Costs

What the work spent. Whether this is visible is a deployment setting — the cost view is hidden by default, so if you cannot see it your operator has not turned it on.

What is counted

Agent spend, costed against the rate card of the provider that actually served the call. A project running Claude through one route and another model through another is costed per provider rather than at one blended rate that would be wrong for both.

A run that picks up where an earlier run of the same agent left off is counted at what it spent itself. Costs recorded before this correction, which included the earlier runs' spend again, are corrected when Repave is upgraded, wherever the run's own spend can be told apart from the earlier runs'. A run that cannot be told apart, such as one on a model with no published price, keeps the cost it was recorded with.

A run's token counts cover everything its cost does, including any subagents the agent handed work to, and a run that picks up where an earlier one left off counts only its own tokens. Runs recorded before this release count the main agent's tokens only, so a run that used subagents shows far fewer tokens than it used, though its cost is right.

A run that fails or is cancelled is counted too. When the agent reported its cost before it stopped, that cost is recorded; when it did not, because it was stopped or its process was killed part-way through, its work up to that point is costed at list price from the agent's own record of the session. That figure is an estimate, usually within a few percent of what the agent would have reported. Runs that failed or were cancelled before this release were not counted, so work from before then reads cheaper than it was.

A run on a model the agent's own price list does not include is costed at that model's published rate, not at the agent's guess; a model with no published price either is shown as unpriced. Runs recorded at such a guess before this correction, including Claude Sonnet 5 runs recorded at 1.5 times their cost before Repave 0.1.123, are corrected when Repave is upgraded.

This includes the agent you drive yourself in a Repave IDE session, which appears in its own Repave IDE column. A session opened on a feature's worktree is counted against that feature; one opened on the modernized checkout is counted against the project, as is work in a worktree shared by a module's features — that session does not belong to any one feature, so its cost is not attributed to one.

Sessions are counted from the release that introduced this onwards. Repave IDE work done before then is not included, so a feature built by hand earlier will read cheaper than a comparable one built today.

Jev usage

If your agents use TypeSafe Jev for small judgments, the cost view shows a Jev Usage panel: how many asks were made, how many questions they carried, and the input and output tokens. Asks are broken down by feature; asks on features that have since been deleted are grouped into a single row, and asks made outside any feature, such as from a terminal on the modernized checkout, are listed as Not attributed to a feature.

Jev usage is shown in tokens, not dollars, and is not counted in the project's cost or token totals.

Failed asks are counted and named by reason: the deployment could not reach TypeSafe, TypeSafe rejected the key, or TypeSafe returned an error. Agents carry on without Jev when an ask fails, so this panel is where a key that has stopped working shows up. The panel only appears once the project has asked Jev at least once.

Per feature, per stage

A feature's cost is attributed to the step that spent it, and broken down by pipeline stage on the cost view. That breakdown is the useful part. "This feature cost X" tells you little; "most of it went on review cycles" tells you where to look.

Cost can also be split by model, which is how you find out whether a more expensive model is earning its place on a particular stage.

The features table

Every feature with its cost, paginated, showing each feature's status alongside. The status column matters when reading the numbers: a feature that cost a lot and is failed is a different story from one that cost a lot and is merged.

The table stands alone — it does not require the rest of the view to make sense — and it downloads as an Excel workbook when you need the numbers somewhere else.

Controlling spend

Costs are a report, not a control. The controls are elsewhere:

• Budgets stop agents before they overspend, rather than after: a budget per run, set per agent, and lifetime budgets per feature and per project — see Budgets.
• Model and advisor choice per agent, and prompt cache lifetime, whose cost the settings table shows you directly.
• Validation passes are capped at a configurable number of agent calls.

See Project settings.