$trace

$trace rebuilds what a session actually did. It calls two MCP tools, trace_timeline and trace_summary, then combines their output into one chronological event log plus the aggregate statistics. From that you can see which keyword triggered a skill, how delegation flowed between agents, and which tool became the bottleneck.

When to use it

• When you want to understand what happened during a complex multi-agent session
• When debugging an unexpected skill activation or mode transition
• When finding slow tools or agents to eliminate bottlenecks
• When auditing what order hooks fired in

Trigger keywords: "trace", "show trace", "what happened", "flow trace", "session trace"

How to invoke

codex
> trace

codex
> $trace

How it works

Calls trace_timeline with no arguments to retrieve the chronological event log for the latest session. The timeline shows hook fires, keyword detections, skill activations, agent delegations, mode transitions, and tool calls in the order they occurred.

Then calls trace_summary to produce aggregate statistics: how many times each hook fired, which keywords were detected, which skills activated, what mode transitions occurred, and which tools were slowest.

The timeline comes first, the summary second. Within it, mode transitions are called out, and any tool or agent that ate the most time gets flagged as a bottleneck. It also traces the causal chain of keyword → skill → agent so you can follow what set off what.

Outputs

• Chronological event timeline of the current session
• Aggregate stats: hook fire counts, keyword detections, skill activations, mode transitions
• Bottleneck report identifying the slowest tools and agents
• Keyword → skill → agent causal flow summary

Related skills

• $deep-interview — interview workflow whose execution flow can be reviewed with this skill
• $ralplan — consensus planning whose agent sequence appears in the trace
• $wiki — knowledge base for saving insights from a trace for future sessions