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:

#PatternRiskMost common trigger1Format duplicationHIGHSame output format defined in 2+ files independently2Cron prompt overrideHIGHCron contains stale instructions; HEARTBEAT.md fix is invisible3Routing mismatchHIGHDifferent session/channel targets for same output4Persona/tone conflictMEDSOUL.md vs AGENTS.md communication style5Trigger/schedule mismatchMEDCron frequency vs HEARTBEAT.md timing6Memory routing conflictMEDSame data type sent to different memory locations7Dead/orphaned rulesMEDOld rule not removed after newer rule added elsewhere8Cross-agent assumption mismatchMEDSender expects X, receiver does Y9Implicit vs explicit behaviorLOW-MEDGap filled differently by different files10Skill-agent conflictLOW-MEDSkill instructions override agent rules when triggered11Stale MEMORY.md directivesLOW"From now on" in MEMORY.md contradicts current AGENTS.md12Environment/config driftLOWHardcoded 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.