0.6

Every published okengine release, newest first.

all series · 0.7 · 0.5

v0.6.1

✨ Added

  • Channel drivers on [sently](https://sently.omqkhafi.dev) 1.0 — sndr (email), taqnyat / msegat / unifonic (SMS); wa-cloud, webpush, and fcm wrap sently transports. Boot resolves resend / sndr / taqnyat / msegat / unifonic from env. SMS fallback chains use sently FallbackTransport / RetryTransport. Re-exports toChannelSendResult, parseSndrWebhook, parseUnifonicWebhook, and toDeliveryEvent from okengine/channel.
  • scripts/webpush-interop.ts — manual Channel → Web Push check against a real browser subscription (reuse sently’s subscribe page; prints push-service HTTP status / headers / body). Default VAPID subject uses @example.com (Apple rejects @oke.local / @localhost with 403 BadJwtToken).
  • Root .env.example (and create-oke template) covering Channel, Store, Vault, AI, and Console keys — copy to .env.local.

♻️ Changed

  • Dependency sently bumped to 1.0.0.
  • webpush driver sends through sently createPushSender wrapping WebPushTransport (same Channel surface; push sender pipeline for hooks).
  • Root README.md redesigned with [shieldcn](https://shieldcn.dev/) — adaptive glow header, shadcn-styled npm / JSR / license / Bun / stars / CI badges, and scaffold commands aligned to create-oke’s standard template (dropped the stale --template hello).
  • Local CI test suite is leaner by default: Docker live suites need OKE_TEST_DOCKER=1, Ollama live needs OKE_TEST_OLLAMA_URL or OKE_TEST_OLLAMA_DOCKER=1 (no auto-probe on :11434), create-oke oke dev boot needs CREATE_OKE_DEV_INTEGRATION=1, and npm pack / JSR dry-run run under bun run gate via PUBLISH_GATE=1.
  • Site homepage motion — proof strip and budgets tree read budgets.json, Flow simulator for the one-law band, settle reveals, spotlights, and reduced-motion gates across the landing bands.
  • Landing polish — hero fact chips share the law-chip shell (border / card), with a TypeScript brand mark in #3178C6 and mono icons for Backend / version; Manifest wiring diagram with packet flow, and icons on proof / budget / surface chips.
  • Hero element lattice sits lower and larger, with the lit-cell accent as a 1px rule on the cell’s bottom edge and a roomier docs footer; the collapse diagram’s idle feed lights a small multi-element operation each beat instead of a single concern.
  • Landing budgets panel: four AGENTS caps as headline metrics (measured of limit, % used), then okengine and okengine/plugins as baselines with leaf entrypoints / per-plugin rows collapsed by default (Base UI Collapsible); plugins grouped auth · security · ops · perf — rounded timings (36 ms, < 1 ms) instead of false-precision decimals; Features panel keeps the shared card chrome.
  • CI budgets measure each official okengine/plugins named module (regression group plugins in budgets.json / BUDGETS.md), not only the plugins barrel.
  • One-law band: full-width claim, then two equal peer panels — a vertical Flow path (matching CodePanel chrome) beside the starter snippet, instead of stacking the diagram under the heading.
  • Element feature cards drop the decorative dot lattice and drifting placeholder icons — content and preview chips only.
  • Local bun run ci runs Budgets first so the cold-start probe is not inflated by Format/Lint/Typecheck/Tests heating the machine.

🐛 Fixed

  • bun run budgets writes oxfmt-aligned BUDGETS.md tables, so Format no longer fails after every budgets refresh.

v0.6.0

✨ Added

  • AI tool-calling through the same path as fx.call — fx.ask(prompt, input, { tools: [flowRef, …] }) and fx.run(agent) share one model-driven tool loop; each tool is a Flow (capability, Manifest calls, Runs portal). No silent extra authority for the model.
  • Real fx.stream — Ollama NDJSON and openai-compatible SSE; cancellation reuses the ambient AbortSignal from fx.all / fx.race (one channel).
  • AI_RATE_PRESETS / aiRateGate("ask" | "agent" | "embed") — Gate rate presets for expensive AI edges (defaults 20 / 10 / 60 per minute, keyBy: "user"). Cost caps stay on prompt/agent budget decls.
  • Optional headers on AI open options (e.g. OpenRouter HTTP-Referer / X-Title) — still one openai-compatible driver, not per-vendor ids.
  • Live Ollama tool-calling integration test (gated ask → real tool → ledger portal; skip-visible without a reachable Ollama).
  • defineLocale / fx.t — register EN/AR (or any) message catalogs; fx.t resolves locale → i18n.default → key. Request locale comes from Accept-Language (matched against oke.config i18n.locales). fx.locale exposes the active tag; channel fx.send defaults to it. Starter ships src/locales/en.ts + ar.ts.
  • ICU MessageFormat for fx.t (FormatJS intl-messageformat) — interpolation, cardinal/ordinal plurals, select / selectordinal, and rich-text tags (<tag>…</tag> with function values).
  • Type-safe message keys — defineMessages, MessagesFor<typeof en>, and declare module "okengine" { interface Register { messages: typeof en } } so fx.t autocompletes keys and rejects typos at compile time.
  • Docs — [i18n](/docs/reference/i18n) reference page (ICU catalogs, typed keys, locale matching, built-in failure messages vs Channel {{field}}).
  • magicLink / emailOtp Channel catalogs include Arabic bodies alongside English (locales: ["en", "ar"]).
  • Built-in EN/AR ICU catalogs for typed failures (errors.*) and OKE codes (oke.*.cause / oke.*.fix). fail / fx.fail attach a localized error.message from the active request locale; thrown OkeError localizes cause and fix. Override via defineLocale; explicit { message } wins.
  • ollama AI driver — thin native POST /api/chat client for any pulled model (OKE_AI_MODEL / ai.model); documented local-dev default qwen3.5:9b (balanced starting point, fully overridable; Apple Silicon can use qwen3.5:9b-mlx). Base URL http://127.0.0.1:11434 via OKE_AI_URL / OLLAMA_HOST. Unreachable servers throw OllamaUnavailableError — never a silent mock fallback.
  • Ollama Docker recipe (images.ai: "ollama/ollama") — pulls the configured model, health-checks, writes OKE_AI_URL. Boot resolves drivers.ai through a shared aiDriverFor switch (same class as store.index).
  • Live Ollama integration test (real completion; skips visibly without a reachable Ollama / explicit Docker opt-in).
  • magicLink / emailOtp deliver challenges via Channel (fx.send + auth-magic-link / auth-email-otp templates and English catalog bodies). exposeDevToken / exposeDevOtp stay for local DX without SMTP. phoneNumber SMS delivery remains deferred.
  • Plugin API .channelCatalog(catalog) — merge template bodies into boot Channel catalog (parallel to .channelTemplate).
  • Mailpit integration tests prove real SMTP capture of magic-link token/link and email OTP (skip with a visible warn when Docker is unavailable).
  • username({ usernamePolicy, passwordPolicy }) — tunable min/max length, charset, letter/number/symbol requirements, and reserved names on sign-up. Policy failures return AuthFailed with reason: "username_policy" / "password_policy" and a reasons array (taken usernames stay enumeration-safe as invalid_credentials).
  • username default reserved blocklist (DEFAULT_RESERVED_USERNAMES) — admin / root / api / auth / anonymous and peers. reserved: [] opts out; a custom list replaces (does not merge).
  • usernamePolicy.extraAllowedChars — append characters to the default (or allowedChars) charset without replacing the whole set.
  • usernamePolicy.extraReserved — append names to the default (or reserved) blocklist without replacing it.
  • usernamePolicy shape rules — mustStartWithLetter, forbidEdgeSymbols (default on), forbidConsecutiveSymbols (default on).

♻️ Changed

  • Ollama documented local-dev default model is qwen3.5:9b (was qwen3:8b; same as qwen3.5:latest on the [Ollama library](https://ollama.com/library/qwen3.5)). On Apple Silicon, set OKE_AI_MODEL=qwen3.5:9b-mlx for the MLX build. Docker recipe / compose control use the non-MLX tag.
  • openai-compatible AI driver is the shared chat/completions client for OpenAI, Groq, Together, OpenRouter, vLLM, LM Studio, and Ollama /v1 — configure via baseUrl + apiKey + model. apiKey is required only for the default OpenAI cloud base; custom bases may omit it. Native ollama remains the first-class local path.
  • Native ollama driver adds tools + streaming (stream: true NDJSON) on /api/chat (still fail-loud; never silent mock fallback).
  • Effect inference: fx.ask(…, { tools: […] }) with static Flow refs adds them to Manifest effects.calls.
  • Rebased export/driver budgets after AI depth (tools, stream, rate presets) and related Unreleased work; refreshed budgets.json / BUDGETS.md.
  • Rebased the Store-only oke() gzip graph baseline (41.4 kB → 50 kB) after shared fx growth (stream + i18n).
  • Rebased the plugins export budget after Channel delivery wiring (13.84 kB → 14.77 kB), and refreshed the published budgets snapshot.
  • Rebased the plugins export budget after username policy wiring (14.77 kB → 15.54 kB), and refreshed the published budgets snapshot.
  • Rebased the plugins export budget after username Gate-auth parity (15.54 kB → 15.97 kB), and refreshed the published budgets snapshot.
  • username() sign-up inherits gate.auth.passwordPolicy, password hash knobs, and breachCheck (same as email/password Gate auth). Explicit plugin opts still override for that method only.

🔒 Security

  • Redacted values in fx.ask / stream prompts are masked with [redacted] before they reach any provider (same class as fx.log).

🐛 Fixed

  • smtp channel driver passes sently's BunAdapter so real SMTP (Mailpit / providers) works under Bun — previously failed with "No socket adapter configured".
  • Mailpit auth-delivery tests use real test.skip when Docker is unavailable (never an empty pass).