audit
AIOS Audit
Check whether this AIOS can find the right information and do useful work reliably, and identify changes that would make the user's work easier. Score Context, Connections, Capabilities, and Cadence out of 25 each, using rubric.md. This is rubric v2. Label the number verified operational reliability, not overall usefulness or percentage of work the AIOS can do. A folder, installed skill, API key, or confident claim is not proof of a working system.
Run locally in the current project. Do not launch subagents unless the user requests them. Audit first; /link repairs a route, and /level-up helps close one meaningful gap. Never inflate or deliberately depress a score to encourage another run. Improvement comes from better evidence and behavior.
Every audit automatically saves its report in the audited project's audits/ folder. Read history.md for prior-report selection, finding transitions, comparison rules, and safe persistence. Saving the local audit record is part of this skill, not a separate approval step. An explicit user request not to save overrides this default. The inspected system remains unchanged.
1. Establish scope and evidence
- Get the current date and project root. Read applicable
AGENTS.md,CLAUDE.md, and local overrides; follow the runtime's instruction hierarchy. Read both manuals when both exist. Check for conflicting routing and promised synchronization, without assuming distinct runtime instructions must be identical. - Load the relevant prior AIOS report and carried finding ledger using history.md, before selecting probes. Recheck important prior findings alongside current priorities; do not assume earlier results remain true. Identify the comparison baseline and scope changes before scoring.
- Use the manual's routing rules first. Read its project and knowledge indexes before listing their immediate child folders. Inspect only relevant scoped manuals. Skip dependencies, caches, generated output, and archives unless a route points there or an archived/current ambiguity needs checking.
- Identify the user's role, main objective, current priorities, and up to three important workflows from actual context. Unfilled template placeholders do not count. Do not require business revenue data from someone whose work has no revenue function.
- Inspect the supported skill/agent registries, connection references, and actual scheduler configuration/run evidence. Accept scripts, MCPs, exports, local data, and equivalent mechanisms equally. Count inventories for orientation only.
- Record each finding as verified, documented but unverified, missing, stale, or conflicting, with a file/line, source URL, or dated run reference. File modification times and last-checked labels alone do not prove a successful refresh or execution.
- Start with targeted reads. If a time budget or local page limit prevents verification, mark it unverified and state coverage. Never award full points because there was no time to check. Do not claim an exhaustive audit from a sample.