fix(claude-team): read seat spend before the billing month closes - #2526
Merged
Conversation
The seat endpoint reports the month in progress and carries no history, so a month's spend freezes at the last read inside it and everything after that read is lost permanently — a later sync answers for the new month and cannot go back. Reading at 04:00 UTC therefore dropped the last twenty hours of every month by construction, and a failed run on the last day dropped a day more. Measured on a populated instance: July's final read landed 2026-07-30 04:02 UTC, leaving 44 hours unread, and the month came out $634.52 short of the reference implementation's $13,415.12 — 13 of the 57 seats with a non-zero amount, every one of them short rather than over. June's final read was 2026-06-24, leaving 164 hours. 23:50 leaves ten minutes instead of twenty hours, and stays off-peak: the other connectors occupy 02:00 to 06:00 and 11:00, and the customer's proxy runs a single browser instance. Both of this connector's streams tolerate the move — the code-metrics stream walks a trailing window and re-reads it. The check names the same failure as a warning on the month that just closed, because losing the tail is silent otherwise: nothing else notices that a figure is short, and no re-sync can fix it once the month is over. Signed-off-by: Gregory Gogin <grigoriy.gogin@constructor.tech>
|
Important Review available on request
Reviews should be triggered manually for repositories with fewer than 10 stars. Select Trigger review above or comment ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
3 tasks
Contributor
Author
|
@coderabbitai review |
|
Contributor
Author
|
@coderabbitai review |
|
Contributor
|
failed checks |
…LECT
The comment explaining why `closed_month` is not aliased sat between two SQL
tokens and used trimming Jinja delimiters. `{#-` eats the whitespace before the
tag and `-#}` the whitespace after, so the rendered SQL read
`closed_monthSELECT` and ClickHouse rejected the check with SYNTAX_ERROR.
An erroring data test fails the whole dbt build, so the ai shard lost the
ai_cost and seat-history fixtures with it — three red tests, one cause. The
source file looks correct; only the compiled SQL shows the defect.
Untrimmed delimiters keep the newline. The comment stays where it is: it
explains the line below it, and moving it would cost the reader the reason.
Signed-off-by: Gregory Gogin <grigoriy.gogin@constructor.tech>
…h-tail Signed-off-by: Gregory Gogin <grigoriy.gogin@constructor.tech> # Conflicts: # src/ingestion/connectors/ai/claude-team/descriptor.yaml
github-merge-queue
Bot
removed this pull request from the merge queue due to failed status checks
Aug 17, 2026
hello1101n
enabled auto-merge
August 17, 2026 16:43
github-merge-queue
Bot
removed this pull request from the merge queue due to a conflict with the base branch
Aug 17, 2026
…h-tail Signed-off-by: Gregory Gogin <grigoriy.gogin@constructor.tech> # Conflicts: # src/ingestion/connectors/ai/claude-team/descriptor.yaml
…h-tail Signed-off-by: Gregory Gogin <grigoriy.gogin@constructor.tech>
hello1101n
approved these changes
Aug 18, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #2525. Part of #2479.
/api/organizations/{org}/overage_spend_limitsreports the billing month in progress and carries no history. A month's spend is therefore frozen at the last read that happened inside it, and everything after that read is lost for good — the next sync answers for the new month and cannot go back.Reading at 04:00 UTC dropped the last twenty hours of every month by construction. A failed or missing run on the last day dropped a day more.
Measured against the reference implementation
July, 145 seats, compared per address:
Our final read of July landed 2026-07-30 04:02 UTC — 44 hours unread. June's landed 2026-06-24, leaving 164 hours.
The transformation is exact: joining bronze to silver over the same month gives zero mismatches on both the amount and the ceiling. Nothing is miscomputed; the rows simply stop before the month does.
What lands
descriptor.yaml—50 23 * * *instead of0 4 * * *, leaving ten minutes of a month unread instead of twenty hours. Still off-peak: the other connectors occupy 02:00-06:00 and 11:00, and the customer's proxy runs a single browser instance. Both streams tolerate the move — the code-metrics stream walks a trailing window and re-reads it, so it is unaffected by the hour.assert_ai_overage_read_before_the_month_closed— a warning when the month that just closed was last read more than six hours before it ended. The schedule puts the final read ten minutes before the boundary; one failed run there makes it a day. Only the just-closed month is judged: an open month is still being read, and an older one can be neither re-read nor acted on.No descriptor version bump: the reconciler applies the CronWorkflow unconditionally (
kubectl apply, no-op when unchanged —lib/reconcile.sh), and ADR-0015's semver governs the connector's spec and data shape, neither of which moves here.Test plan
dbt parse— clean; the new test is in the manifest astest.ingestion.assert_ai_overage_read_before_the_month_closed.hours_unread = 44. It catches the defect that prompted it.pre-commit run --fileson both paths — yamlfmt and the file-hygiene hooks pass.An earlier draft of the check aliased the closed month to
period_monthin its SELECT; the alias shadowed the column, the WHERE compared the alias with itself, and every month collapsed into one group whose latest read belonged to the open month — inverting the result to a silent pass. The comment in the file records it so the shape is not reintroduced.