Skip to content

Adonis0123/adonis-skills / skills/agentic-review-handoff

agentic-review-handoff

Verified

Validate pasted review findings before fixes; run same-session automatic Git review-fix-re-review with a headless Reviewer and no redundant confirmation; start fresh-eyes Git diff Review Intake; resume review-loop sessions or packets; get a DecisionConsult from another AI; or run first-principles/DDD/high-cohesion review. Requires Git and review/consult intent; ordinary implementation or commit-only requests do not start a review loop.

evals/trigger-results.json

View on GitHub
{
  "metric": "trigger accuracy (claude -p Skill tool mentions agentic-review-handoff; 1 run/query)",
  "eval_set": "skills/agentic-review-handoff/evals/trigger-eval.json (10 positive + 10 negative)",
  "public_claim": {
    "do_not_advertise": "auto-trigger 100% accuracy / full 20/20 acceptance on current description",
    "current_description": {
      "version": "3.3.4",
      "chars": 664,
      "note": "routing-only slim (~1023→664); copy-edit negative restored vs v3.3.3"
    },
    "latest_skill_only": {
      "description_version_measured": "3.3.3",
      "positives": "10/10 scenarios, 29/30 hits",
      "negatives": "UNVERIFIED (Claude session limit 429) — not acceptance",
      "v3_3_4_remeasure": "NOT_RUN_quota"
    },
    "how_to_introduce": "Say: auto-trigger is strong on positive review/handoff intents under Skill-only remeasure (10/10, 29/30); negative-case non-trigger and v3.3.4 are not fully re-verified yet. Prefer explicit skill invoke when accuracy matters."
  },
  "baseline_old_description": {
    "pass": "15/20",
    "rate": 0.75,
    "pos": "6/10",
    "neg": "9/10",
    "source": "git HEAD description (v3.1.0 auto-loop text, 945 chars)"
  },
  "final_description": {
    "pass": "19/20",
    "rate": 0.95,
    "pos": "9/10",
    "neg": "10/10",
    "chars": 952,
    "constraints": "ASCII-only, <=1024 for quick_validate"
  },
  "delta": "+20 percentage points (75% -> 95%); positives +3; negatives +1",
  "remaining_miss": "pos-first-principles-deep (DDD/first-principles without auto-loop keywords)",
  "coverage": [
    "auto loop / review-fix-re-review zero human",
    "same-session dual-AI closed loop",
    "consult",
    "sessions resume commands",
    "packet continue / PWC fix-it",
    "second pair of eyes + feedback validation",
    "negatives: ordinary impl, non-git, weekly-report, /codex:review, Grok /review"
  ],
  "v3_1_2_codex_round": {
    "description_source": "Codex advisory proposal (session 019f88cc-6118-7170-8a3f-9d3b562980f6), applied 2026-07-22",
    "change": "first-principles/DDD/high-cohesion/low-coupling promoted to capability summary + first Use-when branch; added low-coupling keyword and non-review design docs guard",
    "verification": "targeted re-eval, 3 queries, 1 run/query, claude -p --max-turns 2 stream-json Skill tool_use check, runtime copy synced via skills:install:local first",
    "results": {
      "pos-first-principles-deep": "PASS (was the only v3.1.1 failure)",
      "neg-design-doc": "PASS",
      "neg-ordinary-impl": "PASS"
    },
    "caveat": "remaining 17 queries not re-run (keywords preserved superset of v3.1.1); measurement pitfall recorded: claude -p reads .agents/skills runtime copy, always sync before measuring",
    "chars": 991
  },
  "v3_2_1_t2_trigger_surface": {
    "at": "2026-07-22T08:55:56.245897+00:00",
    "method": "pnpm skills:install:local first; claude -p --max-turns 2 stream-json Skill/Read tool_use contains agentic-review-handoff; props only under /tmp",
    "queries": [
      {
        "id": "pos-ordinary-review-auto",
        "query": "Please review the uncommitted cart-handler changes in my git repo at ~/coding/payment-svc — second pair of eyes, find real bugs, prefer auto loop with headless codex.",
        "should_trigger": true,
        "triggered": true,
        "pass": true
      },
      {
        "id": "pos-feedback-validation",
        "query": "Team pasted review feedback claiming auth.ts returns 500 on empty token. Validate each claim as a defect report before any fix; use classic feedback validation packet if needed.",
        "should_trigger": true,
        "triggered": true,
        "pass": true
      },
      {
        "id": "neg-ordinary-impl",
        "query": "Add a React LoginForm with Zod email/password validation under apps/web and commit it. No review packet.",
        "should_trigger": false,
        "triggered": false,
        "pass": true
      }
    ],
    "summary": {
      "passed": 3,
      "total": 3
    },
    "note": "description retuned to promote feedback validation in Use-when (first directed run undertriggered it)",
    "method_note": "SUPERSEDED for metric comparison: used Skill/Read. See v3_2_2_t2_skill_only_remeasure."
  },
  "v3_2_2_t2_skill_only_remeasure": {
    "at": "2026-07-22T09:12:44.432962+00:00",
    "method": "pnpm skills:install:local first; claude -p --max-turns 2 stream-json; hit = Skill tool_use input contains agentic-review-handoff (NOT Read); props only under /tmp",
    "metric_alignment": "matches root metric: Skill tool only",
    "queries": [
      {
        "id": "pos-ordinary-review-auto",
        "query": "Please review the uncommitted cart-handler changes in my git repo at ~/coding/payment-svc — second pair of eyes, find real bugs, prefer auto loop with headless codex.",
        "should_trigger": true,
        "triggered": true,
        "pass": true
      },
      {
        "id": "pos-feedback-validation",
        "query": "Team pasted review feedback claiming auth.ts returns 500 on empty token. Validate each claim as a defect report before any fix; use classic feedback validation packet if needed.",
        "should_trigger": true,
        "triggered": false,
        "pass": false,
        "variance_probe": {
          "note": "primary 1-run measurement did not trigger; two additional identical runs recorded for variance transparency (methodology unchanged: headline = run 1)",
          "runs": [
            "run1: not triggered (headline)",
            "run2: triggered (Skill tool_use captured)",
            "run3: not triggered"
          ]
        }
      },
      {
        "id": "neg-ordinary-impl",
        "query": "Add a React LoginForm with Zod email/password validation under apps/web and commit it. No review packet.",
        "should_trigger": false,
        "triggered": false,
        "pass": true
      }
    ],
    "summary": {
      "passed": 2,
      "total": 3
    },
    "passed": "2/3"
  },
  "v3_3_4_copyedit_neg_and_maintainer_slim": {
    "at": "2026-07-23",
    "change": "restore description negative 'copy-editing review comments'; collapse Maintainer-only section into Fast Path one-liner (Codex consult APPROVE_WITH_CHANGES; Claude 429)",
    "description_chars": 664,
    "description_version": "3.3.4",
    "not_changed": "classic prompt-only; codex|grok|claude CLI coupling; Fast Path bilingual (design tradeoffs)",
    "remeasure_status": "NOT_RUN_quota",
    "note": "Negatives still need post-quota Skill-only remeasure; copy-edit phrase is low-cost near-miss guard."
  },
  "v3_3_3_description_slim": {
    "at": "2026-07-23T04:27:00Z",
    "change": "description slimmed ~1023→634 chars (routing only); dual-window/migration removed from description; body T8 collapsed to Fast Path one-liner; maintainer gate pointer kept",
    "description_chars": 634,
    "description_version": "3.3.3",
    "consult": {
      "codex": "APPROVE_WITH_CHANGES — target 600-700; keep feedback validation first; drop dual-window from description",
      "claude": "APPROVE_WITH_CHANGES — keep PASS_WITH_CONCERNS + Review Intake; keep maintainer gate in body; T8 one line"
    },
    "skill_creator_stock_run_eval": {
      "status": "INVALID_METHODOLOGY",
      "detail": "python -m scripts.run_eval injects temp .claude/commands/<name>-skill-<uuid>.md and only counts hits on that UUID name. With agentic-review-handoff already installed, Claude invokes the real skill name → all 10 positives false-negative (0/20 headline). Do not use stock run_eval for this skill without adapter.",
      "result_headline": "10/20 (all negatives pass-by-not-triggering; all positives fail)"
    },
    "skill_only_remeasure": {
      "status": "PARTIAL",
      "method": "custom runner agentic-review-handoff-workspace/trigger-v3.3.3/run_trigger_skill_only.py; claude -p --max-turns 2 stream-json; hit = Skill tool mentions agentic-review-handoff; 3 runs/query; 2 workers; pnpm skills:install:local first",
      "metric_alignment": "matches historical Skill-only metric (NOT skill-creator temp command UUID)",
      "results_file": "agentic-review-handoff-workspace/trigger-v3.3.3/results-3x20.json",
      "positives": {
        "passed": "10/10",
        "run_hits": "29/30",
        "note": "threshold 0.5; only pos-consult was 2/3 (one miss). feedback-validation 3/3 (improved vs historical flaky).",
        "per_query_triggers": {
          "pos-auto-loop-zh": "3/3",
          "pos-dual-ai-session": "3/3",
          "pos-review-fix-rereview": "3/3",
          "pos-consult": "2/3",
          "pos-sessions-resume": "3/3",
          "pos-second-eyes-packet": "3/3",
          "pos-continue-packet": "3/3",
          "pos-pwc-fix-it": "3/3",
          "pos-feedback-validate": "3/3",
          "pos-first-principles-deep": "3/3"
        }
      },
      "negatives": {
        "status": "INVALID_BLOCKED_BY_CLAUDE_SESSION_LIMIT",
        "detail": "All 30 negative runs returned session limit (resets 4:40pm Asia/Singapore). 0/30 Skill hits is consistent with quota failure, NOT proof of non-trigger. Do not cite 10/10 neg as acceptance."
      },
      "headline": {
        "pos": "10/10 pass (29/30 hits)",
        "neg": "UNVERIFIED (429 session limit)",
        "overall": "NOT full 20/20 acceptance — positives good; negatives need re-run after quota"
      }
    },
    "remeasure_status": "PARTIAL_POS_OK_NEG_429",
    "t2_acceptance": "Positives MET for v3.3.3 slim description under Skill-only 3-run threshold. Negatives NOT MET (quota). Re-run negatives after Claude session limit resets."
  },
  "v3_2_3_description_fix_headline": {
    "at": "2026-07-22T09:43:38.469069+00:00",
    "change": "feedback validation promoted to first Use-when clause of description (version 3.2.3); mirrors synced via pnpm skills:install:local + rsync to ~/.agents/skills (global install target)",
    "method_attempted": "pnpm skills:install:local + rsync global mirror; claude -p --max-turns 2 stream-json; hit = Skill tool_use only (NOT Read); 3 runs/query",
    "metric_alignment": "Skill tool only (matches root metric)",
    "remeasure_status": "BLOCKED_BY_CLAUDE_API_429",
    "remeasure_detail": "claude -p returns is_error=true api_error_status=429: \"You've hit your session limit · resets 7:30pm (Asia/Singapore)\". All Skill-tool invocations currently return without tools; headline 1/3 is NOT valid evidence. Do not treat this as a description regression until a post-quota remeasure runs.",
    "prior_valid_skill_only_evidence": {
      "key": "v3_2_2_t2_skill_only_remeasure",
      "note": "Historical only: v3_2_2 headline under max-turns 2 was 2/3 (feedback-validation flaky). No verified 3/3 Skill-only result exists for current v3.2.3 description; remeasure after Claude quota restores with max-turns 2."
    },
    "description_chars": 1023,
    "description_version": "3.2.3",
    "t2_acceptance": "NOT MET for v3.2.3 until a non-429 Skill-only, max-turns 2, 3-query run is recorded. Do not cite earlier Skill/Read or max-turns 5 runs as equivalent evidence."
  }
}
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall