Contradiction Detector

面向 OpenClaw 多代理环境：枚举各代理的指令文件、cron 任务、技能与 hook，按 12 种已知矛盾模式交叉比对，输出含原文证据、影响与修复建议的结构化报告，经用户确认后以 diff 方式修复并复验。

korengast

热度 54.07

最近更新 2026-05-11 17:41:28

分类：效率依赖：读写文件 / 代码执行应用场景：工作流 / 报告撰写

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add contradiction-detector

Metadata

name: contradiction-detector
description: Detects and eliminates contradictions between agent instruction files that cause hallucinations and silent misbehavior. Use when: (1) any agent behaves inconsistently despite correct-looking instructions, (2) a sub-agent format/behavior was fixed in one file but the behavior didn't change, (3) a new agent, cron job, or instruction file was added, (4) running a periodic consistency sweep. Scans AGENTS.md, SOUL.md, HEARTBEAT.md, MEMORY.md, cron job payloads, and skill files for any agent — then cross-references them for contradictions.
metadata: {"openclaw":{"emoji":"🔍","requires":{"bins":["openclaw","python3","find","cat"]},"configPaths":["openclaw.json"],"readAccess":["workspace","workspace/skills","workspace/hooks"]}}

Contradiction Detector

Contradictions between agent instruction files are the #1 cause of silent misbehavior in multi-agent OpenClaw setups. Files are loaded non-deterministically across sessions — the only fix is making all files agree.

Prerequisites

openclaw CLI — listing agents and cron jobs
python3 — parsing JSON config
find, cat — standard POSIX utilities
Filesystem access to the OpenClaw workspace directory

Audit Procedure

Step 1 — Discovery

Enumerate every instruction source: all agents, their workspace files (AGENTS.md, SOUL.md, HEARTBEAT.md, MEMORY.md, IDENTITY.md, USER.md), all cron jobs with inline prompts, all installed skills, and any hook scripts.

→ Detailed procedure: references/discovery.md

Build a complete inventory per agent before proceeding. Missing even one source means contradictions slip through.

Step 2 — Cross-reference for contradictions

Systematically check every pair of instruction sources per agent against 12 known contradiction patterns:

#	Pattern	Risk	Most common trigger
1	Format duplication	HIGH	Same output format defined in 2+ files independently
2	Cron prompt override	HIGH	Cron contains stale instructions; HEARTBEAT.md fix is invisible
3	Routing mismatch	HIGH	Different session/channel targets for same output
4	Persona/tone conflict	MED	SOUL.md vs AGENTS.md communication style
5	Trigger/schedule mismatch	MED	Cron frequency vs HEARTBEAT.md timing
6	Memory routing conflict	MED	Same data type sent to different memory locations
7	Dead/orphaned rules	MED	Old rule not removed after newer rule added elsewhere
8	Cross-agent assumption mismatch	MED	Sender expects X, receiver does Y
9	Implicit vs explicit behavior	LOW-MED	Gap filled differently by different files
10	Skill-agent conflict	LOW-MED	Skill instructions override agent rules when triggered
11	Stale MEMORY.md directives	LOW	"From now on" in MEMORY.md contradicts current AGENTS.md
12	Environment/config drift	LOW	Hardcoded paths, model names, or URLs that changed

→ Full detection procedures and examples: references/contradiction-patterns.md

Step 3 — Report and fix

Produce a structured finding for each contradiction (agent, severity, pattern, evidence with exact quotes, impact, proposed fix). Present all fixes as diffs and wait for user confirmation before applying.

→ Report format, severity guidelines, and fix protocol: references/reporting-and-fixes.md

Post-fix verification

After every fix: re-read both files, verify agreement, and confirm no new contradictions were introduced.

Scheduling (optional)

This skill works as a manual on-demand invocation. Optionally add to HEARTBEAT.md or a cron job for periodic runs.