What shipped, dated.

Chunk usage attribution and loop detection on every trace

Retrieval spans now record which chunks the completion actually used, computed from the output after the fact. Tool spans that repeat the same call with the same arguments three or more times in one trace are flagged as loops, shown on the waterfall and available as an alert condition.

• New used_in_answer field on retrieval chunks; filter traces by “unused context ≥ 50%”.
• loop.count attribute on traces; alert condition loops ≥ N.
• Waterfall renders up to 400 spans per trace without virtualisation hiccups (was 120).

Gates can compare against main with a delta, not just an absolute floor

Judges are noisy; a hard floor of 0.90 blocks honest PRs on a bad judge day. A delta gate says “no worse than main by more than X” and uses the same dataset version and model pins for both sides, so the comparison is apples to apples.

gates:
  - { metric: relevance, delta: -0.01 }   # vs main, same dataset@version

Cached-token accounting for prompt caching was undercounting cost by up to 12%

When a provider reported cached input tokens, the SDK priced the whole prompt at the cached rate. Costs for affected completions have been recomputed for the last 30 days; your dashboards may show a small upward correction dated to the original span.

Signed eval reports and a human review queue

Every gate run now produces a PDF with cases, scores, model versions and dataset hashes, signed with a per-project key you can verify offline. The review queue routes low-confidence or flagged cases to named reviewers; their labels calibrate judges and become dataset ground truth.

• Reviewer agreement with the judge shown as Cohen's κ per judge.
• Queue SLAs with Slack reminders.

Queries over 30-day windows are 3.1× faster

We moved the span index to a columnar layout partitioned by tenant and hour and pushed attribute filters into the scan. p95 for a 6-filter query over 2.4B spans dropped from 1,204 ms to 388 ms. Saved views load in under 200 ms.

Vercel AI SDK and LiteLLM auto-instrumentation

Both libraries are patched on import. Streaming responses record time-to-first-token and per-chunk timestamps; tool calls declared through either SDK become tool.* spans with arguments and results.

Flush on process exit in serverless runtimes

On Lambda and Cloud Run the last batch could be lost when the runtime froze before the 200 ms flush. The SDK now flushes synchronously on atexit / beforeExit and exposes tw.flush() for explicit control. p95 overhead is unchanged at 3 ms.

Anomaly bands for drift alerts

Instead of a static threshold, an alert can now watch a seasonal band learned per feature and hour of day. Alerts fire when pass rate, score distribution, refusal rate or tokens per request leave the band for two consecutive windows. Failing traces are promoted to a dataset tagged with the alert id.

⌘K everywhere

The command palette now jumps to traces by id, datasets by name, saved views and settings. Fuzzy matching, recent items first. Press ⌘K on this site to try the same thing.

Self-hosted Tracewell, generally available

A Helm chart that runs the collector, store, query engine and console in your Kubernetes with your Postgres and object store. Same SDKs, same console, same gates. Upgrades are rolling; the first install takes about forty minutes.

helm install tracewell oci://registry.tracewell.dev/charts/tracewell \
  --set store.bucket=s3://acme-tracewell --set postgres.url=$PG_URL