v0.7.0

v0.7.0 Release Notes

Compare recorded cache reuse before and after a change, and find the keys,
teams or sessions whose uncached input grew. investigate now includes cache
reads, cache writes and their request coverage for LiteLLM exports and collector
summary windows. Significant changes appear directly below the headline.

Upgrade

The LiteLLM SQL export now includes version-pinned cache-read, cache-write and
reasoning columns by default. Fetch the v0.7.0 projection to get those fields;
set -v usage_details=false to retain the previous smaller export. Missing
writer paths stay NULL. The tested LiteLLM v1.104.0 Messages writer can fill
absent detail with zero, which remains recorded zero with unknown origin.
The export guide lists each call type's
verified behavior.

This release adds a report section and changes the export default, hence the
minor-version increment. Existing exports remain readable. JSON adds explicit
cache quantities, coverage and uncached-growth rankings; unknown quantities are
null. Text uses human-readable coverage and omits zero exclusion counts. Use
JSON for integrations rather than parsing text.

Collector summaries already contain aggregate cache totals and quality counts.
Their per-model and contributor cache breakdowns remain unavailable; the report
names that gap. No collector upgrade, wire-format change, hash-key rotation or
window reset is required. Token normalization, reported spend and identity
semantics remain unchanged, using llm-sketchkit v0.2.3 with x/text v0.41.0.

Read The Result

The cache example keeps total input flat while cache-read share
falls from 92% to 61% and uncached input increases by 620 tokens. Both periods
need complete cache-read coverage for the second headline. It appears at a
five-percentage-point share change, or an uncached-input change of at least
100 tokens and 20% relative to the before period. From zero, the 100-token
minimum applies.

Rankings retain keyed identities and deterministic bounds. Missing or invalid
cache details are excluded and counted separately. The report measures usage,
not the cause or price of cache loss; a falling share directs the next step to
provider cache diagnostics and the gateway route's preservation of cache markers.

Inspect Without Go

examples/inspect/synthetic-traces.json is now checked in and reproduced byte
for byte in CI. Run it with the verified executable:

curl --fail --location --proto '=https' --tlsv1.2 \
  https://raw.githubusercontent.com/llm-measurement/fleetdiff/v0.7.0/examples/inspect/synthetic-traces.json \
  --output synthetic-traces.json
./fleetdiff-install/fleetdiff inspect synthetic-traces.json

Install And Verify

Use the verified installer with v0.7.0.
Linux and macOS archives cover AMD64 and ARM64, with dependency licenses,
checksums, an SBOM and signed build provenance. The release workflow exercises
the packaged commands on all four platforms and the pinned LiteLLM writer,
read-only database export and importer on Linux AMD64.