harness-digestlisted

# Harness Digest (Run Telemetry → Harness-Improvement Proposals) Apply the quality-digest principle — cluster recurring evidence, propose promotions, never change anything automatically — to **the harness itself**. Input is one or more finished harnie:dev runs; output is a ranked proposal list the user accepts or rejects. This is the mechanism that turns each run's measured waste into the next version's contract, without self-modifying-harness risk: **this skill writes no files and changes no configuration.** ## Input (all existing artifacts; no new instrumentation) Read via sanctioned paths only (the Read tool; `loop.mjs export` for anything a shell would need): - `execution.json` — per-task timestamps, `codexCalls`, `watchdogExtensions`, `threadRebindings`. - Review-unit state — each `review/<unit>/ledger.json` + `state.json` (round counts, blocking-issue trajectory), and each task's `review-archive/t<id>/` (unit reviews archived at integration). - Delta sidecars (`*.json` next to each `delta.patch`) — `changedCount` per round: the unit's real size. - `design/errata.md` (`errata-list` output) — where the approved design was wrong. - `notepad.md` — recorded path choices (serial-path rationales), handoffs, incident notes. - Optional: session JSONL token stats when the caller provides them (per-call input sizes, cache reads). ## Procedure 1. **Unit cost profile.** For every review unit: rounds × changedCount × model tier (+ watchdog extensions, rebindings). Flag the dist