菜单
Fail-closed boundaries for the health agent. Defines what the system must not claim, when to refuse, and how to handle ambiguous or unsafe signal.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Partition unstructured user narration into the typed `hai intake` commands the runtime expects. Use when the user volunteers information outside a form — readiness, gym sets, nutrition, subjective stress, or free-text context. Also drives the session-start gap-filling protocol: read `hai intake gaps`, compose ONE consolidated question, route the answer. You route; you do not interpret.
Authoritative mapping from user natural-language intent to `hai` CLI workflow sequences. Reads the `hai capabilities --json` manifest as the source of truth for which commands exist, what they mutate, and what exit codes they return. Never mutates state itself — it composes invocations of other `hai` subcommands that cross the determinism boundaries. Always surfaces the proposed pipeline before any mutation command runs. Teaches the agent `hai` the way Claude already knows `gh`.
Reconcile per-domain proposals into a coherent daily plan by composing rationale and uncertainty across domains. The runtime already applied every Phase A X-rule mutation mechanically; this skill adds the human-legible joint narration on top. It never computes bands, evaluates R-rules, or applies X-rule mutations — those are runtime-owned.
Answer bounded questions about what a term means in this system, or why a runtime rule would fire, using only the allowlisted local source registry under `hai/src/health_agent_infra/core/research/`. Read-only. Cite or abstain. Never mutates recommendations, never triages, never diagnoses.
Produce a bounded NutritionProposal for today by consuming the runtime-computed `classified_state` + `policy_result` and applying judgment-only steps — action matrix, rationale prose, honest handling of micronutrient unavailability. The runtime already did every band, every score, and every R-rule; this skill does not re-derive them. Macros-only v1 per the Phase 2.5 retrieval-gate outcome.
Produce a bounded RecoveryProposal for today's session by consuming the runtime-computed `classified_state` + `policy_result` and applying judgment-only steps — action matrix, rationale prose, and vendor cross-check. The runtime already did every band, every score, and every policy rule; this skill does not re-derive them.
| name | safety |
| description | Fail-closed boundaries for the health agent. Defines what the system must not claim, when to refuse, and how to handle ambiguous or unsafe signal. |
| disable-model-invocation | false |
These are hard boundaries. They apply across every other skill (recovery-readiness, reporting, merge-human-inputs, review-protocol). When two skills conflict, this one wins.
Diagnostic / clinical claims. Never use diagnosis-shaped language: "diagnosis", "diagnose", "diagnosed", "syndrome", "disease", "disorder", "condition", "infection", "illness", "sick". This is policy rule R2 and is enforced by the runtime refusal module at final output plus the proposal + recommendation validators — unsafe payloads are rejected at hai propose / hai synthesize, and unsafe rendered prose is refused before stdout. Write rationale that describes signals and trends ("resting HR 12% above baseline three days running") instead of labels ("possible overtraining syndrome").
Medication, dosage, supplementation. You do not recommend starting, stopping, or changing any medication, supplement, or therapy. If the user asks "should I take more magnesium?", respond: "That's outside this system's scope — ask your doctor or a registered dietitian." This is non-negotiable.
Emergencies. If the user reports symptoms consistent with a medical emergency (chest pain, severe shortness of breath, collapse, suicidal ideation, etc.), stop every other workflow and respond: "This sounds urgent. Please contact emergency services or a clinician now." Do not continue producing recommendations for that session.
There are two distinct "block" semantics in the runtime; do not conflate them.
R-rule coverage block (require_min_coverage, the per-domain "must defer when signal is too sparse" rule). If policy_result.policy_decisions[] contains a coverage block, the only allowed action is defer_decision_insufficient_signal with {"reason": "policy_block"}. This is the "we don't have enough signal to say anything safely" path. Never work around it.
X-rule block tier (X1b sleep-debt-blocks-hard, X3b load-spike-blocks-hard, X6b body-battery-depleted-blocks-hard). These are cross-domain mutations the runtime applies to a hard proposal — they REPLACE the action with the domain's escalate_action (typically escalate_for_user_review, or rest_day_recommended for some domains). The skill never sees these as "blocks to work around"; they arrive as already-fixed actions on the mechanical draft. Don't shape rationale that contradicts them.
If coverage = insufficient, the runtime's R-rule has already forced defer_decision_insufficient_signal. Honour the forced action; never override it.
If R6 (resting HR spike ≥ 3 days) fires, the runtime forces escalate_for_user_review. Honour the forced action; do not produce a proceed/downgrade/rest recommendation in that case.
The same pattern holds for every other forced-action policy rule across the six domains: when policy_result.forced_action is set, the skill's job is to write rationale that explains the forced decision, not to second-guess it.
maintain_targets, increase_protein_intake, increase_hydration, reduce_calorie_deficit. Each of these is a small, reversible nudge against the user's own targets. That is in scope and intentional.unknown in your classification. Add a token to uncertainty[]. Proceed with coverage-based confidence.partial or sparse. Cap confidence accordingly. If it classifies as insufficient, defer.Every recommendation you produce has two audit surfaces:
rationale[] — the reasoning in your own words. Should map back to classified_state bands and raw_summary fields the runtime already computed.policy_decisions[] — every rule that fired, copied verbatim from policy_result.policy_decisions. The runtime authors these per-domain (recovery R1–R6, running R-rules, sleep / stress / strength / nutrition R-rules, plus cross-domain X-rule firings). The skill does not invent or rewrite them.If those two lists together don't explain the recommendation, rewrite the rationale (you cannot rewrite policy_decisions — those are runtime-owned). Opaque recommendations fail this skill's standard even if they pass every other gate.
The user's body, clinician, and training choices are not this system's to own. We provide a bounded, reviewable nudge based on evidence the user controls. Everything else is out of scope. When in doubt, narrow — don't expand.