dispatch-token-auditlisted

# Dispatch Token Audit This skill parses recent Claude session transcripts and emits a ranked report of token-reduction opportunities. It is report-only: it creates no GitHub issues, writes no control artifacts, and edits no workflow files — the user decides what to act on from the report. (The learned phase→model routing policy this skill used to write was retired in #2872: the phase orchestrator is now always Sonnet, and Opus tiering lives at the `agent()`/subagent layer inside each phase's Workflow, not in an audit-written policy.) Two cost figures appear in the output: - **`price_proxy_usd`** — a uniform Opus-list-price rate applied to every token regardless of the actual model. Holding price constant isolates token count, so this figure ranks opportunities by relative magnitude. It is **not** the actual bill. - **`cost_usd`** — the truthful per-model bill. Each model is priced at its real rate (Sonnet, Haiku, or Opus), so this figure measures real dollars and shows the actual savings from model-routing decisions. Ranking (step 5) stays on `price_proxy_usd`. `cost_usd` is reported alongside it to show the real bill. 1. **Parse the window argument.** The skill accepts an optional argument of the form `Nd` (e.g. `2d`, `14d`). Parse it as follows: - If `ARGUMENTS` matches `^[0-9]+d$`, strip the trailing `d` and use that integer as `N`. - Otherwise default to `N=7`. - Reject any value where `N` is 0 or non-numeric with a clear message: `error: window must be a po