AGENT SKILL

harness-curator

Reads the friction tags that agents filed in the harness-friction wing, clusters them, and opens one evidence-backed feedback issue per cluster on the repositories named in the components' provenance, after a health check of the wing. Also sweeps transcripts for untagged frictions (`--deep`) and runs on a schedule. Use when the user explicitly asks to curate, sweep or cluster frictions, to open the feedback issues, to check the health of the friction wing or to schedule the curator. Trigger terms: harness curator, curate frictions, friction wing, feedback issues, deep sweep. Not for tagging a single friction (harness-report) or for ordinary bug reports.

License 许可
Apache-2.0
所属仓库
crewrig/crewrig
最近更新
2026-11-15
Stars
★ 24
Forks
6

01 前言 Front Matter

name
harness-curator
type
skill
claude
{ "allowed-tools": [ "Read", "Bash", "Grep", "Glob" ], "user-invocable": true }
license
Apache-2.0
metadata
{ "provenance": { "version": "2.1.0", "feedback": "${CANONICAL_REPO}", "canonical": "${CANONICAL_REPO}" } }
description
Reads the friction tags that agents filed in the harness-friction wing, clusters them, and opens one evidence-backed feedback issue per cluster on the repositories named in the components' provenance, after a health check of the wing. Also sweeps transcripts for untagged frictions (`--deep`) and runs on a schedule. Use when the user explicitly asks to curate, sweep or cluster frictions, to open the feedback issues, to check the health of the friction wing or to schedule the curator. Trigger terms: harness curator, curate frictions, friction wing, feedback issues, deep sweep. Not for tagging a single friction (harness-report) or for ordinary bug reports.
compatibility
Requires Node.js 24 or later, a forge CLI (setup-labels.ts uses gh; --apply selects gh, glab, or tea per the target repository's forge), and the mempalace Python package (pipx install 'mempalace>=3.6.0,<3.7').

02 资源文件 13

03 SKILL.md 正文

Harness Curator

Closes the feedback loop: the frictions that agents tagged during real work become one descriptive issue per cluster, on the repository that owns the offending component. The issue carries evidence, not a fix; the fix lands later as a human-authored change that closes it.

When to run it

Only on an explicit request: the user runs /harness-curate, asks what frictions the crew has accumulated, asks for a deep sweep, or a schedule triggers it. During unrelated work, reaching for this skill means you are off task.

The script

Reading the wing through MCP would take thousands of calls, so a bundled script reads it with the MemPalace Python library, parses every FRICTION: payload and clusters them. Run it from the skill directory (or task harness-curate -- <options>); node scripts/curate.ts --help lists every option.

Exit codes of curate.ts: 0 ok, 1 a usage error or an unhealthy --check, 2 python or MemPalace not found (or the backend failed), 3 --apply without an authenticated forge CLI. A usage error prints Error: or Unknown option on stderr, which tells it from an unhealthy wing.

Test mode: --from-stdin reads a JSON list of drawers (drawer_id, room, content) instead of MemPalace, so a dry run needs no server and no forge. assets/sample-frictions.json is a sample.

Step 1: dry run and health gate

node scripts/curate.ts --check

It runs a dry run and prints healthy or unhealthy: exit 0 healthy, exit 1 when skipped_malformed exceeds 25 percent of the drawers (--max-malformed-ratio moves the limit) or routing_failures is above zero. An unhealthy wing means the filing source is rotten (payloads missing writer_agent or evidence) or a friction names no canonical repository; fix that before opening issues, so that the issues are not opened blind or partial. Without --check, node scripts/curate.ts --dry-run (the default) prints the full JSON report: stats, clusters (each with cluster_key, target_repo, title, body, labels, frictions), skipped and routing_failures. Spot-check at least one body to see that it reads sensibly.

Step 2: open the issues

Two equivalent paths; either way open one issue per cluster, because independent clusters deserve independent triage.

  • The script: node scripts/curate.ts --apply opens one issue per cluster through the forge CLI picked from the cluster's repository host (gh, glab or tea), with the three labels from the report.
  • By hand: iterate the report and run gh issue create --repo <owner>/<repo> --title <title> --body <body> --label <label> per cluster (the glab and tea equivalents on GitLab and Gitea). Choose this when you want to enrich a body first. Forge access is CLI-only; the framework ships no forge MCP server.

What the script decides

  • A cluster is proposed at 2 or more reports (--threshold), or at 1 report of severity: high, which bypasses the threshold by design.
  • The cluster key is the subcategory, else the room. The target repository is the most frequent canonical in the cluster (--target-repo forces one).
  • Labels are harness-feedback, room:<dominant room> and severity:<worst>. The title carries a Gitmoji per dominant room (tool 🐛, process 📝, behavior 🚸, format 🎨, prompt 💬, other 🔧).
  • A cluster with no resolvable canonical is a routing failure: it is reported, never opened blind.
  • --max-issues N keeps the N highest-ranked clusters (severity, then size, then key) and --dedup skips a cluster whose key already has an open harness-feedback issue.

Labels must exist first

--label fails on a label the target repository does not have. Create the nine labels once per receiving repository; the script is idempotent and safe to re-run after a vocabulary change:

node scripts/setup-labels.ts --repo <owner/repo>        # add --dry-run to preview

A missing label shows up as a failures: entry of the run summary; create it and retry.

Deep mode

node scripts/curate.ts --deep [--deep-window 1000]

It scans the most recent transcript drawers (default 500) with a fixed set of heuristic patterns and prints a Markdown checklist of candidates grouped by pattern, for a human to triage. It never opens issues and excludes --apply. To promote a candidate, tag it with the harness-report skill; it joins the next sweep.

Scheduled mode

node scripts/schedule-curator.ts installs a weekly Monday 09:00 run of node scripts/curate.ts --apply --dedup --max-issues 5 (launchd on macOS, cron on Linux; re-running replaces the entry; --dry-run previews, --uninstall removes it). It runs on the maintainer's machine, never in CI, because the MemPalace state is local. Run the same command by hand right after filing a severity: high friction.

Step 3: report the run

node scripts/curate.ts --summary

prints the counts as bullets: frictions read and skipped, clusters formed, proposed and parked, routing failures and the proposed clusters. Post it, adding the issues opened with their links. Even a zero-issue run is worth reporting: it tells the user the wing is healthy.

Friction

If the curation led to a bad cluster, a wrong target or an unactionable issue, tag it with the harness-report skill; the shared rule is in references/shared-friction-reporting.md.

04 同仓库更多 Skills