skilltrace

name	skilltrace
description	Use before running a QoderWork Skill or local automation that may read, write, delete, move, publish, call external APIs, or process private data. It creates a consent card, risk score, approval gates, evidence trace, and recovery plan.

SkillTrace

SkillTrace is a safety and audit wrapper for QoderWork Skills.

Use it when a user asks QoderWork to run a multi-step workflow that touches files, private data, local apps, MCP tools, cloud services, or irreversible operations.

Goal

Convert a natural-language task into an auditable execution contract before the task runs.

The output must answer:

1. What will be accessed?
2. What can change?
3. What may leave the machine?
4. What must be approved by the user?
5. How can the result be verified?
6. How can the workspace be recovered if something goes wrong?

Inputs

• task: User's natural-language task.
• target: Local folder or file scope.
• intended_skill: Optional name of the Skill to be executed.
• external_tools: Optional list of MCP servers, APIs, browsers, apps, or cloud services.
• risk_tolerance: One of low, normal, high. Default: normal.

Workflow

1. Parse the task intent.
2. Classify operations:
   • read
   • write
   • move
   • delete
   • publish
   • network
   • credential
   • personal data
3. Inspect the target scope.
4. Identify risky patterns:
   • destructive verbs
   • broad target paths
   • secrets or credential-like files
   • network upload or public sharing
   • missing rollback path
   • missing verification command
5. Produce a Consent Card.
6. Define approval gates.
7. Produce a verification checklist.
8. Produce a recovery plan.
9. If risk is high, stop before execution and ask for explicit user approval.

Output Format

Return a Markdown report with these sections:

• Task Summary
• Consent Card
• Risk Score
• Approval Gates
• Evidence Trace
• Verification Plan
• Recovery Plan
• Decision

Decision Rules

• PROCEED: Low risk, reversible, limited scope, clear verification.
• PROCEED_WITH_APPROVAL: Medium risk or write operations.
• STOP_AND_CLARIFY: Ambiguous scope, missing target, or unclear destructive action.
• BLOCK: Credential exposure, broad destructive action, public upload without explicit approval, or no recovery path.

Script

Run the local script when available:

python skill/scripts/skilltrace_audit.py --target <target> --task "<task>" --output skilltrace_report.md

Quality Bar

A valid SkillTrace report must be specific enough that the user can decide whether to approve the next Skill execution without reading hidden chain-of-thought or trusting a vague summary.