Imported from jluo41/Tools (plugins/haipipe-toolkit/skills/task/4_individual/haipipe-individual-inference-report/SKILL.md). Install upstream with npx skills add jluo41/Tools --skill haipipe-individual-inference-report. Copyright stays with the author.

Skill: haipipe-individual-inference-report

Per-individual prediction → interpretation → audience-tailored report.

  📥 individual data        🌐 endpoint prediction       🤖 LLM compose
  (parquet)              (haipipe-end-deploy-local)   (claude_agent_sdk)
       │                          │                          │
       └─ ctx ───┬───── forecast ─┴───── system_prompt ───────┘
                │                       (persona)
                ▼
           📨 Report{json, nl}
              + telemetry

Sibling progression in task/4_individual/:

Layout

src/
  compose_report.py     SDK call, XML extract, parse
  report_schema.py      pydantic Report model
  persona_loader.py     resolve --persona name | path → system_prompt + meta

personas/                   ← shipped reference personas (1-2)
  patient-friendly/
    persona.yaml          metadata: audience, tone, model, safety_rules
    system.md             system prompt
    schema.md             <report> XML schema description

scripts/
  make_report_cli.py    end-to-end CLI: individual + persona → report

tests/
  (smoke against Subject-18, when written)

Quickstart

1. Start the local prediction endpoint (sibling skill):

ENDPOINT_PATH=_WorkSpace/6-EndpointStore/endpoint_cgm_patchtst_ohio_v0001 \
PORT=8765 \
    python Tools/plugins/haipipe-toolkit/skills/task/3_end/haipipe-end-deploy-local/scripts/serve_local.py

2. Generate a report:

python Tools/plugins/haipipe-toolkit/skills/task/4_individual/haipipe-individual-inference-report/scripts/make_report_cli.py \
    --individual Subject-18 \
    --persona patient-friendly

Output: _WorkSpace/7-AgentWorkspace/reports/<individual_id>/<persona>/<ts>/

report.json     structured payload (matches Report pydantic)
report.txt      patient-facing NL (3-6 sentences)
response.xml    raw <report> block from the LLM
meta.json       telemetry: model, cost, duration, session_id, ...

Persona system

A persona is a folder with three files:

<persona-dir>/
├── persona.yaml      audience, tone, model, language, safety_rules
├── system.md         system prompt
└── schema.md         <report> XML schema description

--persona accepts:

This lets external persona libraries (Samsung-internal, IRB-approved templates, etc.) live outside haipipe-toolkit and still be invoked without forking the skill.

Required fields in persona.yaml:

• audience (e.g. patient, clinician, parent)
• tone Optional: model, language, safety_rules, anything else the persona author wants to track (logged into report meta.json).

LLM call mechanics

Uses claude_agent_sdk (subprocess to local claude CLI). Auth flows through ~/.claude OAuth — same login the user did in this Claude Code session. Cost is reported (cost_usd_equiv in telemetry) but not billed when subscription auth is active.

The script unsets ANTHROPIC_AUTH_TOKEN and ANTHROPIC_BASE_URL before the SDK call to avoid the project's CRS proxy diverting the request away from OAuth (see repo memory reference_crs_proxy_gotcha).

Output schema (XML the model emits)

<report>
  <basics>{individual_id, dataset, gender, year_of_birth, disease_type}</basics>
  <current>{last_obs_dt, last_bg_mg_dl, recent_window_n, recent_min/max/mean}</current>
  <forecast_summary>{horizon_minutes, n_windows, pred_min/max/mean}</forecast_summary>
  <interpretation>
    <verdict>rising|stable|falling|mixed</verdict>
    <why>...</why>
    <actions><action>...</action></actions>
    <confidence>high|medium|low</confidence>
    <safety_flag>none|hypo_risk|hyper_risk|hypo_and_hyper_risk</safety_flag>
  </interpretation>
  <nl>... patient-facing prose ...</nl>
</report>

Failure modes

Reuses

• haipipe-individual-inference for load_patient_ctx, build_payload, client.call_predict
• haipipe-end-deploy-local for the prediction endpoint
• claude_agent_sdk for the LLM call (subprocess of claude CLI)
• Pattern reference: Physician-SPACE/.../tasks/A3_cross_family_judge/run_sdk_judge.py