zenml

Breaking Changes

• The langchain integration now supports langchain>=0.3.5,<1.0 instead of the previously pinned 0.3.0. If you rely on a fixed LangChain version, update your environment constraints accordingly, and note that LangChain 1.x is still not supported by ZenML and may require you to stay on a compatible pre-1.0 release. PR #5341

Triggers and automation

• Dynamic webhook event filters: Webhook triggers can now match a provider’s raw event type and fields directly from the authenticated JSON body. This gives you an escape hatch for events or payload fields that ZenML’s typed event catalog does not expose yet, such as matching newly opened GitHub pull requests targeting a specific branch. PR #5247
• Forward-compatible run.trigger resolution: Snapshots updated with this release can resolve trigger types that are introduced in future ZenML versions. This reduces the need to refresh snapshots just to use newly added trigger types such as platform or webhook triggers. PR #5281
• Safer snapshot deletion for triggered pipelines: ZenML now checks whether snapshots are attached to triggers before deletion and warns users about unintended removals. This helps prevent accidentally breaking automations that still depend on a snapshot. PR #5284

Kubernetes, Run:ai, and workload ownership

• Synchronized OIDC Linux identities for Kubernetes workloads: ZenML Pro deployments using an enterprise identity provider can now carry synchronized linux_uid and linux_gid claims into workloads created by the Kubernetes orchestrator and Kubernetes step operator. This keeps the process UID and primary GID aligned with organization-managed Linux accounts, which is especially useful for shared filesystems and cluster security policies. PR #5301

• Workspace account annotations on Kubernetes and Run:ai workloads: Kubernetes Jobs, Pod templates, scheduled and dynamic executions, and Run:ai training workloads now include a resolvable zenml.io/workspace-user-id annotation. This gives platform and audit tooling the actual workspace account UUID while preserving custom annotations and leaving ZENML_USER_ID unchanged. PR #5326

• Owners for managed Kubernetes sandboxes: Kubernetes-managed Sandbox sessions can now be assigned owners. This makes sandbox sessions that are closely related to a step easier to track and helps avoid leaving orphaned sessions behind. PR #5336

• Workspace service account activation controls: Activation and deactivation are restored for deprecated workspace service accounts and legacy API keys, including keys retained after adoption. This supports migration cleanup and credential rollback while preserving the existing deprecation restrictions and deletion checks. PR #5340

• Transformers v5 support in the Hugging Face integration: The Hugging Face integration now supports Transformers v5 while retaining compatibility with v4. ZenML only registers the TensorFlow model materializer when running on Transformers v4, and the Accelerate integration now uses cross-version CPU configuration so users can upgrade their Hugging Face stack more smoothly. PR #5321

• TypeSafe Jev hooks integration: A new typesafe integration adds jev_failure_triage_hook and jev_run_summary_hook for storing Jev’s typed, probability-scored answers as ZenML run metadata. Users can attach structured failure triage and run summaries to their runs without building custom metadata plumbing. [PR #5312](https://github.com/zenml-io/

Breaking Changes

• Resource pools have been reworked to use ZenML Pro Resource Manager v2, and the old OSS resource-pool control-plane objects and related API/CLI/RBAC behavior have been removed. If you used legacy resource pool management in OSS or automation built on those endpoints/commands, you will need to migrate to the new resource request model and update any scripts or integrations accordingly. PR #4977
• ZenML now requires Click >=8.3.3,<=8.5.0. If your environment, plugin, or custom CLI extension pins an older Click version or depends on older Click behavior, update those dependencies to a compatible version before upgrading ZenML. PR #5145
• Dictionary outputs with non-string keys are no longer materialized through the JSON path. If you relied on reading such artifacts back as JSON-compatible data with keys converted to strings, update your code to preserve original key types or change outputs to use string keys for JSON interoperability. PR #5124
• Log retrieval now uses a paginated/filterable entries API instead of assuming a single full log blob, especially for external log stores. If you consume ZenML log APIs directly or built tooling around the previous response shape, update it to use the new log entries endpoint and pagination/filter parameters. PR #5158

• Reduced database growth from unused pipeline snapshots: ZenML now cleans up anonymous pipeline snapshots that are no longer referenced by runs, deployments, schedules, or templates. Snapshot creation is also atomic, so failed or interrupted creation attempts no longer leave partial snapshot records behind, and replacing named snapshots no longer keeps unused superseded snapshots around. This helps long-running ZenML deployments keep database size under control with less manual maintenance. PR #5166

• Dynamic pipeline DAG loading: Dynamic pipeline run DAGs now load correctly even when step records are returned out of dependency order. This fixes cases where the UI or API could fail to render a dynamic run graph despite the underlying run data being valid. PR #5263
• Larger cached API transactions: API transaction results are now validated against the compressed storage format instead of the uncompressed Base64 JSON representation. This allows large but highly compressible responses to be cached correctly, reducing unnecessary retries when an original request completes after the client timeout. PR #5262
• Capped exponential backoff overflow: Retry loops using capped exponential backoff no longer keep calculating ever-growing delay values after the maximum delay has been reached. This prevents OverflowError failures in very long-running retry scenarios while preserving the configured maximum delay behavior. PR #5265

What's Changed

• Add 0.96.4 to the migration tests by @github-actions[bot] in https://github.com/zenml-io/zenml/pull/5233
• Add version 0.96.3 to legacy docs by @github-actions[bot] in https://github.com/zenml-io/zenml/pull/5231
• Bump the minor-and-patch group across 1 directory with 5 updates by @dependabot[bot] in https://github.com/zenml-io/zenml/pull/5242
• Software factory example by @schustmi in https://github.com/zenml-io/zenml/pull/5246
• Fix lint failures introduced by the software factory example by @safoinme in https://github.com/zenml-io/zenml/pull/5251
• Fix the c465032f0ec0 DB migration script to be idempotent by @stefannica in https://github.com/zenml-io/zenml/pull/5252
• Su

Webhook-driven automation

• Webhook triggers for pipeline snapshots: ZenML now supports first-class webhook-driven automation for secure, project-scoped integrations with GitHub or custom systems PR #5169. You can attach webhook triggers to pipeline snapshots and launch them when matching external events arrive, including typed GitHub filters for merged pull requests, completed workflow runs, pushes, and published releases.

• ClickUp webhook provider: ZenML now includes a built-in ClickUp webhook provider for triggering automation from ClickUp task and list events PR #5216. The provider authenticates deliveries with ClickUp’s raw hex HMAC signature and supports string-based filters, so teams can connect ClickUp activity directly to ZenML workflows.

• Slack webhook provider: ZenML Pro now supports inbound Slack Events API callbacks as a first-class webhook provider PR #5221. Slack apps can authenticate events such as mentions, messages, reactions, and other automation-focused activity to trigger pipeline snapshots or related workflows; this is separate from the existing Slack alerter that sends notifications to Slack.

• AWS RDS IAM authentication for MySQL stores: The MySQL ZenML store now supports AWS RDS IAM database authentication while preserving the existing password-based behavior PR #5113. ZenML generates fresh IAM tokens for database connections and enforces TLS hostname and certificate verification in IAM mode, giving AWS users a passwordless database authentication option aligned with RDS security best practices.

• Faster pipeline startup with reused build checksums: ZenML now precalculates and reuses Docker build checksums while resolving pipeline builds PR #5154. This avoids repeated build-configuration work during startup, especially for large pipelines with many steps, and also helps when skip_build=True because ZenML no longer repeats unnecessary build preparation checks.

• Safer workload-scoped tokens: Workload-scoped API tokens can no longer be exchanged for persistent, unscoped API tokens PR #5152. This keeps credentials bound to the pipeline run, schedule, or deployment that created them, so workload permissions expire with the workload authorization instead of being escalated into longer-lived generic API access.

• Production/stable package metadata: Future ZenML releases are now marked as Development Status :: 5 - Production/Stable on PyPI instead of beta PR #5186. This aligns the published package metadata with ZenML’s current maturity and makes the package classification clearer for users and automated tooling.

• Failed dynamic pipelines now report failure correctly: Dynamic pipelines running with ExecutionMode.STOP_ON_FAILURE no longer complete successfully when an isolated step fails PR #5182. ZenML now propagates the failed step state correctly so the overall pipeline run status matches the actual execution outcome.

What's Changed

• Add version 0.96.2 to legacy docs by @github-actions[bot] in https://github.com/zenml-io/zenml/pull/5147
• Fail agent examples on provider errors by @strickvl in https://github.com/zenml-io/zenml/pull/5144
• Fix AWS Strands GPT-5 temperature by @strickvl in https://github.com/zenml-io/zenml/pull/5151
• Pin setup-python and extend Dependabot to composite actions by @strickvl in https://github.com/zenml-io/zenml/pull/5118
• Escape $ in generated Docker Compose files by @kounelisagis in https://github.

Runtime and orchestration

• Multi-pod Kubernetes step operator jobs: Command steps can now run across multiple Kubernetes pods with the Kubernetes step operator. Set pod_count in KubernetesStepOperatorSettings to launch the step as an indexed job, making it easier to distribute command-style workloads across pods. PR #5104
• Local Docker sandbox: ZenML now includes a local Docker sandbox, plus a unified settings model for containerized sandboxes. Local sandbox workflows also support file upload and download, making it easier to test containerized ZenML behavior locally before moving to remote infrastructure. PR #5102

Deployment, security, and artifact integrity

• Enrollment keys from Kubernetes Secrets: The Helm chart now supports server.pro.enrollmentKeySecretRef, so ZenML Pro enrollment keys can be injected from an existing Kubernetes Secret instead of being stored inline in Helm release values. The secret reference is applied consistently to the server, migration, and worker containers. PR #5123
• Cloudpickle artifact hash validation: Cloudpickle-materialized artifacts now store a SHA-256 hash when written and validate that hash before loading. This helps detect corrupted or unexpectedly modified artifact files earlier and fail with a clearer integrity signal. PR #5103

CLI and dashboard UX

• Select a stack during login: zenml login now accepts --stack, allowing you to connect to a server and immediately set the active stack in one command. When used together with --project, ZenML applies the project first so the stack is resolved in the intended project context. PR #5125
• Improved secret value display: Secret values in the dashboard no longer expand indefinitely in the UI and are truncated for readability. A direct copy action is now available, making it easier to work with long secret values without disrupting the page layout. PR #1104
• Timeline updates for cancelled runs: The dashboard timeline now shows steps that were not started because a run was cancelled. The timeline filter also supports filtering for Not Started, making cancelled or partially executed runs easier to inspect. PR #1108

• Fewer SQLite lock failures for local stores: Local SQL stores backed by SQLite now wait up to 60 seconds for write locks instead of using Python’s 5-second default. This reduces sqlite3.OperationalError: database is locked failures when concurrent steps, such as mapped dynamic pipeline steps, finish and publish artifacts at the same time. PR #5096
• Artifact deletion on deployed servers: Artifact version deletion now correctly applies project scope when checking whether an artifact is unused. This fixes deletion through the API on remote deployed servers where the default project is disabled. PR #5100
• Complete exception tracebacks: Exception reporting now includes the full traceback lineage instead of only the final exception traceback. This makes chained failures easier to debug because the original cause is preserved alongside the final error. PR #5098
• Smarter Kubernetes dynamic pipeline retries: Kubernetes orchestrator pods are no longer retried when the pipeline run is already in a finished state that cannot be retried. This avoids unnecessary pod restarts that would immediately exit with Run is already finished. PR #5107
• **Co