菜单
Use when intended architecture must be compared with implementation: dependency policies, ADRs, ownership rules, drift, erosion, recovered models, and conformance classifications.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when locating, summarizing, auditing, or debugging Codex session rollout JSONL logs by CODEX_THREAD_ID, cwd, query, issue key, project path, child or inherited-history symptoms, malformed or huge log symptoms, permission concerns, or "what happened in this Codex thread" questions.
Use when auditing agent skills or plugins for safety, coverage, duplicated logic, token/context cost, prompt-contract risk, validation readiness, marketplace visibility, dependencies, network behavior, or install/update risk. For structural split/merge/move/delete/router decisions across skills or plugins, use capability-portfolio-architect instead.
Use when auditing agent skill/plugin portfolio architecture for cross-plugin capability overlap, weak routing, duplicated or missing skills, split/merge/delete/move decisions, shared capability extraction, reference/script extraction, or plugin boundary redesign. For safety, coverage, or validation review of a single artifact without boundary changes, use capability-auditor instead.
Use when performing broad external-first discovery, vetting, scoring, distillation, or synthesis of agent skills or plugin packs from the public web, GitHub/public repositories, OpenClaw/ClawHub, marketplaces, research sources, community implementations, user references, or local skills. Use for well-vetted capability synthesis, cross-skill comparison, plugin-pack synthesis, global capability acquisition, skill strengthening, and adoption/rejection changelogs.
Route and orchestrate agent skill and plugin lifecycle work. Use when work involves agent skills, plugins, marketplace packages, capability acquisition, synthesis, creation, installation, packaging, audit, or context-density optimization.
Use when creating, updating, validating, optionally installing/cache-refreshing, or handing off marketplace-backed agent plugins. Use for plugin manifests, local marketplace entries, plugin-pack synthesis, triggerable skill bundles, install visibility gates, and Codex app deeplinks.
| name | architecture-conformance |
| description | Use when intended architecture must be compared with implementation: dependency policies, ADRs, ownership rules, drift, erosion, recovered models, and conformance classifications. |
Bundled commands use $PLUGIN_ROOT ($env:PLUGIN_ROOT in PowerShell; same path suffix) for the plugin root. Set it once: use the host's plugin-root variable when defined (Claude Code: PLUGIN_ROOT="$CLAUDE_PLUGIN_ROOT"), otherwise the absolute path of this plugin's root directory.
Use when the question is whether implementation still matches intended architecture.
Triggers: compare code to ADR/docs/policy, detect drift or erosion, turn intent into checks, classify edges as convergence/divergence/absence/unknown, recover observed architecture before updating docs.
If intent is thin, label it incomplete. Do not convert recovered source shape into design intent.
python3 "$PLUGIN_ROOT/scripts/architecture_probe.py" <repo-path> --json --policy <policy.json>
architecture_intelligence.policy.v1 supports forbidden edges, required edges, and required documents. See references/contracts.md.
The probe is conservative: top-level static imports and exact document paths only. Reflection, generated code, runtime calls, full CODEOWNERS semantics, actual team communication, and deployment edges need separate evidence.
convergence: intended and observed relation agree.divergence: observed source violates intent.absence: required relation/document is missing.unknown: evidence or policy is insufficient.For each mapping record source, target, intended relation, observed relation, classification, and evidence path/command/doc/trace.
Compact: conformance summary, key convergences/divergences, missing intent evidence, priority fixes, validation checks.
Durable: architecture_intelligence.conformance.v1.
Do not report a violation without an intended rule and observed evidence. Do not enforce a new rule without intent, owner, exception path, and rollout mode.