4d235fee47
Spend per canonical project x branch, keyed by the project identity the spend-flow and By project views already use, so identical branch names in different projects never merge. Reuses the aggregateByBranch rules: last-seen branch carried forward across turns (the parser resolves the carry across a range boundary from pre-range evidence), only ever-branched sessions participate, and turns before a session's first branch stay an explicit Unknown row. Rows carry cost, behavioral calls, token components, distinct sessions (identity-based, rows overlap), activity window, per-session contributions, recorded worktree evidence (historical working directory, never today's checkout), and a coverage split: known / Unknown / no-branch-data sources. Same --period/--provider/--from/--to/--project/--exclude semantics as spend --format flow-json. docs/by-branch.md documents the lens.
2.8 KiB
Spend by branch and worktree
The Spend view in CodeBurn Desktop has a By branch panel: pick a project and see where its AI spend went — per git branch, with the recorded worktrees and the individual sessions behind each branch.
What each branch row shows
- Cost attributed turn by turn. A session that switched branches mid-way
(for example
feat/auth→fix/parser) contributes each turn's cost to the branch that was active when it happened — the same session appears on both rows with its own slice, never the whole cost twice. - Calls (behavioral requests) and tokens per component (input, output, reasoning, cache read, cache write) summed from the same turns.
- Distinct sessions and the activity window (first–last attributed call in the selected range).
- Expanding a row shows the recorded worktrees (from the session's provider-recorded working directory — the worktree path itself, not today's checkout) and each session's contribution. Click a session to inspect its real detail: id, provider, working directory, models, tokens, activity.
Coverage (the fine print, kept visible)
Branch metadata is captured per turn and, today, only Claude transcripts carry it. The coverage note under the rows therefore splits the project's spend:
- On branches — spend attributed to named branches.
- Unknown, before first branch — spend inside a branch-bearing session that
happened before its first recorded branch. A branch set before the selected
range still carries forward into the range; only spend that genuinely
predates every recorded branch lands here. It is Unknown, never relabeled
as
main. - No branch data — spend from sessions that never recorded a branch (non-Claude sources). Listed with the providers and session count.
A session on several branch rows counts once in the distinct-session total (rows overlap; they are not summed). Costs do reconcile: known + unknown + no-branch-data equals the project's total in-range spend.
The lens always respects the period, provider, and calendar range selected at
the top of the window; the project picker narrows the display (choose All
projects to see every project's branches, labeled project / branch, so two
projects that both have a main never merge).
CLI
The same report is available from the terminal:
codeburn spend --format branch-json # whole period default
codeburn spend --format branch-json --period week --provider claude
codeburn spend --format branch-json --from 2026-07-01 --to 2026-07-11 --project alpha
The JSON carries projects[] (canonical project id, branch rows, per-session
contributions, worktree evidence, coverage) and totals (identity-based
distinct sessions plus the same coverage split). It reuses the --project /
--exclude filtering introduced for the CLI in #1285.