What work costs

Every turn a chat finishes is counted against the branch that was checked out when it ended, and against the autopilot item it worked for, if any. Tori adds those up into what a branch, an item, an issue and a pull request cost, shows the figure where you already look, and lets you cap what one autopilot item may spend.

The figures count from the version that brought them in. Turns from before that are in your session and project totals but count toward no branch or item, so nothing is guessed after the fact.

Where the figures show

Each figure is the dollars spent, for example $2.25. Hover it for the sentence, for example $2.25 spent on this branch across 3 turns. A branch with nothing recorded shows nothing.

The figures update a moment after each turn ends, with no reload. A worktree’s spend still counts after the worktree is deleted, since it is kept by repository rather than by folder.

A chat that switches branch partway splits its spend: the turns before the switch stay on the first branch.

To hide every figure, turn off Show what work costs in Settings > Chat > Spending. That hides the three above and a waiting item’s estimate. The spending limits still apply, and spend is still recorded, so the figures come back as they were when you turn it on again.

Agents that report no cost

Claude reports what each turn cost. Codex, OpenCode and the other agents report how full the context is but not what it cost, so their figure is the turns instead, for example 5 turns. Hovering it says no cost reported: this agent reports context, not dollars. Tori never shows those as $0.00, which would read as free.

From a shell

tori budget takes two more flags:

tori budget [<id>] [--folder <path>] [--item <id>] [--pr <n|url>] [--json]

• --item <id> adds an item line with what that autopilot item cost. For an item that came from an issue it also adds an issue line, which counts every item from that issue plus the branches linked to it, each turn once.
• --pr <n|url> adds a pr line with what the pull request’s head branch cost, named by number or by its URL.

item     $4.50  item-k3j9
issue    $6.10  212
pr       $4.50  #318 fix-launch-flicker

An item or branch whose agent reports no cost prints 12 turns, no cost reported, and one with nothing recorded prints nothing recorded. With --json the same figures come back under item, issue and pr, each with costUsd, which is null when no cost was reported, plus turns and tokens. The budget tool takes item and pr the same way.

Cap what one autopilot item spends

Stop an autopilot item after, in Settings > Chat > Spending, is a dollar ceiling across every session the autopilot runs for one item, retries and second opinions included. Leave it blank for no limit, which is the default.

A project or a single item can set its own ceiling, as itemUsd. Tell the autopilot, for example cap items in this project at $8 or let #212 spend up to $20, and it records the ceiling on the project’s contract or on the item. The first one set wins, in this order:

1. The item’s own ceiling.
2. Its project’s ceiling.
3. Stop an autopilot item after.

The ceiling is checked between turns, like the other ceilings: the turn that crosses it finishes. After that:

• The autopilot can no longer steer that item’s workers or start a new one for it. It is told once that the item went over, with what it spent against the ceiling, for example $8.40 of $8.00, and brings it to you.
• A pull request change on the item does not wake its worker.
• Raise the ceiling and the item can carry on.

The item ceiling applies while the autopilot drives. A chat you run by hand is held only by the per chat and per project ceilings.

What a waiting item will likely cost

A proposed or queued item that came from an issue shows a projection on its row under Waiting at the dock: ~$3.50, the median of what finished items like it cost. Hover it for how many items it is drawn from.

“Like it” means the same project and an issue body of the same size, counted in characters:

Only finished items count. A failed item, an item someone took over, and an item one of whose sessions reported no cost (an ACP agent) are left out, as are items from before this version. Until five items of the same size have finished in the project, the row says no history yet.

The size is read once, when the item is made, from the issue as it was then. Items from schedules, loops and review requests have no issue body, so they show no projection.