Pre Flight Tooling Validation

Skill: /pre-flight-tooling-validation

Model preference: #investigate (per horizon_aios_model_prefs.md; overridable by a prompt directive).

Coordinate a fleet of investigation agents — one per major platform — that each verify the AIOS, as it would arrive from a fresh GitHub clone, contains all the tooling required to run its full lifecycle. If every capability is covered, emit a single self-contained prompt per platform that a freshly installed Claude Code (running with Administrator / sudo) can execute to test-run those capabilities on the target machine.

This is a read-only validation of the repo's tooling surface on the dev machine. It does not install, provision, switch, or delete anything here — the actual exercising happens on the target machine via the emitted prompt.

When to invoke

• "run a pre-flight tooling check", "validate the AIOS tooling before I test on another machine", "are we ready to dog-food a fresh clone".
• Before shipping a clone to a clean test box (pairs with getting_started/lifecycle_test.md).

Platforms

The full design fans out one agent per major platform: Windows, Linux, macOS. This prototype run targets Windows only — dispatch only the Windows agent. The skill is structured so Linux/macOS agents drop in unchanged once those platforms are promoted past prototype.

The five capabilities each agent must validate

For its platform, an agent confirms the repo ships working tooling for each of:

"Validate" means: the file exists, is non-empty, and exposes the expected interface (subcommand / flag / argparse option) — grep the source for the flag or add_argument, or run --help where safe and read-only. It does not mean executing the install/provision/switch/delete here. Anything not positively verified is a FAIL (no false greens).

Execution protocol

1. Resolve the repo root

Use $HORIZON_SYSTEM / $HORIZON_ROOT. Never hardcode paths.

2. Dispatch one investigation agent per in-scope platform

For the prototype, dispatch one general-purpose subagent for Windows (use the Agent tool, subagent_type: general-purpose). Subagents start cold — give each the capability table above, the repo root, its platform, and these instructions:

For platform <P>, verify each of the five capabilities by locating its tooling under $HORIZON_SYSTEM and confirming the expected interface (flag / subcommand / add_argument) via Read/Grep, or a read-only --help. Return a verdict table: capability → PASS/FAIL → positive evidence (file:line, matched flag, or --help line). Do not install, provision, switch, or delete anything. Do not commit. If a capability's tooling is missing or lacks its interface, mark FAIL with what's absent.

3. Gate on the results

• If ANY capability on ANY in-scope platform is FAIL → stop. Do not write any test prompt. Notify the user: which platform, which capability, and the missing/incomplete tool. That is the skill's terminal output.
• If ALL capabilities PASS on all in-scope platforms → proceed to step 4.

4. Emit one admin/sudo test-run prompt per platform

For each platform that fully passed, write a single self-contained prompt that a freshly installed Claude Code with Administrator/sudo on the target machine can run end-to-end. Write it to the canonical path $HORIZON_DOCS/development_tools/<platform>_install_switch_uninstall_test_prompt.md (e.g. windows_install_switch_uninstall_test_prompt.md) and register it in documentation/index.md in the same commit (CC-G4). These prompts are durable, shared lifecycle-test artifacts — useful to the owner and any future contributors — not throwaway scratch. Each prompt must:

• State it requires elevation and a clean machine; never run on the dev box.
• Not prescribe install locations. Open with a CONFIGURE step that ASKS the operator which directory to use for AIOS install #1 and #2 (and the clone URL + backup remote), waits for answers, and validates them (the two roots must differ, must not be an existing AIOS root or owner checkout, parents writable). Never default or hardcode a root path (e.g. C:\devroot).
• Walk the full lifecycle in order: clone → bootstrap → create a brain → register + create a second AIOS in another dir → switch back and forth → back up user data to the operator's own remote → uninstall / delete (brain then AIOS) → verify clean.
• Include a dev-environment leakage / hardcoded-path scan right after install (and re-run it after the switch step). Since the test box installs to an operator-chosen root that is NOT the canonical dev root (C:\devroot), any occurrence of a build/dev path or owner identifier in a GENERATED artifact (~/.horizon active_env + registry, the harness wrapper in ~/.claude, generated instance *.conf) is a hardcoded-path leak = FAIL. Scope the scan to generated artifacts/instance config, NOT docs (which legitimately cite the canonical dev root). This is a first-class check because the prompt is run by a development tool whose job includes catching dev-env leakage.
• Use the platform's real commands (PowerShell for Windows; sudo bash for Unix), --dry-run before each destructive step, and reference system/uninstall.md for the authoritative verification checklist.
• Tell the agent to report PASS/FAIL per lifecycle step and stop on first failure.

5. Report

Summarize per platform: the verdict table, the gate decision, and (on success) the emitted prompt(s). Keep the dev-machine side read-only throughout.

Notes for the executing agent

• Fail closed. Missing tooling or an unverifiable interface is FAIL, not "probably fine". The whole point is to catch a gap before it wastes a test-machine run.
• Capability 3 ("new AIOS in another directory") has no dedicated script by design — it is clone → bootstrap → aios register. Validate that composite path, don't fail it for lacking a one-shot tool.
• This skill is read-only on the dev machine. The emitted prompt is the only thing that mutates state, and only on the target machine.
• Keep the fleet structure even in the Windows-only prototype, so promoting Linux/macOS is just enabling their agents.
• Cross-reference getting_started/lifecycle_test.md (operator runbook) and system/uninstall.md (authoritative teardown + verification) in emitted prompts.