Ai Trace Runner

Transform ambiguous AI tasks into auditable execution traces with verified evidence and AI-smell detection.

Install
cmdop skills install agensi-ai-trace-runner

AI Trace Runner is a skill that wraps AI agent execution in a structured pipeline designed to eliminate unverifiable task completion claims. Instead of relying on bare prompting, it creates a request contract at the start of each run, generates an explicit execution plan, logs every material action taken, and assembles a final evidence bundle where every claim is tied to a specific file path, command output, or observation.

A core problem it addresses is hallucinated confidence — agents reporting success when a task actually failed or was silently bypassed. AI Trace Runner counters this by forcing explicit verification at each step rather than accepting an agent’s self-reported outcome at face value.

The skill includes built-in AI-smell detection, which strips robotic filler language from outputs so that results remain precise and audit-ready. It also integrates with project.yaml for dependency and execution topology management, making it suitable for multi-step workflows with defined ordering constraints.

Human-in-the-loop gates are built in for destructive actions, ensuring an agent cannot proceed through high-risk operations without explicit approval. The skill works with standard CLI and shell environments and with any file-system or API-based tool within the agent’s scope.

This skill is aimed at developers who need to debug complex agent behaviors, produce evidence of code changes, or demonstrate compliance through a verifiable audit trail.

Use cases

  • Generate an auditable run record for an agent performing a multi-step code refactor
  • Detect and remove hallucinated success claims from agent task outputs
  • Enforce human approval gates before an agent executes destructive shell commands
  • Manage agent task dependencies using project.yaml execution topologies
  • Assemble an evidence bundle linking every agent claim to a specific file or command output
  • Debug agentic workflows by reviewing the logged sequence of material actions

When to use it

  • When an agent's task completion claims need to be independently verifiable
  • When working in environments that require an audit trail of AI-driven changes
  • When a workflow involves destructive actions that need human approval before execution
  • When project.yaml is already used for dependency and topology management
  • When diagnosing why an agent silently failed or bypassed a required step

When not to use it

  • When no tool list is provided, the skill has no exposed MCP tools to call
  • When the task is simple and single-step with no auditability requirement
  • When the agent runtime does not support skill-based execution wrappers
  • When there is no need for evidence bundles or structured execution plans