What shipped, versioned honestly.

1. v2.3.0

   minor Jul 8, 2026

   Guardrail dry-run mode & policy diffs

   • + New `dry-run` enforcement mode: policies evaluate and log verdicts without blocking, separately from `observe` (which skips enforcement telemetry).
   • + Policy diff view: compare two policy versions and see which historical runs each would have blocked.
   • + Console: would-have-blocked report is exportable as CSV.

2. v2.2.0

   minor Jun 16, 2026

   Eval gates in CI

   • + `trellis eval --gate` exits non-zero when a candidate regresses beyond per-suite thresholds — wire it into any CI as a test step.
   • + Per-suite regression thresholds in `trellis.yaml` (absolute or relative).
   • + GitHub Actions and Buildkite examples in the docs.

3. v2.1.1

   patch Jun 2, 2026

   Trace ingestion fixes

   • + Fixed a backpressure condition where bursty fleets (>2k spans/s from one workspace) could delay trace availability by up to 90 s.
   • + Span deduplication for at-least-once SDK delivery — replayed checkpoints no longer double-count tokens in rollups.

4. v2.1.0

   minor May 19, 2026

   Forensic retention tier

   • + New observability tier: 90-day retention with full request/response payloads and eval snapshots.
   • + Payload encryption at rest with customer-managed keys (CMK) on the forensic tier.
   • + Retention is configured per workspace; changes are audit-logged.

5. v2.0.0

   major Apr 28, 2026

   Runs API v2

   • + BREAKING: `step` objects are now `span` objects across the API and SDKs, aligning our trace model with OpenTelemetry semantics.
   • + BREAKING: `run.cost` moves from cents (integer) to a decimal-string USD amount with per-span breakdown.
   • + v1 endpoints continue to work until 2026-11-01; the migration guide maps every renamed field.

6. v1.9.0

   minor Apr 7, 2026

   Human-approval steps

   • + Runs can pause on an `approval` span — parked runs hold no compute and cost nothing while waiting.
   • + Approvals via console, API, or Slack; the approver's identity and timestamp land in the trace.
   • + Configurable timeout with fallback branch (auto-reject, escalate, or continue).

7. v1.8.0

   minor Mar 17, 2026

   Cost attribution rollups

   • + Tag runs with arbitrary metadata (`tenant`, `feature`, `experiment`) and get cost/latency/outcome rollups per key.
   • + Monthly invoice line items can be split by one metadata key — finance teams asked, we shipped.

8. v1.7.1

   patch Mar 2, 2026

   Python SDK retry fix

   • + Fixed a retry storm where a 429 on checkpoint write could trigger immediate un-backed-off retries in `trellis-python` 1.7.0.
   • + SDKs now honor `Retry-After` on all control-plane endpoints.

9. v1.7.0

   minor Feb 18, 2026

   OpenTelemetry export

   • + Native OTel export: every Trellis span forwards to your collector with standard GenAI semantic-convention attributes.
   • + Prebuilt dashboards for Datadog and Grafana in the integrations gallery.

10. v1.6.0

    minor Feb 4, 2026

    Cache-aware token accounting

    • + Cache reads are now metered separately from fresh input tokens in traces and invoices — multi-step agents typically see 60–90% of input tokens served from provider prompt caches.
    • + Cost projections in the console use your workspace's actual trailing cache-hit rate.

Older releases (v1.0.0 – v1.5.x) live in the console's release archive.