traces-and-audit

SKILL.md

Skill: TraceMem Traces and Audit

Purpose

This skill explains the concept of the Decision Trace as an artifact. Understanding this helps you write better "evidence" into the system.

When to Use

  • When you need to understand what TraceMem is actually recording.
  • When generating reports or answering questions about past actions ("Why did you do that?").

When NOT to Use

  • You generally do not "use" this skill to execute actions, but to inform how you execute them.

Core Rules

  • The Trace is the Truth: If it's not in the trace, it didn't happen (legally/audit-wise).
  • Append-Only: You cannot go back and fix history.
  • Complete Picture: A trace includes your ID, the time, the policy version, the data schema version, and the exact outcomes.

Correct Usage Pattern

  1. Design for Readability: When running a decision, imagine a human reading the trace 6 months later.

    • "Why did this agent delete this user?"
    • Look at the intent, look at the context you added, look at the policy result.
    • If the trace answers the question, you succeeded.
  2. Linking: If you chain decisions (one decision triggers another workflow), reference the parent decision_id in the child's metadata or context.

Searching Past Decisions

Use decision_search to query your agent's previous decisions:

  • Find precedent: Search by text, category, or tags before making a new decision
  • Check supersession chains: Results include supersedes and superseded_by indicators -- follow the chain to find the current active decision
  • Filter by status: Use status: "committed" to find only finalized decisions
Tool: decision_search
Parameters:
  - query: "authentication"  (free-text search)
  - category: "architecture"  (optional)
  - tags: ["jwt", "auth"]  (optional, all must match)
  - status: "committed"  (optional)
  - limit: 10  (optional, default 20, max 100)

This is particularly valuable for:

  • Answering "Why did we decide X?" questions
  • Avoiding duplicate or contradictory decisions
  • Building on prior context when making related decisions

Common Mistakes

  • Phantom Actions: Doing side effects (like calling an external API) without recording it in TraceMem or via a Data Product. This creates "dark matter" — actions that have no record.
  • Incomplete Evidence: Reading data via a side-channel (not a Data Product) and then acting on it. The trace will show the action but not the data that justified it.
  • Not searching before deciding: Always check decision_search for existing decisions on the same topic before recording a new one.

Safety Notes

  • Exoneration: A good trace protects you (the agent). If a policy was wrong, the trace proves you followed the policy correctly. If data was bad, the trace proves you acted on the bad data you were given.
Weekly Installs
15
GitHub Stars
1
First Seen
Jan 23, 2026
Installed on
gemini-cli13
opencode13
cursor11
codex10
github-copilot9
claude-code7