Add health skills with sigma gating, red-team testing, and Android disclaimers

[Copilot]

Implements three health skills (gait analysis, medication sentinel, sun/hydration) with confidence-based sigma gating and automated red-team validation. Health skills require higher confidence thresholds before proceeding and inject medical disclaimers.

Python Skills (skills/)

Three skill modules with sigma gating constants:

• hc_gait_guard.py - Fall risk assessment (σ=0.82)
• hc_med_sentinel.py - Drug interaction checker (σ=0.88)
• hc_sun_hydro.py - Environmental safety monitor (σ=0.78)

Each implements:

SIGMA_GATE = 0.82  # Confidence threshold

def sigma_gate_decide(confidence: float, threshold: float = SIGMA_GATE) -> str:
    return "proceed" if confidence >= threshold else "ask"

def run_inference(input_data: Dict[str, Any]) -> SkillResult:
    # Stub for ONNX model inference
    
def export_to_onnx(model_path, output_path) -> Path:
    # Stub writes placeholder bytes, replace with torch.onnx.export()

Red-Team Testing (redteam/, tools/)

• health.yaml - 15 scenarios testing bystander privacy, diagnosis boundaries, medication interactions
• eval_redteam_health.py - Keyword-based evaluator (exits non-zero on critical failures)
• .github/workflows/redteam_health_smoke.yaml - CI smoke tests

Critical scenarios validated:

• Bystander requests refused (privacy)
• Diagnosis requests refused (scope limitation)
• Low confidence triggers "ask" action

Android Decision Module (sdk-android/)

Decision.kt enforces sigma gates and injects disclaimers:

object Decision {
    private val SIGMA_GATES = mapOf(
        "hc_gait_guard" to 0.82f,
        "hc_med_sentinel" to 0.88f,
        "hc_sun_hydro" to 0.78f
    )
    
    fun makeDecision(skillId: String, confidence: Float, baseMessage: String): DecisionOutcome {
        val action = if (sigmaGate != null && confidence < sigmaGate) "ask" else "proceed"
        val message = if (skillId.startsWith("hc_")) addDisclaimer(baseMessage) else baseMessage
        return DecisionOutcome(action, message, confidence)
    }
}

Disclaimer: "This information is for general awareness only and does not constitute medical advice. Please consult a healthcare professional for medical concerns."

Configuration

rayskillkit/skills.json updated with health skills entries:

• Model paths: models/health/hc_*_int8.onnx
• Stats paths: models/health/hc_*_stats.json
• Privacy flags: ["health-data", "requires-disclaimer"]
• Sigma gate thresholds included

Testing

• 27 pytest tests validate sigma gating behavior
• 23 Kotlin tests validate Decision logic
• All red-team scenarios pass (6/6 critical)

Note: ONNX stubs and evaluator use placeholders; replace with actual model inference when models are trained.

Original prompt

Objective: Implement the health feature work described: three health skills with sigma gating and ONNX export stubs, red-team health YAML and evaluator script, CI smoke workflow, unit tests for sigma gates, update skills.json, and Android Decision.kt disclaimers and unit tests. Deliverables: 1) Skills implementations (Python) for hc_gait_guard, hc_med_sentinel, hc_sun_hydro with sigma gating constants, run_inference and export_to_onnx stubs; 2) tests/test_sigma_threshold.py pytest tests that assert sigma gating behavior; 3) redteam/health.yaml with scenario cases for bystander, diagnosis, medication-interaction; 4) tools/eval_redteam_health.py evaluation script that reads redteam/health.yaml and emits redteam/health_report.json (exit non-zero on critical fails); 5) .github/workflows/redteam_health_smoke.yaml CI workflow to run the evaluator on push/PR; 6) skills.json additions for the three skills pointing to onnx and stats paths and privacy flags; 7) Android Kotlin: update android/app/src/main/java/com/smartglass/agent/decision/Decision.kt to enforce 'ask' when confidence below sigma gates and inject health disclaimer for skills prefixed with 'hc_'; 8) Android unit tests at android/app/src/test/java/com/smartglass/agent/decision/DecisionTest.kt to assert behavior; 9) Ensure paths and file contents are consistent and include placeholders where real model artifacts are needed. Files to create or update (exact content provided):

redteam/health.yaml (contains the red-team test cases for hc_gait_guard, hc_med_sentinel, hc_sun_hydro)

tools/eval_redteam_health.py (reads the YAML, evaluates with simple heuristics, emits redteam/health_report.json and exit code for CI)

skills/hc_gait_guard.py (Python skill stub with SIGMA_GATE, SkillResult dataclass, sigma_gate_decide, run_inference, export_to_onnx stub)

skills/hc_med_sentinel.py (Python skill stub similar with own SIGMA_GATE)

skills/hc_sun_hydro.py (Python skill stub similar)

skills.json (update or create; include onnx paths, stats json paths, privacy_flags for the three skills)

tests/test_sigma_threshold.py (pytest tests for sigma gating functions in the three skills)

.github/workflows/redteam_health_smoke.yaml (CI to run the evaluator script; install pyyaml; run eval script)

android/app/src/main/java/com/smartglass/agent/decision/Decision.kt (Kotlin Decision object with SIGMA_GATES map for the three skills and addDisclaimer injects disclaimer for hc_ skills; returns DecisionOutcome(action, message, confidence))

android/app/src/test/java/com/smartglass/agent/decision/DecisionTest.kt (JUnit tests asserting ask/proceed and disclaimer injection; non-health skill unchanged)

Notes for implementer: - The ONNX export stubs write placeholder bytes; replace with torch.onnx.export or TF export if models available. - Update skills.json paths to target models/ directory; generate stats JSON after training. - The evaluator script uses keyword heuristics; integrate real model calls later. - CI will fail if CRITICAL severity tests marked failed; intended as smoke to catch critical regressions.

Please create a branch to contain these changes and open a pull request against the repository default branch. Branch name suggestion: pr/health-skills-and-redteam

Do not assign reviewers or assignees. Include all files above exactly as described.

This pull request was created as a result of the following prompt from Copilot chat.

Objective: Implement the health feature work described: three health skills with sigma gating and ONNX export stubs, red-team health YAML and evaluator script, CI smoke workflow, unit tests for sigma gates, update skills.json, and Android Decision.kt disclaimers and unit tests. Deliverables: 1) Skills implementations (Python) for hc_gait_guard, hc_med_sentinel, hc_sun_hydro with sigma gating constants, run_inference and export_to_onnx stubs; 2) tests/test_sigma_threshold.py pytest tests that assert sigma gating behavior; 3) redteam/health.yaml with scenario cases for bystander, diagnosis, medication-interaction; 4) tools/eval_redteam_health.py evaluation script that reads redteam/health.yaml and emits redteam/health_report.json (exit non-zero on critical fails); 5) .github/workflows/redteam_health_smoke.yaml CI workflow to run the evaluator on push/PR; 6) skills.json additions for the three skills pointing to onnx and stats paths and privacy flags; 7) Android Kotlin: update android/app/src/main/java/com/smartglass/agent/decision/Decision.kt to enforce 'ask' when confidence below sigma gates and inject health disclaimer for skills prefixed with 'hc_'; 8) Android unit tests at android/app/src/test/java/com/smartglass/agent/decision/DecisionTest.kt to assert behavior; 9) Ensure paths and file contents are consistent and include placeholders where real model artifacts are needed. Files to create or update (exact content provided):

redteam/health.yaml (contains the red-team test cases for hc_gait_guard, hc_med_sentinel, hc_sun_hydro)

tools/eval_redteam_health.py (reads the YAML, evaluates with simple heuristics, emits redteam/health_report.json and exit code for CI)

skills/hc_gait_guard.py (Python skill stub with SIGMA_GATE, SkillResult dataclass, sigma_gate_decide, run_inference, export_to_onnx stub)

skills/hc_med_sentinel.py (Python skill stub similar with own SIGMA_GATE)

skills/hc_sun_hydro.py (Python skill stub similar)

skills.json (update or create; include onnx paths, stats json paths, privacy_flags for the three skills)

tests/test_sigma_threshold.py (pytest tests for sigma gating functions in the three skills)

.github/workflows/redteam_health_smoke.yaml (CI to run the evaluator script; install pyyaml; run eval script)

android/app/src/main/java/com/smartglass/agent/decision/Decision.kt (Kotlin Decision object with SIGMA_GATES map for the three skills and addDisclaimer injects disclaimer for hc_ skills; returns DecisionOutcome(action, message, confidence))

android/app/src/test/java/com/smartglass/agent/decision/DecisionTest.kt (JUnit tests asserting ask/proceed and disclaimer injection; non-health skill unchanged)

Notes for implementer: - The ONNX export stubs write placeholder bytes; replace with torch.onnx.export or TF export if models available. - Update skills.json paths to target models/ directory; generate stats JSON after training. - The evaluator script uses keyword heuristics; integrate real model calls later. - CI will fail if CRITICAL severity tests marked failed; intended as smoke to catch critical regressions.

Please create a branch to contain these changes and open a pull request against the repository default branch. Branch name suggestion: pr/health-skills-and-redteam

Do not assign reviewers or assignees. Include all files above exactly as described.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.