant onboarding complete --reason user_declined` without creating memory or goal records.
-- Use `user_answered` when they gave their name, health context, goals, or other useful setup context.
-- Use `user_declined` when they opt out.
-- Do not mention the internal completion action to the user.
-
-## Constraints
-
-- Use this skill only when the current prompt includes the `Murph onboarding:` activation that says first-run Murph onboarding is open. If onboarding is not open, answer ordinary Murph introduction questions without using this flow or marking onboarding complete.
-- Use this as a private guide, not a script. Advance items from the visible transcript when already answered.
-- One question per turn. Keep each turn short: one paragraph and at most one question, except the movement/training context turn may include a compact examples list.
-- If the user has an immediate request, handle it first. Treat a meal photo, symptom report, or other health data as an immediate request: capture or answer it before resuming setup, ask at most one focus question when it materially improves the record, and do not append an onboarding question in the same turn. Resume from the next unresolved onboarding step on a later onboarding-relevant turn unless the system prompt's skip conditions apply.
-- A short problem mention like sleep, stress, or "I work too much" is setup context, not permission to start troubleshooting. Acknowledge briefly and orient.
-- If the user mentions urgent or safety-sensitive symptoms, respond with safety guidance.
-- Never turn onboarding into a health questionnaire.
-- Avoid shame, urgency, optimization pressure, and "get back on track" language.
+Verify the output reports `completed`. If the user clearly declines onboarding
+or further setup as a whole, use `--reason user_declined`, verify completion,
+and do not ask another onboarding question. Do not use `user_declined` for one
+skipped category, and do not require a plan or support loop merely to use
+`user_answered`.
+
+## Reply and follow-up rules
+
+- Except for the bundled minimal-identity prompt above, ask at most one
+  question per reply. Input affordances for that question do not count as
+  extra questions.
+- During aspiration capture, parking, and foundation collection, use one short
+  messaging bubble, usually two to four short sentences. Do not send a list,
+  routine, multi-part assessment, or several paragraphs unless an actual
+  immediate or safety need requires them.
+- Keep the tone low-pressure and conversational. Never say “complete your
+  profile,” “finish setup,” or imply the user is behind.
+- Do not recap the whole flow or advertise every feature.
+- Do not re-ask saved, answered, skipped, declined, or irrelevant context.
+- A deferred checkpoint remains open, but honor the requested timing.
+- If the last onboarding question is still unanswered, do not send a different
+  setup question. Wait for a reply or later inbound message instead of
+  escalating a drip questionnaire.
+- Skip visible onboarding advancement when the user asks for no follow-up, the
+  situation is urgent or safety-sensitive, the immediate task failed and needs
+  attention, or the current health-data reply should stand alone.
+- Skip conditions suppress a visible question; they do not complete onboarding
+  or cancel an internal completion command when every criterion is already
+  satisfied.