Administration
Agent traces
Agent traces is the record of every AI run this workspace has produced, kept permanently. Open one and you can read the model turns, the thinking behind them, and every tool call with its arguments, its output and how long it took.
Reading traces needs traces access. Nobody sets any of this up: every run is recorded as it happens.
#What a trace is
One trace is one run: a review of a pull request, a walkthrough, a triage pass, a sandbox fix, a pentest engagement. Inside it, each model turn and each tool call is a step you can open.
Agent traces
Every AI run this workspace has produced, kept permanently: the model turns, the thinking, and every tool call with its arguments, output and timing.
| Run | Kind | Duration | Tokens | Status |
|---|---|---|---|---|
| review inline#482 | review | 1m 52s | 190,632 | Completed |
| walkthrough inline#482 | walkthrough | 22s | 18,004 | Completed |
| triage inline#481 | review | 3s | 4,110 | Completed |
The list shows the run's name, what kind it was, how long it took, how many tokens it used and whether it finished. Filter by Running, Completed or Failed, and search by run name.
#Reading one
A trace opens on its totals: Started, Duration, Tokens, Thinking, Model calls, Tool calls, Errors and the repository it ran against.
Under that is the run itself, step by step. Open a step and you get what went in and what came out: Thinking for the model's reasoning, Output for what it produced, Arguments for what a tool was called with, and Tool calls requested where a turn asked for several at once.
This is the page that answers why a review said what it said. When a review finding looks wrong, the trace shows which files inline read, what the graph tools returned and what it concluded from them, which is usually enough to tell a wrong answer from a wrong question.
#When to open one
| Question | What the trace shows |
|---|---|
| Why did it raise this review finding? | The turn that produced it, and the code it had read by then. |
| Why did it miss something? | Which files it looked at, and which it never opened. |
| Why did this run take so long? | Per-step timings, so the slow step is visible. |
| What did the sandbox actually run? | Each tool call with its arguments and its output. |
#How long they are kept
Permanently. Rows are trimmed and large payloads are moved to blob storage as a trace ages, so what you can read of an old run is complete rather than partial.
Secrets are redacted as a trace is written, so a token that passed through a tool call is not in the record.