+Composite AND policies now short-circuit on tool-pattern scope, an out-of-scope call no longer depletes the policy's token/USD/step budget, so a later in-scope call trips exactly when its own usage warrants
+USD budget counters pipeline their INCRBYFLOAT + EXPIRE atomically (parity with the integer counters), no window where a USD key could be left without a TTL
+The enforcement gateway returns a generic 502 on an unreachable provider (the upstream error detail is logged server-side, not echoed to the caller)
+Docs: clarified that gateway streaming calls enforce token budgets pre-call (via max_tokens) only, and that gateway trace_created is metered after the provider responds
1.24.0
Onboarding & growth
+Guided quickstart, the empty dashboard now walks you from API key → install → env var → first trace, with copy-paste Python/TypeScript snippets and a live "waiting for your first trace…" state that flips the moment one arrives
+One-click demo workspace, load sample traces, a detection, and a tape into your own project to explore every feature before shipping a span; clearly badged, one-click removable, and never counted toward your usage
+Trending tapes, a public /tapes/trending gallery (cached, public-only) plus "forked N times" social proof on shared tapes; fork any public tape to replay it in your own account
+Clearer empty states across the app explaining what each feature does and the one action to populate it
+Activation funnel, first-party milestone tracking (signup → SDK installed → first trace → detection → replay → tape shared) with an internal admin view; no third-party analytics
1.23.0
Starter plan + pricing refresh
+New Starter plan ($29/mo · $290/yr), 10,000 traces/mo, 30-day retention, 100 fork replays, 25 prove-the-fix runs, and Cassette VCR deterministic replay. The cheapest way to get the replay workflow
+Enterprise is now a self-serve annual plan ($24,000/yr · billed annually), no monthly Enterprise product
+Pricing, landing, and billing pages now render every plan from a single shared source of truth, so quotas shown always match what the API enforces
+Existing Pro/Teams subscribers are grandfathered, invoices always reflect the amount actually charged, regardless of list-price changes for new customers
+Billing fix: every Dodo product now maps to the correct plan (a Starter purchase grants Starter, never Pro); unmapped products alert the team instead of silently defaulting
1.22.0
Reliability Platform
+Enforcement (circuit breakers), budget ceilings, loop breakers, debounce windows, and tool-pattern action policies that block or hold an agent action BEFORE it runs (not just flag it after). Server-side decision engine on atomic, fail-closed Valkey counters, a low-latency /enforcement/check pre-call gate, policy versioning + a hold-for-approval queue, and a new Enforcement page
+SDK pre-call gate (Python + TypeScript), local step/token/USD-per-run ceilings enforced offline with zero network, optional server-policy consult, and a typed RetraceEnforcementError that stops the run instead of silently skipping the call
+MAST failure taxonomy, failed traces are auto-classified into the 14 MAST failure modes (an LLM judge metered separately from your AI quota, idempotent per trace), surfaced as detections and an Insights → Failure Taxonomy breakdown
+Prove the fix, re-run a trace (fork → replay → first-divergence diff → judge verdict) to confirm a fix improved/regressed/unchanged; hypothesis testing via typed substitutions (swap model/prompt/tool-output), a dashboard Verify-fix button, a `retrace traces verify-fix` CLI gate, and an AI-widget tool
+Multi-agent traces, record which agent produced each span (Python `retrace.agent(...)`, TypeScript `withAgent`, OTel gen_ai semconv), see the agent topology graph on multi-agent traces, and catch agent ping-pong loops (all plans) plus reasoning-action mismatch (FM-2.6) and task derailment (FM-2.3) detectors (Pro+)
1.0.0
Advanced AI Agent Analytics Engine
+Deep Fork Replay, context injection flows fork output into subsequent prompts, tool output mocking, batch sweeps with multiple variants
+Runtime Guardrails, cost budgets, loop detection, context overflow, latency budgets with halt/alert/throttle actions enforced in real-time via WebSocket
+Sessions, group multi-turn conversations by session_id, execution DAG graph endpoints with causal ordering
+Eval CI/CD Gate, POST /evaluations/:id/gate returns pass/fail with regression detection against baselines
+Multi-Agent Tracing, agent_id on spans, vector clocks for causal ordering, cross-agent interaction graphs
+Hallucination Detection, tiered pipeline using KL divergence, grounding scores, and entropy analysis on every LLM call
+Adaptive Guardrails, LinUCB contextual bandit learns optimal thresholds per trace-type from trigger history
+Trace Similarity, Sinkhorn optimal transport (Wasserstein distance) finds similar failures across all traces
+Embedding-Influence Scores, per-span heuristic ranking from embedding similarity to the final span (experimental; a similarity proxy, not causal inference)
+LTL Safety Verification, formal property checking (always/never/eventually/precedes) on traces
+Spectral Gap Heuristic, Laplacian eigenvalue gap of the agent interaction graph compared to a fixed baseline (experimental; a threshold heuristic, not learned anomaly detection)
+Fork Preferences & RLHF Export, record preferred fork outcomes, export as DPO training pairs
+Fork-Position Baseline, estimates fork success/cost from fork POSITION via inverse-distance weighting over past forks (experimental; a position baseline — not counterfactual prediction, and not nearest-k)