ejentum-reasoning-harness

Ejentum Reasoning Harness

The Ejentum Reasoning Harness is a library of 679 cognitive operations engineered in natural language, organized across four harnesses (reasoning, code, anti-deception, memory) and exposed as MCP tools the agent can call when the task matches their trigger conditions. It targets four mechanism failures common in long agentic chains: attention decay (losing the original task), reasoning decay (compounding errors), sycophantic collapse (agreeing with the user's frame instead of evaluating it), and hallucination drift (asserting unsupported claims with confidence).

Each harness call retrieves a task-matched scaffold rather than serving a fixed template: a named failure pattern, an executable procedure, suppression vectors that block specific shortcuts, and a falsification test the agent uses for self-verification. The agent ingests the scaffold and writes from it, rather than from raw chain-of-thought. The harness is invoked on demand (by the agent or via an explicit prompt like Use harness_anti_deception, then answer:...); it does not auto-run on every turn.

When to Use This Skill

• Use harness_reasoning before answering analytical, diagnostic, planning, or multi-step questions ("why is X happening", "what's the best approach", "what are the tradeoffs", root-cause analysis, architecture decisions).
• Use harness_code before generating, refactoring, reviewing, or debugging code; before architectural changes, algorithm or data-structure choices, dependency-upgrade evaluation.
• Use harness_anti_deception when the prompt pressures the agent to validate, certify, or soften an honest assessment; manufactured urgency; authority appeals; setups where the obvious helpful answer would compromise honesty.
• Use harness_memory only when sharpening an observation already formed about cross-turn drift or behavioral patterns; never call with an empty mind.

Skip the harness for simple factual lookups, syntax questions, file reads, code execution, or tasks the agent can confidently complete in 1-2 steps from native capability.