メニュー
Adversarially review the last 24h of multi-agent work by combining git history, GitHub issue state, generated analysis artifacts, governance tests, and duplicate-checked follow-up issue creation.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/vamseeachanta/workspace-hub --skill agent-work-adversarial-reviewコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SOC 職業分類に基づく
SKILL.md を表示中
| name | agent-work-adversarial-review |
| description | Adversarially review the last 24h of multi-agent work by combining git history, GitHub issue state, generated analysis artifacts, governance tests, and duplicate-checked follow-up issue creation. |
| version | 1.0.0 |
| category | coordination |
| tags | ["audit","adversarial-review","github","governance","artifacts","issues"] |
Use when asked to review recent work done by multiple agents across the ecosystem, especially the last 24h. This is not a normal progress summary — the goal is to find regressions, contradictions, stale claims, enforcement gaps, and missing follow-through.
Produce an evidence-backed review of recent agent work and create high-value follow-up GitHub issues without spamming duplicates.
Run:
pwd
git rev-parse --show-toplevel
git remote get-url origin
-u
gh auth status
Do not rely on only git log or only session logs.
Use:
git log --since='24 hours ago' --date=iso --pretty=format:'%h%x09%ad%x09%an%x09%s' --stat --no-merges
Also inspect:
.claude/state/session-signals/YYYY-MM-DD.jsonllogs/orchestrator/claude/session_*.jsonlgh issue list --state all --limit 30 --json number,title,state,createdAt,updatedAt,labels,author,url
Treat generated result docs as first-class review targets.
Check recent files under patterns like:
docs/plans/*/results/*.mddocs/handoffs/*.mdLook for:
Do not stop at document review. Re-run focused tests or scripts for the changed area.
When reviewing Deckhand/customer-channel behavior, include an interaction inconsistency pass using references/deckhand-interaction-inconsistency-audit.md. This pass must compare channel logs, scope/routing config, audit rows, and Claude/session claims across five axes: channel fit, domain scope, result-delivery state, engineering credibility, and live-readiness/canary evidence.
Good pattern for governance/runtime work:
uv run pytest <focused test subset> -q
Also exercise both human-facing and machine-facing entrypoints when a tool claims automation support:
--json / structured-output mode separatelyAdversarial check for governance/checker work:
Adversarial check for scheduled governance/cron wrappers:
fail vs FAIL, etc.)conformance or registry-health are definedIf one file fails in a combined run but passes alone, record it as a possible invocation-context/import-path problem rather than claiming a stable failure.
Delegate independent subreviews for parallel adversarial pressure, for example:
Ask subreviewers for:
Always search GitHub before opening follow-up items.
Use targeted searches such as:
gh issue list --state open --search '<keywords>' --limit 20
Important: distinguish exact duplicates from umbrella issues. If an umbrella exists, reference it in the new issue instead of skipping automatically.
If a previously closed issue is directly contradicted by a reproduced live failure, prefer reopening the original issue instead of creating a duplicate regression ticket.
Use this when:
Pattern:
gh issue reopen <number>
gh issue comment <number> --body-file /tmp/repro.md
Your comment should include:
Create issues for systemic gaps, not every symptom.
High-value categories:
Avoid filing noise issues unless the evidence is concrete and reproducible.
Return a concise summary with:
Good output is short but evidence-backed. Keep the detailed proof in issue bodies or internal notes; keep the user summary compact.
Write outbound email and external messages in Vamsee Achanta's voice — a subtle offer to help, never bold or rash claims. Load before drafting ANY email, LinkedIn/Collide reply, proposal note, or outreach sent under his name.
Save/publish analysis or computation results from ANY ecosystem repo to Hugging Face as a queryable, viewer-renderable dataset. Use when the user wants to "save results to hugging face", "publish dataset to HF", "hugging face data saving", "save analysis results", "hf dataset", "make results queryable", or "render via datasets-server API". Reshapes nested results into flat parquet tables, writes a dataset card with a viewer `configs:` block and provenance, applies license/public-vs-private routing, enforces a domain data-quality gate (faithful-to-source != correct), publishes to `aceengineer/<repo>-<projection>`, and verifies via the datasets-server API.
Clone, create, fork, configure, and manage GitHub repositories. Manage remotes, secrets, releases, and workflows. Works with gh CLI or falls back to git + GitHub REST API via curl.