dokugent trace¶
The dokugent trace command is used to analyze and explain agent behavior. It generates a trace from the agent’s certificate and plan, outputting step-by-step reasoning, token estimates, constraint checks, and logic flow.
Unlike dokugent simulate, which predicts output and next actions, trace focuses on introspective explainability for auditing and debugging purposes.
Usage¶
Options¶
Example¶
Output Format¶
The trace command returns the full contents of the certified .doku.json file. If the file contains agent metadata, task plans, constraints, and conventions, all will be included. The example below shows a minimal file with only agent metadata.
In future versions, this will expand to include full plan analysis, step-by-step logs, and context deltas.
What It Does¶
dokugent trace provides introspective explainability for certified agent plans. Unlike simulate, it does not run the plan in real-time, but instead reveals:
Trace = inspect behavior | Simulate = run behavior
- Step-by-step intent breakdown
- Constraint adherence or violations
- Token estimates per action
- LLM reasoning traces
How It Works¶
Doku URI Resolution¶
When passed a doku://<agent>@<born-timestamp> URI, Dokugent resolves the slug to a secure MCP URL. This lets trace work with remote agents or certified third-party plans.
Remote Fetch & Auth¶
If the URI points to a Supabase-protected resource, dokugent trace uses your API key (or JWT) to authorize the request before downloading the .doku.json.
Note: You must provide a valid
SUPABASE_SERVICE_ROLE_KEYin your.envfile or pass--tokenmanually. Without this, remote trace will be denied.
Trace Logic Internals¶
Once loaded, the trace runner:
- Parses the
plan,criteria, and any constraints - Applies mutations from
--change-constraintsor--violate - Executes LLM-based dry introspection
- Produces a memory-safe audit trail
Remote Powered by MCP¶
All trace operations are routed through a secure MCP (Model Context Protocol) server. When you run dokugent trace, the CLI sends:
- The resolved
dokuUri(points to a compiled agent certificate) - A signed
intenttoken (JWT for authentication)
The MCP endpoint validates, fetches, and parses the agent data before returning a structured response to Dokugent CLI.
Description¶
This command analyzes the reasoning process of an agent by:
- Resolving
doku://<agent>@<born-timestamp>URIs into remote certificate locations - Applying optional constraint mutations or overrides
- Generating a trace file based on what the agent would do
- Logging decision-making steps and projected token cost
Use this when you need to:
- Explaining agent behavior before execution
- Debugging missing context or misaligned plans
- Conducting Human-in-the-Loop reviews