Unreleased

0.1.3 - 2026-09-09

Fixed

• Build and run on native Win32 while retaining protected credential handling: require secret 0.1.3 for direct Win32 secret I/O, use a native monotonic clock, use selectable socket pairs for internal wakeups, and use the platform null device for exec credentials.

0.1.2 - 2026-09-07

Fixed

• Require ppxlib 0.36.0 or newer so the kube.ppx deriver compiles when opam selects lower dependency bounds, and exercise lower-bound installs in project CI to prevent regressions.

0.1.1 - 2026-09-06

Fixed

• Declare the minimum compatible tls and yojson versions so opam lower-bound builds select dependencies that provide the APIs used by kube.
• Skip loopback-server integration tests under opam-repository CI, whose build sandbox intentionally disallows binding local sockets. The tests continue to run normally in development and project CI.

0.1.0 - 2026-09-06

Added

• Http.Sensitive and Client.Sensitive protected request paths backed by Secret.t, including scoped token-file authorization, bounded protected response bodies, direct protected JSON/base64 Secret creation and atomic patching, and protected ServiceAccount TokenRequest results. Existing string APIs remain source-compatible and retain their original memory semantics.
• A batteries-included Kube.Operator runner that centralizes standard CLI flags, configuration, signal cancellation, client ownership, diagnostics, health, metrics, and optional Lease leadership for examples and scaffolds.
• Kube.Reconcile helpers for typed Server-Side Apply and safe controller-owned children, plus cancellation-aware Kube.Retry.on_conflict optimistic retry.
• Kubernetes-standard custom-resource conditions with transition-preserving updates, strict codecs, and structural schemas.
• Compilation-checked common operator patterns, application instrumentation, an ecosystem parity audit, and a candid guide to choosing OCaml over Go or Rust.
• Native OCaml 5 Kubernetes HTTP and TLS transport.
• Authenticated WebSocket upgrades; versioned exec and attach channels with resize, half-close, and structured exit status; and multiplexed Pod port-forward streams with a local TCP forwarder.
• Kubeconfig, in-cluster, token-file, basic, client-certificate, and exec-plugin authentication.
• Typed and dynamic CRUD, patch, Server-Side Apply, pagination, discovery, status, arbitrary subresources, and watch support.
• Safe collection deletion with explicit cross-namespace opt-in; complete raw subresource POST, PUT, PATCH, DELETE, and streaming GET operations; typed autoscaling/v1 Scale access; and bounded or streaming Pod logs across typed, dynamic, and cache-backed clients.
• Resource-version-aware reflectors, local stores, deduplicating work queues, delayed requeue, exponential retry, finalizer helpers, and concurrent controllers.
• Controller cache-sync readiness checks that cover primary and watched cache synchronization, initial queue seeding, worker startup, and shutdown.
• A typed custom-resource operator and repeatable kind integration test.
• Compatibility lanes for active Kubernetes minor releases.
• Checksum-pinned generated built-in API packages for Kubernetes 1.34 through 1.37, with manifest-driven regeneration and offline drift checks.
• Kubeconfig merging and client-go-compatible exec-plugin validation, relative command resolution, and token-file precedence.
• Thread-safe persistent HTTP/1.1 connection pooling with stale-connection detection and explicit client shutdown.
• A configurable monotonic connection-establishment deadline covering isolated DNS lookup, nonblocking TCP connect, and interruptible TLS handshake.
• Independent monotonic request-write and response-header deadlines while leaving streaming response bodies under explicit cancellation control.
• Kubeconfig and programmatic user impersonation, including UID, repeated group, and percent-escaped extra headers with explicit per-request override behavior.
• Explicit kubeconfig HTTP forward/CONNECT proxies and authenticated SOCKS5 tunnels with cancellation and connection-deadline coverage.
• Structured multi-controller supervision, shareable typed reflector caches, cache-sync barriers, predicates, arbitrary secondary watches, and typed owner-reference watches.
• Kubernetes Lease leader election with local observation-time expiry, optimistic conflict handling, per-attempt deadlines, jittered acquisition, renewal supervision, and release only after protected work is joined.
• Native Prometheus counters, gauges, and histograms; controller and leadership instrumentation; named liveness/readiness checks; and a supervised HTTP diagnostics component serving health, readiness, and metrics endpoints.
• Strict admission.k8s.io/v1 codecs, typed resource handlers, validating and mutating outcomes, warnings, audit annotations, and complete RFC 6902 patch responses.
• Checked apiextensions.k8s.io/v1 CRD conversion reviews with ordered mapping, desired-version validation, and name/namespace/UID identity preservation.
• A supervised TLS 1.2/1.3 webhook server with optional mutual TLS, bounded connections/headers/bodies, monotonic request deadlines, cancellation, readiness integration, per-path Prometheus metrics, and joined shutdown.
• A deterministic Kubernetes Swagger 2.0 compiler that computes transitive schema closures and emits strongly typed OCaml records, constructors, JSON codecs, common metadata helpers, and Core.Resource descriptors. Stable resource selection, plurals, and scopes are derived from upstream LIST/WATCH operations rather than maintained by hand.
• The kube.api.v1_36 generated sublibrary, pinned to Kubernetes v1.36.2 and covering all 60 stable LIST/WATCH resources and 473 dependent definitions, with a machine-readable all_resources registry.
• Reproducible pruned OpenAPI input, generator golden tests, generated-codec round-trip tests, and the @codegen-check drift gate.
• Reconciliation requests carrying the controller lifetime token so in-flight client operations can be interrupted during structured shutdown.
• A default cache-synchronization deadline, optional per-reconciliation deadlines with linked child cancellation, prompt cancellation-aware sleeps, and timeout-labelled reconciliation metrics.
• Atomic, thread-safe secondary indexes on reflector stores and server-side field selectors for primary, secondary, and owner-reference watches.
• An explicit typed cache-backed client with synchronized GET, LIST, and index reads, classified cache misses, deliberate fresh-read methods, live mutation delegation, and automatic availability on every controller request.
• A thread-safe lazy discovery mapper for exact and preferred GVK/GVR resolution, plural/singular/short-name aliases, ambiguity detection, discovered scope, verbs and subresources, shared in-flight requests, generation-safe invalidation, and opt-in complete refresh.
• Configurable per-controller error policies with finite exponential backoff, retry budgets, and drop semantics.
• A cancellation-aware native token-bucket API-client rate limiter with shared, custom, and unlimited configurations.
• A finalizer lifecycle state machine that separates apply from cleanup and removes a finalizer only after successful cleanup.
• A deterministic scripted loopback API-server harness covering coalesced and byte-fragmented watch events, truncated chunk framing, HTTP and in-band 410 recovery, reconnect continuity, and atomic compaction relists.
• The kube.crd sublibrary with a typed structural OpenAPI schema AST, validation, CRD names/versions/status/scale/printer-column support, deterministic JSON and YAML manifests, and a typed custom-resource functor.
• Typed custom-resource construction, strict spec/status decoding, status merge patch helpers, and a generated-manifest drift check used by the Greeting operator.
• The kube.ppx [@@deriving kube] rewriter for records, aliases, options, lists, arrays, string maps, structural tuples, nullary and payload variants, external derived types, and explicitly schema-bounded recursive types.
• Deriver attributes for stable field and constructor wire names, descriptions, and complete schema overrides, plus generated .mli declarations and compile-time duplicate-name diagnostics.
• Kubernetes-aware API error classifiers, structured retry-delay extraction from Status details and HTTP Retry-After, and server-directed reflector and leader-election backoff.
• An events.k8s.io/v1 recorder, canonical ObjectReference serialization, UID-checked owner-reference constructors, proof-operator Event reporting, and a real-cluster Event assertion.
• A pluggable thread-safe structured logger with deterministic JSON output, immutable scoped fields, dynamic levels, explicit redaction, sink-failure isolation, sanitized HTTP request diagnostics, and lifecycle/error events across managers, reflectors, controllers, leader election, diagnostics, and webhook serving.
• A public injectable Client.Transport boundary and kube.test package that retain the production authentication, retry, request, decoding, watch, cache, and controller paths while providing deterministic responses, programmable handlers, concurrent request capture, streaming, cancellation, and strict script-completion checks.
• Explicit multi-namespace reflector and controller scopes with one independent LIST/WATCH resource version per namespace, aggregate readiness, namespace-local compaction relists, atomic scoped store replacement, and matching primary, secondary, and owner-watch plumbing.
• An installed ocaml-kube scaffold CLI that imports a selected served version from a v1 CRD and atomically generates a standalone typed operator project, controller/finalizer skeleton, package metadata, RBAC, deployment, sample, and container artifacts without overwriting an existing path.
• An ocaml-kube init type-first workflow whose starter OCaml models derive JSON codecs and structural schemas, generate the checked-in CRD manifest, expose a drift-check alias, and compile as an independent operator project.
• Schema-to-OCaml mapping for records, optional fields, primitive formats, arrays, typed maps, empty objects, Int-or-String, and string enums, with exact source schemas and path-specific raw-JSON diagnostics for unsupported structural shapes. --deny-raw turns those diagnostics into a generation failure.
• Generated operator entry points and deployments with optional Lease leader election, native metrics and probe endpoints, two-replica defaults, graceful supervision, and restricted non-root container security settings.
• A scaffold integration proof that generates and externally compiles a fresh project, runs it against kind, observes typed reconciliation and finalizer cleanup, and verifies graceful shutdown.
• An offline, pinned scaffold acceptance corpus covering Gateway API, KEDA, and Prometheus Operator CRDs, plus codec-only PPX derivation so imported schemas compile without recursively duplicating unused generated schemas.

Security

• Bound buffered HTTP response bodies and individual watch frames.
• Reject newline injection in request targets and headers.
• Refresh cached exec credentials once after an HTTP 401 response.
• Validate HTTP header grammar, reserve framing headers to the transport, and reject unencoded request-target whitespace.

Fixed

• Reliable SIGINT and SIGTERM cancellation while the operator supervisor is blocked waiting for components, including generated operators on Linux.
• Leader-election standby readiness now reports healthy while participating in election, preventing multi-replica rolling deployments from deadlocking while preserving the active-leader metric and single-reconciler execution.
• Recoverable per-connection port-forward stream failures, with bounded and diagnostic integration probes instead of indefinite CI hangs, and portable regression coverage for both clean EOF and connection-reset closure.
• Use a process-local monotonic clock for cancellation deadlines, controller latency, discovery cooldowns, queue scheduling, rate limiting, reflector recovery, exec-helper timeouts, diagnostics, and Lease expiry observation, so civil-clock adjustments cannot distort elapsed-time decisions.
• Preserve previous values for overlapping resources during an atomic relist, allowing secondary-watch and owner mappings to reconcile both the old and new relationships after resource-version compaction.
• Reconcile both the previous and current primary keys when a secondary watch changes relationships, including dependent-resource ownership transfers.
• Keep polling and draining exec credential helpers until both their process and output pipes complete, instead of returning after an initial nonblocking poll.
• Omit unknown resource versions from finalizer patches instead of sending JSON null.
• Decode JSON integer literals in common object metadata so generated int64 generations remain visible to reconciliation logic.
• Retry transient initial LIST failures with jittered exponential backoff, protect against immediate clean-EOF watch reconnect loops, and propagate non-410 in-band watch Status failures into the reflector retry path.
• Apply watch size limits to individual JSON event frames rather than arbitrary transport chunks, so coalesced valid events are not rejected.
• Resume a reflector reconnect from the newest resource version already delivered to its cache after a transport failure.
• Classify socket EOF and framing failures as cancellation when an in-flight request's cancellation token shut down the connection.
• Fill missing TypeMeta from the requested API descriptor when decoding dynamic LIST/WATCH objects, matching API servers that omit apiVersion and kind inside collection items.
• Make the kind integration test recover stale finalizers left by interrupted runs and use group-qualified resource names when several CRDs share a kind.