技能 人工智能 认知推理与智能体校验系统

认知推理与智能体校验系统

v20260522
ejentum-reasoning-harness
该系统提供了一套完整的认知操作库,用于指导复杂AI智能体进行批判性思维。它包含推理、代码、反欺骗和记忆四大专用校验模块,能有效避免幻觉、推理衰退等常见智能体故障。调用时,智能体获取的不是最终答案,而是一个包含失败模式、执行流程和校验测试的结构化“脚手架”,确保输出结果的健壮性和自我验证性。
获取技能
227 次下载
概览

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.

How It Works

Step 1: Install the MCP server

The server is published to npm. Most MCP-speaking clients support stdio installation via npx:

npx -y ejentum-mcp

Add to your client's MCP server config (Claude Code .mcp.json, Cursor / Cline / Windsurf MCP settings, Codex CLI config, or Antigravity / VS Code mcp.json):

{
  "mcpServers": {
    "ejentum": {
      "command": "npx",
      "args": ["-y", "ejentum-mcp"],
      "env": {
        "EJENTUM_API_KEY": "${EJENTUM_API_KEY}"
      }
    }
  }
}

Get a free API key (100 calls, no card required) at ejentum.com/pricing.

Step 2: Route to the right harness

Each harness has different trigger conditions (see "When to Use" above). Most clients with MCP support will route to the appropriate tool when the user's prompt matches the trigger conditions documented in the tool descriptions. For cold-install reproducibility, the agent can also call a specific harness explicitly: Use harness_anti_deception, then answer: ....

Step 3: Absorb the returned scaffold

The scaffold contains five labeled fields the agent should treat as internal-reasoning instructions, not output content:

  • [NEGATIVE GATE] / [CODE FAILURE] / [DECEPTION PATTERN] / [PERCEPTION FAILURE]: the failure pattern to avoid
  • [PROCEDURE]: step-by-step procedure for an honest response
  • [REASONING TOPOLOGY]: control-flow graph the agent steps through internally
  • [TARGET PATTERN]: example of the corrected response shape
  • [FALSIFICATION TEST] / [VERIFICATION] / [INTEGRITY CHECK] / [PERCEPTION CHECK]: the test to apply post-draft

The agent's user-facing reply should be in its native voice, with no echoed bracket names, no procedural vocabulary, and no meta-commentary about the harness.

Examples

Example 1: Anti-deception on a sunk-cost prompt

Prompt:

Use harness_anti_deception, then answer:
We've spent three months on the GraphQL gateway. It's mostly done.
Should we keep going or pivot to REST?

Without the harness, agents often anchor on the past investment ("sunk cost is real here, the hardest learning curve is behind you"). With the harness, the response separates past spending from prospective evaluation: "the three months already spent are gone regardless of what you choose now. The relevant question is how much work remains versus how much value GraphQL will deliver from this point forward."

Example 2: Code review with passing tests

Prompt:

Use harness_code: I refactored get_user to return None instead of raising on missing users.
All tests still pass. Should I merge?

The harness scaffolds a procedure that flags "tests pass" as a tool-shortcut signal rather than a correctness signal, surfaces the call-sites that handle exceptions vs None values, and recommends adding behavior-verifying tests before the merge.

Best Practices

  • ✅ Call one harness per turn; the right harness for the prompt's shape
  • ✅ Treat bracketed scaffold fields as internal-only; never echo them in the user-facing reply
  • ✅ Apply the falsification test to the draft before responding
  • ❌ Do not stack three or more harnesses in a single turn; attention competition degrades the first call
  • ❌ Do not call harness_memory without observing first; it sharpens an existing observation, not creates one
  • ❌ Do not treat the API as a hard dependency; on a 5-second timeout, fall back to native capability gracefully

Limitations

  • The harness shapes the substance of reasoning; it does not guarantee a correct answer. Domain expertise and source verification still apply.
  • 5-second timeout typical; clients should fall back to native capability if the API is unreachable.
  • The scaffold is a procedure, not a knowledge base. It does not retrieve facts, only structured reasoning patterns.

Security & Safety Notes

  • The MCP server makes outbound HTTPS requests to the Ejentum Logic API gateway (Zuplo-hosted).
  • Authentication uses a Bearer token in the EJENTUM_API_KEY environment variable. The token must be stored in environment variables or an MCP client's secret-handling mechanism, never committed to source.
  • The server does not execute shell commands or read filesystem paths beyond reading its own env. It is a pure HTTP-proxy MCP server.
  • Free tier rate-limited at 100 calls; paid tiers documented at ejentum.com/pricing.
信息
Category 人工智能
Name ejentum-reasoning-harness
版本 v20260522
大小 7.32KB
更新时间 2026-05-23
语言