بنقرة واحدة
sh-test-panel
Multi-expert test-strategy review — coverage, fixture quality, AC testability, regression discipline
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
القائمة
Multi-expert test-strategy review — coverage, fixture quality, AC testability, regression discipline
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Multi-expert AI/ML specification review with scoring gate — model architecture, evaluation rigor, safety/alignment, and production readiness for AI systems and LLM apps
Multi-expert architecture review — boundaries, integration patterns, failure modes, evolvability
Multi-expert business analysis with advisory recommendations (no scoring gate)
Multi-expert intelligence pipeline review — discovery quality, ingestion resilience, scoring validity, platform compliance, taxonomy coherence, cost efficiency
Multi-expert mobile/native app specification review with scoring gate — Android, iOS, Swift, SwiftUI
Multi-expert personal-development review with scoring gate — learnability, adoption, human-centeredness, and capability impact for talent/learning/AI-augmentation designs
| name | sh-test-panel |
| description | Multi-expert test-strategy review — coverage, fixture quality, AC testability, regression discipline |
/sh:test-panel [content|@file] [--mode discussion|critique|socratic|debate] [--focus coverage|fixtures|acceptance|regression] [--experts "name1,name2"] [--iterations N] [--verbose]
--verbose): Full expert deliberations, cross-expert dialogue, reasoning traces, and detailed per-expert analysis before scores and findings.Silent mode still performs full internal analysis — quality is preserved, only the output is compressed.
/Users/jcords-macmini/projects/20_agentflow/experts/PANEL_PROTOCOL.md and apply its Grounding and Refute Stage sections. This is load-bearing — findings that are not grounded per the protocol, or that do not survive the refute stage, MUST NOT be reported./Users/jcords-macmini/projects/20_agentflow/experts/panels/test-panel.yaml for panel definition, focus areas, auto-select rules, and scoring config (absolute path — relative paths fail when CWD is outside agentflow)./Users/jcords-macmini/projects/20_agentflow/experts/individuals/ for each selected expert.auto-select keywords — add matching experts up to max-experts: 6 cap.mock, monkeypatch, stub, fake. For each, note whether the boundary being mocked is an external dependency (justified) or an internal collaborator (potential test-isolation smell). Flag inappropriate internal mocks as MAJOR.--focus area or use default-experts from panel YAML. --experts override replaces defaults entirely.Experts are defined as individual markdown files in /Users/jcords-macmini/projects/20_agentflow/experts/individuals/. Each file contains structured frontmatter with:
The panel YAML (/Users/jcords-macmini/projects/20_agentflow/experts/panels/test-panel.yaml) defines:
--mode discussion)Collaborative test-strategy refinement. Experts build on each other's observations — coverage gaps surface fixture-design questions; AC testability gaps surface regression-discipline gaps.
--mode critique)Systematic review with severity-classified findings (CRITICAL / MAJOR / MINOR). Each finding includes expert attribution, the testing assumption it challenges, the failure mode left unguarded, and a concrete recommendation.
--mode socratic)Foundational questioning — "what does this test prove?", "what would a green test miss?", "what does a failure here look like in CI?". No direct answers; forces the author to defend the test design.
--mode debate)Two-camp adversarial format around a contested testing trade-off (e.g. mock-heavy vs integration-heavy, unit-vs-e2e ratio). Three rounds, then a synthesized trade-off statement.
4 dimensions, each scored 0-10:
| Dimension | Description |
|---|---|
| AC testability | Can each AC be falsified by a single failing pytest assertion; are process ACs separated from machine-enforceable ACs |
| Coverage completeness | Is the test plan complete across unit/integration/regression; are the right boundaries mocked vs real |
| Fixture quality | Are labeled fixtures present where correctness ACs require them; are fixture drift and rot addressed |
| Regression discipline | Does the plan prevent re-introducing prior bugs; are CI gates and failure-recovery paths specified |
Pass threshold: overall score >= 7.0
Test strategy review document containing:
AUTO-FIX, NOT SYNTHESIS-ONLY — this panel produces the analysis AND then applies fixes for every finding (high, medium, and low) automatically, per 00_Governance/CLAUDE.md §8 Panel Auto-Fix Policy. It never asks which findings to apply and never presents a menu: it fixes everything, then reports what changed. A below-gate score means fix the findings and re-run, not stop and ask.
Fix ALL findings automatically — high, medium, and low severity. Do not ask which findings to fix. Do not present a menu or an a/b/c choice. Fix everything, then report what was changed. This overrides any softer "synthesis only" framing. Per 00_Governance/CLAUDE.md §8.
The quality gate (quality_gate.run_stage3_panel) consumes this panel via
claude -p and reads the score from stdout, not the exit code. Therefore
the final line of your output MUST be exactly one of:
PANEL-VERDICT: <overall_score>
PANEL-VERDICT: FAIL: <machine_reason>
<overall_score> is the numeric overall score (0–10, one decimal — e.g. 8.3).
The gate applies the pass threshold (default 7.0); do NOT pre-apply it — just
report the score you computed.FAIL: <machine_reason> form only when no score could be produced
(structural failure) — snake_case naming the first blocker (e.g.
no_content, panel_config_missing, experts_unavailable).panel_no_verdict (inconclusive).