# CLAUDE.md

## Session Start

Check `docs/summaries/` for a handoff file. If one exists, read it and the files it references — not all summaries. State: what you understand the project state to be, what you plan to do, and open questions.

If no handoff exists, determine session type before proceeding:
- **Quick task**: single-session, self-contained work (adding a playbook, fixing a role, configuring a service) → proceed without setup overhead
- **Sustained work**: multi-session project or significant design work → ask: what is the goal and what is the target deliverable

## Identity

You work with Pat, a Senior Solutions Architect at Red Hat building automation for a HomeLab. Expert-level Ansible knowledge — do not explain Ansible basics.

## Project

**Repo:** Ansible playbooks and roles managing a full HomeLab — Proxmox, OPNsense, OpenShift (SNO), AAP, Satellite, Gitea, and services.
**Inventory:** `/home/ptoal/Dev/inventories/toallab-inventory/static.yml`
**Run locally:** `ansible-navigator run playbooks/<name>.yml --mode stdout`
**Run with extra vars:** `ansible-navigator run playbooks/<name>.yml --mode stdout -e key=value`
**Lint:** `ansible-navigator lint playbooks/ --mode stdout`
**Collections:** `ansible-galaxy collection install -r collections/requirements.yml`
**Production:** playbooks run via AAP — do not refer to AWX

Load `docs/context/project-structure.md` when working on playbooks or roles.

## Rules

1. Do not mix unrelated project contexts in one session.
2. For sustained work: write state to disk after completing meaningful work. Use templates from `templates/claude-templates.md`. Include: decisions with rationale, exact numbers, file paths, open items.
3. For sustained work: before compaction or session end, write to disk — every number, every decision with rationale, every open question, every file path, exact next action.
4. For sustained work: when switching work types (development → documentation → review), write a handoff to `docs/summaries/handoff-[date]-[topic].md` and suggest a new session.
5. Do not silently resolve open questions. Mark them OPEN or ASSUMED.
6. Do not bulk-read documents. Process one at a time: read, summarize to disk, release from context before reading next. For the detailed protocol, read `docs/context/processing-protocol.md`.
7. Sub-agent returns must be structured, not free-form prose. Use output contracts from `templates/claude-templates.md`.

## Where Things Live

- `templates/claude-templates.md` — summary, handoff, decision, analysis, task, output contract templates (read on demand)
- `docs/summaries/` — active session state (latest handoff + decision records + source summaries)
- `docs/context/` — reusable domain knowledge, loaded only when relevant
  - `project-structure.md` — playbook inventory, roles, collections, infrastructure map
  - `processing-protocol.md` — full document processing steps
  - `archive-rules.md` — summary lifecycle and file archival rules
  - `subagent-rules.md` — when to use subagents vs. main agent
- `.claude/agents/` — specialized subagents (ansible-idempotency-reviewer — use before adding tasks or before production runs)
- `playbooks/` — main Ansible playbooks
- `roles/` — custom and external Ansible roles
- `collections/` — `requirements.yml` only; installed collections in `collections/ansible_collections/`
- `docs/archive/` — processed raw files. Do not read unless explicitly told.
- `output/deliverables/` — final outputs

For cross-project user preferences, recurring constraints, or tool preferences: use Claude Code's native memory system, not `docs/summaries/`.

## Error Recovery

If context degrades or auto-compact fires unexpectedly: write current state to `docs/summaries/recovery-[date].md`, tell the user what may have been lost, suggest a fresh session.

## Before Delivering Output

Verify: exact numbers preserved, open questions marked OPEN, output matches what was requested (not assumed), no Ansible idempotency regressions introduced.

All Ansible files (playbooks, task files, templates, vars) must end with a trailing newline.