Instruction file imported from SFenton/ha-sfenton-react-dash (.github/instructions/simulated-user-panel.instructions.md). Copyright stays with the author.

Simulated user panel instructions

• Keep routine/standard coordination pinned to gpt-5.6-sol, reasoning effort medium, and context tier default. Sol max/long_context is conditional on an evidence-bound panel-deep-safety-adjudication or panel-material-disagreement-adjudication receipt.
• Keep all operator-requested core personas in the default panel and preserve the exact model, effort, and context settings unless a change is supported by measured panel calibration.
• Keep participant agents read-only. They must not edit files, write code, read secrets, use browser tools directly, call Home Assistant services, mutate external systems, or spawn agents. Launch them through isolated non-interactive Copilot sessions with tool filtering and explicit deny rules; block the run instead of falling back to shell-capable participant agents.
• Keep interaction mock-first. Live Home Assistant is observation-only inside a panel run; participant-driven device actuation is not supported. Require the no-proxy eval preview config, localhost URL, fresh browser context, network audit, and window.__mockHass preflight before guided tasks. Do not use npm run dev:mock or the repository's default preview config for panel interaction because both inherit live proxy routes.
• Keep persona definitions behavioral and explicitly synthetic. Do not claim demographic authenticity, equate model cost with human intelligence, or generalize capabilities or preferences to real groups. Require interface properties and mechanisms instead of demographic capability claims.
• Keep independent first passes, evidence-backed consensus, adversarial cross-critique, calibration of model-sensitive findings, and trigger-scaled Sol adjudication.
• Keep participant output strict JSON with exact allowlisted surface ids, non-empty evidence for every finding, lens attestation, severity justification, and explicit out-of-scope notes.
• Keep severity anchored. Unknown evidence gaps cannot be high or blocker, and missing packet evidence is not a product blocker.
• Keep calibration and adversarial review cross-vendor when practical, and keep specialist selection deterministic with omitted-trigger reporting.
• Keep implementation closed unless the original invocation explicitly asks for changes. Follow-on work resolves through the hierarchical project pipeline and does not imply permission for Home Assistant mutation, commit, push, deployment, or restart.
• Keep reports privacy-sanitized. Never print or persist tokens, credentials, live camera images, household task text, names, URLs, coordinates, or volatile entity attributes.
• Do not add package scripts, dependencies, or app source merely to run this skill. Keep eval assets under the skill's evals/ directory and artifacts under ignored artifacts/.