Skip to content

v0.8.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 12 Oct 03:40
· 9 commits to main since this release
v0.8.0
This tag was signed with the committer’s verified signature.
kwisatzh Vijay
GPG key ID: 9272956FD1758DDC
Verified
Learn about vigilant mode.
44967bd
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v0.8.0 Release Notes

Review a LiteLLM session budget from the spend logs you already retain. See
which sessions exceed a supplied USD cap, how recorded spend and call counts
are distributed, and which session IDs appear under multiple keys, teams or
agents. Everything runs locally on a request-level export.

Review Your Sessions

fleetdiff investigate --litellm-spend spend.csv --session-budget-usd 1.00

The synthetic example
opens with:

3 of 4 sessions with complete recorded spend exceed the supplied USD 1.00 cap.
  • Exact decimal recorded-spend and call distributions show p50, p95 and maximum,
    each with its own sample count. Largest sessions use keyed aliases.
  • Cap review lists sessions above the supplied cap and their recorded rejections.
    Incomplete spend subtotals are counted separately.
  • Shared session IDs and long recorded spend spans help review the pinned
    LiteLLM limiter's counter unit and configured window.
  • Sessions with rows outside the selected period are labeled clipped. Activity
    within one counter window of an uncovered file edge is labeled possibly clipped.
  • --period selects a day or UTC interval and also works without a cap. Otherwise
    the report selects complete UTC days, falling back to the labeled observed
    request-start span for shorter files. Both flags imply session grouping.

Upgrade

Fetch the matching SQL projection to
include agent_id and rejection_kind. The latter classifies session-budget,
session-iteration and other budget rejections without exporting the raw error
message. These fields are pinned to the tested LiteLLM v1.104.0 writer.
Older exports remain readable, with rejection coverage shown as unknown;
fleetdiff v0.7.0 safely ignores the additional columns.

Single-period review groups by the literal session ID alone, matching the pinned
limiter's counter key. Existing two-period --group-by session comparisons stay
key-scoped and byte-identical. This release changes no summary wire format,
normalization, hashing dependency or collector configuration.

Reading the Review

These are observed session totals. A session ID need not identify a completed
task run. Missing spend stays outside the complete-spend percentiles; sessions
with recorded rejections remain in the observed distributions and are counted
explicitly. Missing identities remain unattributed.

An above-cap total is a review item, not an enforcement verdict. Check last-call
or parallel-call overshoot, the counter window, agent budget configuration and
failure logging. Recorded duration alone does not establish that a reset occurred.
Source-reported spend is not an independently verified invoice.

The reader accepts up to 10M rows and 16 GiB. Exact session review retains the
250,000-session and 1,000,000-membership limits, with guidance when singleton
sessions suggest request-level IDs. See the session guide.

Install and Verify

Use the verified installer with v0.8.0.
Linux and macOS archives cover AMD64 and ARM64. The release workflow checks
packaged commands on each platform, includes dependency licenses and an SPDX
SBOM, and signs build provenance for the archives and checksums.

Assets 10
Loading