agentic-life-os: Instructions file for Codex

Agent guidance

This repo contains a portable Agent Skills based personal advisor OS.

Rules

• Always use the grill-me review pattern for non-trivial product/design changes: interview the plan one question at a time, recommend an answer for each question, inspect the repo instead of asking when code/docs can answer it, and walk the decision tree until the design is clear enough to implement. Keep the questions about the public repo/product, not the user's personal install.
• Keep the repo public-safe: no personal data, real chat IDs, secrets, tokens, local logs, or private runtime config.
• Private state belongs in $HOME/.life-os by default, or LIFEOS_DATA_DIR when explicitly set. Never commit private state files.
• Life OS is a helper/coordination layer over the active runtime, not the owner of the user's life data. By default, real data stays in Hermes/OpenClaw/external sources; each Life OS skill stores its own source decisions, pointers, access notes, operational state, caches, and Life-OS-specific preferences.
• Runtime credentials, delivery routing, memory, vault, mail/calendar credentials, tasks, contacts, birthdays, and cron ownership belong to the runtime or external source unless the user explicitly created a Life OS note/preference/state item.
• Prefer small skills and lazy loading over one giant prompt.
• Validate Markdown, JSON, YAML, frontmatter, and schemas before committing.
• Add runtime-specific instructions under runtimes/ only after checking the actual runtime docs or code.
• Treat Hermes and OpenClaw as first-class supported runtimes. For any workflow that mentions skill availability, install/register paths, scheduling, delivery, or global skill visibility, document both runtime paths or explicitly mark the missing runtime path as pending with the fact that must be verified.
• Keep scripts strictly mechanical and deterministic: validation, schema/frontmatter checks, public-safety scans, boring file-layout/state creation, and smoke tests. Do not encode semantic product decisions in scripts.
• Runtime discovery and integration decisions belong to the agent playbooks, not helper scripts. The LLM should inspect the active runtime with runtime-native tools/docs, then decide where tasks, crons/schedules, reminders, memory, delivery routes, birthdays, contacts, or other systems live, and record each domain source decision in the owning skill's data file. Ask the user before changing any runtime-owned system. Generic helpers may report configured pointers and runtime inventory freshness, but must not infer arbitrary runtimes from filesystem/command heuristics.
• When tempted to add a script, first classify it: mechanical guardrail or state helper is acceptable; heuristic discovery, policy choice, prioritization, migration strategy, routing, or integration ownership is agent semantics and should be written as skill instructions instead.
• Treat skills/life-os/references/data-integrity.md as the universal runtime contract. Every umbrella or directly invoked subskill must load and preserve its source provenance and uncertainty, correction history, explicit missing-value semantics, read-before-write plus authoritative readback, native concurrency protection where available, and verified backup/recovery before risky data operations. Domain guidance may be stricter but must not weaken that contract or create a competing source of truth.