Ask your agent why.
Not just what a work item cost, but why — why it cost that much across every session it took, why a decision was made, whether a skill was even worth reaching for.
Point your agent at its own history
There's no separate traceknot screen for this. A whole work item — every session ever claimed against the same ticket — or just one session exports as a full trace, and you hand it to the same agent you're already running in your terminal. Ask why something cost what it did, why a call was made, or how well a skill performed, and it answers by citing the exact node.
# PROJ-482 — 3 sessions
> why did PROJ-482 end up costing more than we scoped, and could any of it have been avoided?
⏺ traceknot export work-item --provider jira --key PROJ-482
⎿ overview.md, 3 sessions (session_a1b2f9e2, session_c3d461af, session_e3147585)
Most of the overrun is one thing: the auth refactor got rebuilt from scratch in the second session after the first attempt's approach was abandoned — nothing in the ticket said why. The third session's flaky-test retries account for the rest; flags.md already caught that one as a repeated-call outlier.
see: overview.md, 02-session_c3d461af/nodes/004-decision.md, 03-session_e3147585/flags.md
> was that abandoned first attempt agreed with me, or the agent's own call?
> just for the third session — was the flaky-test subagent worth spinning up?
Illustrative — this runs in your own Claude Code, Codex CLI or Copilot CLI terminal, not a traceknot screen
Every session on a work item — or drill into just one
“Which of the three sessions on this ticket actually moved the implementation forward?”
“Why did we end up building the retry queue with backoff instead of what the ticket originally asked for?”
“Was the reasoning-token spend on that chat turn normal, or is this session an outlier?”
“Did the subagent actually need to run, or could a single tool call have done it?”
Three kinds of question
The analysis skill reads the export differently depending on what you're actually asking — and you can point it at a single session or at a whole work item, every session ever claimed against that ticket, so the answer isn't limited to today's run.
Cost & duration
“Why did this ticket cost more than we scoped? Did any of its sessions stall, loop, or redo work?” — flags statistical outliers first, then the full timeline.
Decision rationale
“Why was X implemented when it wasn't anywhere in the spec — was that agreed with me, or the agent's own call?” — walks the tool calls and turns that led there, even if that happened in an earlier session on the same ticket.
Skill & subagent effectiveness
“Was a skill used well? Should it have had more context upfront? Why didn't it use the obvious tool?”
How it actually runs
- 1
You ask, in your normal terminal
No separate traceknot screen for this — you're talking to the same Claude Code, Codex, or Copilot session you already run, with the skill enabled.
- 2
It exports the trace
Either the one session, or every session ever claimed against the same work item — whichever the question actually needs.
- 3
It reads top-down, not everything at once
Summary and flagged outliers first, then the timeline, then only the specific nodes that matter — the same way you'd triage it yourself.
- 4
It answers with citations
Every conclusion names the specific node or file path it rests on, so you can go check it yourself instead of taking the agent's word for it.
The skill is opt-in and off by default —enable it for an agent.