ulw-report

name	ulw-report
description	Render a markdown digest of recent harness activity — sessions, gate fires, top reviewers, classifier misfires, Serendipity catches, and finding/wave outcomes — over a chosen time window. Use to answer "is the harness actually helping me?". Pass `--share` for a privacy-safe shareable card (numbers + distributions only; no prompt text or free-text reasons).

ULW Report

Parse any argument passed with the skill invocation (stored in $ARGUMENTS). Treat the arguments case-insensitively.

Window selection:

• last — the most recent session only
• week — sessions in the last 7 days (default when no argument given)
• month — sessions in the last 30 days
• all — every available row across the cross-session aggregates

Modifiers:

• --share (v1.31.0) — privacy-safe digest. Sessions / quality-gate-blocks / specialist-dispatches / Serendipity-fires counts plus the top-10 gate-name distribution. Suppresses ALL free-text fields (prompt previews, gate reason payloads, Serendipity fix text). Suitable for posting to Slack, PRs, or social. Combine with the window argument: /ulw-report week --share.
• --field-shape-audit (Bug B post-mortem hardening) — runs a runtime field-shape sanity check over ~/.claude/quality-pack/gate_events.jsonl and reports any rows whose typed detail fields fail their declared shape contract. Catches Bug B-class leaks (positional misalignment landing prompt-text fragments into typed fields) directly from runtime artifacts instead of relying on source-only review. Bypasses the verbose body — emits a focused audit report and exits 1 if violations are present so it can be wired into CI / pre-release checks. Combine with the window argument: /ulw-report all --field-shape-audit. Detail field invariants checked: state-corruption.archive_path (path-shaped), state-corruption.recovered_ts (Unix epoch), wave-plan.{wave_idx,wave_total,finding_count} (non-negative int), finding-status.{finding_id,finding_status} (F-NN regex + enum), bias-defense.directive (enum). Other gates today have no typed detail invariants and pass through bounded only by the per-event 1024-char value cap (256 for state-corruption).
• --merge <dir> (v1.48-pre, repeatable; also --merge=<dir>) — fold ANOTHER MACHINE's quality-pack ledgers into this report's view. <dir> may be a quality-pack dir itself, an extracted archive root containing .claude/quality-pack, or a dir with a quality-pack/ child. Read-only — sources are never written; merging happens in temp copies cleaned on exit. Rows lacking host (pre-v1.48) are labeled at read time: local rows get this machine's hostname, external rows the merge dir's basename. Composes with --sweep. Banner lines report the merged dir labels and a Sessions by host breakdown. Motivated by the 2026-07-04 sampling error: per-machine ledgers with no aggregation let a one-machine audit ("12 sessions ever") masquerade as the full record (728 sessions across three machines).
• --sweep (v1.36.0) — fold currently-active session dirs (under ~/.claude/quality-pack/state/) into the in-memory view this report uses. Closes the gap where /ulw-report run during an active session missed that session's gate events because session_summary.jsonl / gate_events.jsonl only populate at the daily TTL sweep. Read-only: never writes to the cross-session ledger and never claims (deletes) the source dirs. Synthesizes per-session session_summary rows using the same jq formula as sweep_stale_sessions and tags them _live: true for downstream consumers. Combine with the window argument: /ulw-report --sweep last (peek at the active session) or /ulw-report --sweep week (active session merged into 7-day rollup). Banner line [--sweep] Including N active session(s) in this view (read-only; ledger not modified). prefaces the report so the user knows the cross-session aggregate was extended on the fly.

Run the report:

bash ~/.claude/skills/autowork/scripts/show-report.sh ${ARGUMENTS:-week}

Display the output as-is. The report ends with a ## Patterns to consider section that surfaces 2–3 actionable interpretations from the data (high gate-fire density, skip rate, classifier misfires, archetype convergence, reviewer find rate); when no thresholds trip, the section emits a clean "no patterns to call out" line. The interpretations are heuristic, not assertions — treat them as starting points the user can accept, refine, or ignore. Do not paraphrase or re-summarize above the script's output — the report itself is the deliverable.

The report joins data from ~/.claude/quality-pack/:

• session_summary.jsonl — per-session outcome rows (gates fired, skips, serendipity, finding/wave outcomes)
• serendipity-log.jsonl — Serendipity Rule applications
• classifier_misfires.jsonl — classifier misfire annotations
• timing.jsonl — per-session walltime, token categories, and bounded role/model/native-dispatch attribution
• agent-metrics.json — reviewer invocation counts and verdicts
• defect-patterns.json — defect category histogram

If a section comes back empty, it means no data has accumulated for that window — point the user at /ulw-status for an in-flight session view, or suggest letting more sessions run.