Release notes¶

Unreleased¶

Made workflow failures visible directly in replay and distinguished diagnostic attention from successful step completion.
Added readable goal revisions and native, URL-backed filters to execution lists.
Improved workflow-canvas contrast with a navy background and high-visibility connectors for normal, branch, retry, and fallback paths.

Replaced configuration v1 with a smaller vendor-neutral v2 project index, external contract and workflow files, and explicit named requirement facts.
Added contract-authored failure diagnostics and validated workflow investigation targets; the dashboard can open the declared evidence node without generating an explanation.
Added a deterministic CUAD configuration and direct/LangGraph example that exercises both achieved and failed contract requirements without external provider calls.
Split dashboard routes and stable vendor libraries into cacheable production chunks, reducing the initial JavaScript bundle below the build warning limit.
Added stable URL-backed dashboard drilldowns for filtered goals, runs, system health, measured spend, charts, and issues.
Added a public, deterministic evidence-bundle v1 export for canonical execution records and existing OSS diagnostics.
Exposed the same neutral contract through AnalyticsRepository and the versioned dashboard read API.
Added an oldest-supported v1 compatibility fixture and contract regression coverage.

Added configuration v2 project, contract, workflow, requirement-fact, and contract-authored investigation support.
Made witdem-sdk init generate a minimal .witdem/skills/witdem coding-agent skill with opt-in .agents/skills/witdem discovery.

Added concrete OpenAPI response schemas for every documented dashboard endpoint.
Documented run summaries, execution graphs, operation identity, measurements, workflows, evaluations, and issue responses while preserving additive fields.
Grouped Swagger operations by system, metadata, analytics, runs, workflows, and evaluations for easier API discovery.

Added native MCP protocol normalization for connection initialization, capability discovery, resource reads, prompt retrieval, and tool execution.
Preserved MCP as the invocation interface while retaining the semantic operation family, type, role, provider, and nested downstream work.
Counted canonical MCP tool executions in run summaries even when telemetry arrives as generic OpenTelemetry operations.

Separated execution containers, control-plane orchestration, and work-plane operations so workflow roots no longer appear as unknown AI operations.
Added explicit model applicability, participant deduplication, and linked child activity to operation profiles.
Added safe taxonomy reprocessing from immutable raw telemetry with witdem taxonomy reprocess.
Reclassified the preserved CUAD stress dataset with taxonomy v2 and verified direct Anthropic and OpenAI SDK paths alongside framework-emitted operations.

Added an extensible, versioned AI operation taxonomy that keeps operation type separate from interface, provider, implementation, framework, and role.
Added first-class direct OpenAI SDK instrumentation for Responses, Chat Completions, embeddings, synchronous and asynchronous calls, streaming usage, and tool-call identifiers.
Preserved nested provider operations inside LangGraph nodes and expanded canonical measurements across retrieval, tools, quality, media, memory, human work, orchestration, and custom operations.
Verified direct Anthropic and OpenAI SDK calls, with and without LangGraph, against the CUAD contract-review workload using live provider APIs.
Updated the dashboard to display unfamiliar operation families and types without requiring provider-specific UI changes.

Added the raw-first immutable corpus and Duckle raw-to-serving worker.
Added canonical serving projections and repository-backed dashboard reads.
Bundled the React dashboard and versioned FastAPI dashboard API.
Added OpenTelemetry-only and SDK-enriched integration paths.
Added one-point SDK wrappers across LangGraph, LangChain, OpenAI Agents, Anthropic, Haystack, Claude Agent streams, and generic provider calls, with explicit result mapping and LangGraph 1.x support.
Added explicit runtime/application/product-goal outcome semantics.
Added managed pricing with unavailable-cost diagnostics.
Added authenticated ingestion and an opt-in TLS remote deployment profile.
Added runtime examples for OpenAI, Anthropic, LangChain, LangGraph, Haystack, Azure, Bedrock, Vertex, and Ollama.
Consolidated overview reads and redesigned Compare, Workflows, Issues, and Runs pagination.
Removed the retired Streamlit dashboard and reconciled repository documentation.

Fixed the worker container's false-negative health status under DuckDB lock contention.
Added an explicit lock-free process-liveness probe for workers in local and NPX Compose deployments.
Replaced a machine-speed timing assertion with a deterministic event-loop concurrency check.

Kept SDK and OTLP ingestion responsive during parallel durable writes and ordinary ELT processing.
Added bounded grouped persistence, explicit retryable backpressure, and configurable SDK delivery deadlines.
Added parallel durability and maintenance-contention regression coverage with no immutable-corpus changes.

Rotated the production release-manifest signing key after the original private key was unavailable.
Preserved the 0.1 protocol, workflow, corpus, and SDK compatibility contracts.

Hardened package and container reproducibility with locked runtime constraints and pinned build tools.
Made platform and SDK release identities tag-verified and immutable across PyPI, npm, and GHCR.
Added cross-version SDK/server compatibility tests and dependency security automation.
Assigned the shared witdem command exclusively to witdem-analytics; SDK-only commands use witdem-sdk.