sleep-quality

name: sleep-quality description: Produce a bounded SleepProposal for today by consuming the runtime-computed classified_state + policy_result and applying judgment-only steps — action matrix, rationale prose, vendor cross-check. The runtime already did every band, every score, and every R-rule; this skill does not re-derive them. allowed-tools: Read, Bash(hai state snapshot *), Bash(hai state read *), Bash(hai propose *), Bash(hai intent list *), Bash(hai target list *) disable-model-invocation: false

Sleep Quality

All arithmetic happens in code. Your job is: read the bundle, honour the policy result, pick an action, write the rationale, persist.

Load the bundle

hai state snapshot --as-of <today> --user-id <u> --evidence-json <hai clean output>

Under snapshot.sleep you receive these blocks:

• today — last night's accepted_sleep_state_daily row (sleep_hours, score, minute breakdowns, awake_count, avg_sleep_respiration, …), or null.
• history — recent nights of sleep rows for context.
• signals — the runtime-derived dict the classifier consumed: sleep_hours, sleep_score_overall, sleep_awake_min, sleep_start_variance_minutes, sleep_history_hours_last_7. Context only; never re-derive.
• classified_state — sleep_debt_band, sleep_quality_band, sleep_timing_consistency_band, sleep_efficiency_band, coverage_band, sleep_status, sleep_score, sleep_efficiency_pct, uncertainty. Source of truth.
• policy_result — policy_decisions[], forced_action, forced_action_detail, capped_confidence, and (v0.2.0 W-PROV-2) optional evidence_locators[]. The runtime always emits a row-level locator citing today's accepted_sleep_state_daily; an R-chronic-deprivation firing additionally cites sleep_hours per night in the trailing window. Source of truth.
• missingness — per state_model_v1.md §5.

Protocol

1. If the policy forced an action, use it

If policy_result.forced_action is set, action is that value and action_detail is policy_result.forced_action_detail. Confidence: low for defer_decision_insufficient_signal, else moderate. Skip the action matrix; jump to rationale.

Note: when R-chronic-deprivation fires, it forces sleep_debt_repayment_day and the corresponding policy_decisions[] entry carries an escalate tier. Copy the decision verbatim — severity is recorded in the audit record, not in the action.

2. Otherwise, pick from the action matrix

Keyed on classified_state.sleep_status:

3. Confidence

Default from classified_state.coverage_band: full → high, partial → moderate, sparse → moderate, insufficient → low. If policy_result.capped_confidence is set, it lowers the default but never raises it. The vendor cross-check below may lower further.

4. Vendor cross-check

The snapshot carries Garmin's own sleep signals under sleep.today:

• sleep_score_overall vs classified_state.sleep_quality_band — if the vendor score lands in a meaningfully different band than the classifier reports (e.g. vendor 88 "Good" but classifier band is "fair" after a threshold tweak), cap confidence at moderate and add agent_vendor_sleep_quality_disagreement to uncertainty.
• avg_sleep_respiration, avg_sleep_stress — informational only in rationale; never action-bearing.

5. Rationale (5–8 lines)

One line per band that informed the decision. Name the band; do not re-derive it.

Examples: sleep_debt=<band>, sleep_quality=<band>, sleep_efficiency=<band> (<pct>%), sleep_timing_consistency=<band>, sleep_status=<status>, chronic_deprivation_detected (if R-chronic fired), agent_vendor_sleep_quality_disagreement (if applicable).

6. Uncertainty

Start with classified_state.uncertainty (already sorted + deduped). Append any tokens you added (e.g. *_unavailable_at_source derived from the snapshot's missingness token, vendor-disagreement tokens). Re-sort alphabetically; deduplicate.

7. Follow-up

Sleep emits a SleepProposal, not a recommendation, so it has no follow_up field. Synthesis assigns review semantics per finalised plan. Skip this step.

On defer_decision_insufficient_signal, synthesis uses the sleep-domain template "Did anything shift in your sleep last night worth noting?" (owned by core.narration.templates.DEFER_REVIEW_QUESTION_TEMPLATES).

Output

Emit a SleepProposal JSON and call hai propose --domain sleep --proposal-json <path>. The propose tool validates the shape and appends to proposal_log; it is your determinism check.

proposal_id = prop_<for_date>_<user_id>_sleep_01 (idempotent on (for_date, user_id, domain); re-running on the same night does not produce a new row).

Copy policy_result.policy_decisions into the output's policy_decisions verbatim — the runtime decided them; you do not re-edit or add new ones. v0.2.0 W-PROV-2: if policy_result.evidence_locators is present, copy that list verbatim into the proposal's evidence_locators field — do NOT derive locators yourself; the runtime computed them. If the field is absent, omit evidence_locators from the proposal.

Invariants

• You never compute a band, a score, or a ratio. classified_state is the source of truth.
• You never evaluate an R-rule (require_min_coverage, no_high_confidence_on_sparse_signal, chronic_deprivation_escalation). policy_result is the source of truth; you honour forced_action and capped_confidence.
• You never apply X-rule mutations. Synthesis owns all cross-domain reasoning; this skill emits one domain's bounded proposal.
• You never emit an action outside the v1 enum (maintain_schedule, prioritize_wind_down, sleep_debt_repayment_day, earlier_bedtime_target, defer_decision_insufficient_signal). hai propose enforces this. Chronic-deprivation severity is recorded in the policy_decision tier, not in the action.
• You never fabricate values for missing evidence; missing stays missing.
• If a decision isn't reasoned in rationale[] or policy_decisions[], it didn't happen.