Everything for using and developing kvm-pilot, in one place. This folder is the source of truth; the project wiki is an auto-generated, nicely formatted mirror of these pages (built by .github/workflows/wiki-sync.yml on every push to main — edit the files here, never the wiki directly). The sections below mirror the navigation manifest (PAGES in build_wiki.py); CI fails if this index, llms.txt, and the wiki sidebar drift apart (#221). Adding a doc? See "Adding a doc" in Contributing.

• New here? Drive a KVM from your agent → Getting started
• Overview, install, quickstart → README
• Something's broken? → Troubleshooting & FAQ
• Release history → CHANGELOG
• AI agents: the repo root carries an llms.txt doc map (per llmstxt.org).

Task-oriented: how to accomplish something.

• Intel AMT onboarding runbook — the operator/agent guide to bringing an AMT box online: expectations, the ordered bring-up steps (provision → healthcheck → enable listeners), the ME-firmware-update hazard, and symptom→fix troubleshooting.
• Unattended Linux installs — prefer text mode + SSH over driving a graphical installer via KVM HID: the per-distro boot-arg matrix (Anaconda inst.sshd/inst.text, d-i network-console, Subiquity autoinstall, linuxrc ssh=1) and the SSH handoff (#129).
• Remote firmware update — the GL /api/upgrade/* surface, the reliability/risk model, and the gated firmware-update command (#92).
• Troubleshooting & FAQ — symptom-first fixes: GLKVM API 404, snapshot failures, approval cancel, dark-host recovery, mouse calibration, SOL noise.

Descriptive: what exists and how it behaves.

• CLI reference — every kvm-pilot subcommand: capability required, destructive gating, key flags.
• Configuration — the config file, every KVM_PILOT_* env var, and precedence.
• Driver features — the complete per-driver capability list with per-feature reliability + testing level (honest maturity; points at the support-matrix source of truth) (#171).
• Architecture — the driver-plugin design, capability protocols, and diagrams.
• Redfish reference — the BMC driver: hypermedia navigation, auth, and firmware quirks.
• Intel AMT / vPro reference — the AMT driver: WS-Man power/boot/inventory, SOL serial, and RFB KVM-redirection (firmware-level BIOS/GRUB screenshot + HID) (#211).
• Firmware registry — the firmware-currency check, the community registry data model, and the GitHub-based single-source-of-truth + ingestion design (#80 follow-up).
• Claude skill — the bundled skill for driving kvm-pilot from Claude: the core rules plus a map of when to read each playbook. Its playbooks (read at need-time; also re-served mid-session by the MCP doctrine tool, #222): interfaces — recovery — setup & gates — Linux installs — target context — Python library.
• MCP server — the bundled Model Context Protocol server (kvm-pilot-mcp).
• Hardware-Compatibility list — generated from the community run ledger: what has actually been exercised live, per device + firmware + capability.

Procedures executed against real hardware.

• Hardware reliability test plan — the reusable full-fleet sweep procedure: exercise every CLI + MCP function across device states, cross-check each result against independent ground truth to catch false reports, and feed the support matrix (#172).
• Hardware test plan: iLO / iDRAC — quick-execute runbook to validate the Redfish driver (boot-device + power) against an HPE DL380 G9 (iLO4) and a Dell R710 (iDRAC6, no-Redfish → IPMI) (#200/#29).

Decisions and RFCs, not how-tos.

• Design decisions — the "looks wrong but is intentional" record, newest first.
• Reflexes (RFC) — the post-GA edge-autonomy playbook runner: act locally on known steps, escalate surprises to the agent (#117).

• Contributing — dev setup, the pre-PR checklist, and engineering principles.
• Writing a first-class driver — the procedural guide for adding a new device driver/plugin to the fleet-wide quality bar: capabilities, the five test layers, every doc surface, and the reliability/maturity evidence chain, with a Definition-of-Done checklist (#211).
• Security policy — reporting a vulnerability and operational guidance.

Session-level review narratives — what was reviewed, how, what was found, and what changed. Individual judgment calls live in decisions.md; these are the stories around them.

• 2026-07-01 deep review — the top-to-bottom multi-agent review: 86 verified findings, 10 same-day fix commits, issues #37–#65, and the milestones that came out of it.
• 2026-07-03 RM1PE firmware + encoder — the first real-hardware run: the firmware-update live no-op (#94/#95) and the H.264 encoder wedge behind the snapshot 503s.
• 2026-07-08 a13→a14 performance — the first performance baseline: honest before/after library-level latencies for the a14 interface router + persistent SSH, measured on the real fleet.
• 2026-07-08 a13→a14 end-to-end tasks — whole operator tasks timed start→result: persistent SSH's ~2.2–2.5× health-check win is real; every interactive KVM-plane task is flat.