Changelog
Changelog
What we've shipped, in order. We version Diaflow as we build it — pre-1.0 means we're still adding core capabilities. Expect rapid iteration.
v0.9 — Latest
v0.8
v0.7
...and 5 earlier releases
Parallel execution and subagent billing
- NewParallel branch execution in pipelines. Fan-out/fan-in patterns now supported.
- NewSubagent run attribution in billing dashboard. See cost per subagent per parent run.
- ImprTrace explorer: collapse/expand branches in parallel runs. Much easier to read.
- FixMemory read/write race condition in high-concurrency runs. Resolved with write-ahead log.
- FixGemini 1.5 Flash tool calling returning empty string on complex schemas.
SSO, audit logs, and enterprise controls
- NewSSO/SAML support for Pro and Enterprise plans. Works with Okta, Azure AD, and generic SAML 2.0.
- NewAudit log export (CSV + webhook). Every API call, run start/end, and user action logged.
- NewAPI key scopes: read-only, execute-only, admin. Fine-grained access for CI/CD pipelines.
- ImprBudget alerts now support per-agent and per-pipeline thresholds, not just account-level.
Qdrant and pgvector memory backends
- NewQdrant memory backend. Drop-in alternative to Pinecone for self-hosted setups.
- Newpgvector support. Store embeddings in your existing PostgreSQL instance.
- ImprEpisodic memory retrieval 2.3x faster with compound indexing strategy.
- FixMemory TTL expiry not applying correctly to session-scoped keys.
Webhook triggers and scheduled runs
- NewWebhook triggers: fire agent runs from any external event source with a HTTP POST.
- NewCron-scheduled runs. Define a schedule in cron syntax — no external scheduler needed.
- ImprTool call retry backoff now exponential with jitter by default (was fixed 1s).
- BreakPipeline.run() now returns a RunResult object instead of raw dict. See migration guide.
Latency heatmaps and cost tracking
- NewLatency heatmap view: per-step p50/p95/p99 across the last 1,000 runs.
- NewLLM cost tracking: token counts and USD cost per run, per model, per step.
- NewBudget alerts: set a monthly LLM spend cap and get notified before you hit it.
Anthropic Claude and Mistral support
- NewClaude 3.5 Sonnet and Haiku support. Pass your Anthropic API key, works immediately.
- NewMistral Large and Mixtral support via Mistral AI API.
- ImprTrace view now shows model name and provider per LLM step.
Node.js SDK and REST API v1
- NewNode.js/TypeScript SDK (beta). Full parity with Python SDK for core features.
- NewREST API v1: programmatic run trigger, status polling, and trace retrieval.
- FixCircular reference in pipeline definition now caught at definition time with clear error.
Visual trace explorer and Pinecone memory
- NewVisual trace explorer in the dashboard. Step-by-step replay for every run.
- NewPinecone memory backend integration. Semantic retrieval for multi-turn agents.
- NewRetry policy config per step: max_retries, backoff_seconds, fallback_agent.
- ImprPython SDK 0.2.0: type hints, async/await support, context managers for runs.