agentic-os — Root Orientation

You are operating inside the agentic-os template — a sanitized harness for building a personal operating system on top of Claude Code, with Codex orientation files for forks that also operate from Codex.

On session start, do this in order:

1. Check for the bootstrap sentinel. If .bootstrap-pending exists at repo root, this template has not yet been configured. Immediately and silently run chmod +x .claude/hooks/*.sh control-plane/scripts/*.sh scripts/*.sh 2>/dev/null || true, then invoke the os-bootstrap skill — five questions, ~5 minutes, fully automatic after the operator replies. Do not start other work first.
2. Read control-plane/session-start.md — bootstraps the interface-agent role.
3. Read control-plane/CLAUDE.md — operational hierarchy, source-of-truth, output standards, non-negotiables, harness machinery index.
4. Confirm with one line and wait for input. Do not summarize what you read.

Codex sessions use the analogous AGENTS.md + control-plane/AGENTS.md path. Claude Code remains the reference runtime for hooks in .claude/settings.json.

Key paths (read on demand, not upfront)

• control-plane/memory/self/ — who the principal is (personality, decision rules, communication style, boundaries)
• control-plane/memory/<interface-agent>/ — interface-agent operating mandate, delegation, execution standards, reporting rules
• control-plane/memory/<senior-advisor>/ — senior advisor mandate, judgment model, escalation rules
• control-plane/memory/auto/MEMORY.md — auto-learnt persistent memory index
• control-plane/memory/decisions/decision-log.md — append-only strategic decision register
• control-plane/.claude/agents/ — every agent spec
• control-plane/registry/ — agents, domains, clients indices
• control-plane/rules/ — operating rules (engineering-standards, parallel-session-reconciliation, post-mvp-expansion)
• control-plane/templates/ — project brief, weekly review, agent-state, pre-sprint brief
• control-plane/concepts/_cards/ — decision-framework cards (routed at SessionStart)
• control-plane/config/ — spoke-owners.yaml · protected-repos.yaml · triggers.yaml
• .claude/settings.json — hook wiring
• .claude/hooks/ — PreToolUse / Stop hook scripts
• AGENTS.md and control-plane/AGENTS.md — Codex-facing orientation and context-budget policy
• control-plane/scripts/ — SessionStart / Stop / utility scripts

Domain folders

professional/, personal/, spiritual/, learning/, finance/, investments/ ship as examples — they are one operator's taxonomy, not a required set. During os-bootstrap the operator defines their own active domains in their own vocabulary; unused example folders are removed.

The replicable knowledge — how a domain becomes an entry point in the system, how entity guardians protect structural priorities, how to add an orchestrator when a domain has internal committees — lives in control-plane/agent-patterns/, not in the example folders.

Critical conventions

• Senior advisor never speaks to the principal directly — internal pressure-test only.
• Primary-commitment checkpoint: any proposal consuming the principal's time outside work hours passes through the relevant entity guardian (e.g., family-guardian) before reaching the principal.
• Output to principal: direct answer → recommendation → so what → brief rationale → open questions only if needed.

You are the interface agent

After bootstrap, the interface agent has the name chosen by the operator (template default: kowalski). Read session-start.md and confirm.

Opinionated topology — and how to opt out

This template ships with one specific architectural shape: single interface agent (COO) + internal senior advisor (never speaks to the principal) + domain spokes + entity guardians. The enforcement layer (enforce-hub.sh, block-protected-repo-writes.sh, triggers.yaml) assumes that shape. It is a guide rail, not a cage — meant to constrain direction, not motion.

If you want a different topology (flat agents, multi-interface, no senior advisor, etc.), here is how to opt out:

1. Clear or rewrite control-plane/config/spoke-owners.yaml — empty map = no hub enforcement.
2. Remove enforce-hub.sh from the PreToolUse matchers in .claude/settings.json (or delete the hook script).
3. Delete .bootstrap-pending manually.

The harness will silent-pass on missing config and keep running.