Track record

A space's track record shows how each agent's tasks in a space went, so you can see which agent and model does the work well on that space's code, and at what cost. It reads what already happened on the space's tasks; nothing extra runs.

Open a space and choose its Track record tab. Members, viewers and admins see it; guests don't.

The table

Each row is a harness and a model, such as OpenCode with opencode-go/kimi-k3. The model is the one the harness reported running, or the one its runner was set to use when the harness doesn't say. A runner on a workspace model shows the endpoint's model ID. A runner left on its harness's own model shows as its default model.

A task counts toward the agent and model that ran its latest build. A task that was never built counts toward the one that wrote its latest plan.

  • Tasks: how many tasks the row covers.
  • Merged: tasks whose pull request was merged, out of those whose pull request was merged or closed. Pull requests still open aren't counted until they're decided.
  • Built from first plan: tasks built from the agent's first plan, out of the tasks built from a plan. Plans count by version, as in the plan's history: a turn that only asks or answers a question doesn't make a new plan, and neither does your own edit.
  • Questions: questions the agent asked per task.
  • Taken over: tasks where an engineer took over the branch.
  • Review comments: review comments per pull request.
  • Cost per merge: what the row's agent and model spent on the row's tasks, divided by its merged pull requests. Runs another agent or model made on one of these tasks, such as an earlier build with a different runner, don't count here or in any other row. "est." means some of it is the harness's flat estimate, not a reported cost. Runs whose harness reports no cost, such as Codex, Kimi Code and Qwen Code, add nothing, and a row where no run reported a cost shows —. A run that failed before reaching the model has no cost; one that failed later counts what its model calls had used. A run on a workspace model has no cost either: harnesses price tokens from their own provider's list, which doesn't cover the endpoint's model.

A row with fewer than five tasks says "Too few tasks to say" instead of showing rates.

A space's track record: Claude Code and OpenCode side by side, with their merges, plans, questions, takeovers, review comments and cost per merge

People choose which agent takes each task, and they may give the hardest tasks to the agent they trust most. Keep that in mind when one row looks worse than another. For a fair comparison on one task, use Compare agents.

One agent's tasks

Choose a row to list its tasks, newest first, each with how it went: the pull request's outcome, how many plans came before the build, the agent's questions, a takeover, failed runs, review comments and what the row's agent and model spent on it.

Previous
Schedules