Skip to content

riddlg Release Notes

The current release is 0.6.0 (20 July 2026). Install or upgrade with brew upgrade riddlg, or see Installation for direct downloads.

Version Date Headline
0.6.0 2026-07-20 Eight new model-transform generators
0.5.0 2026-07-17 Rich docs in four formats; streaming chat
0.4.0 2026-07-15 Bring-your-own-key AI providers; MCP over HTTP
0.3.1 2026-07-13 Linux GPU builds (CUDA and Vulkan)
0.3.0 2026-07-10 Reliable code fill; public distribution
0.2.0 2026-07-06 First release as riddlg

0.6.0

20 July 2026

A RIDDL model describes a system precisely enough to derive far more than documentation from it. This release acts on that: eight new generators turn your model into schemas, catalogs, and structured documentation.

New generators

Generator Command Tier
AsyncAPI 3.0 — event-driven API specs with in/out/stream channels and request/reply gen api -f asyncapi Free
JSON Schema 2020-12 — standalone validation schemas for contract enforcement gen api -f json-schema Free
SQL DDL — normalized DDL from entities and repositories, in five dialects gen sql Free
DBML — dialect-free logical schemas for dbdiagram.io gen dbml Free
Backstage — software catalogs with ownership resolution gen backstage Free
EventCatalog — event-driven doc sites with per-message schemas gen catalog Free
Confluence — storage-format pages plus a REST importer gen confluence Pro
DocBook + DITA — structured XML documentation gen docs -f docbook / -f dita Pro

Every one of them is also exposed over riddlg serve at POST /generate/…, so Synapify and your own tooling get them too. See Generators for what each produces.

Model-carried generation options

riddl-lib 1.31.0 registers the options these generators read, so a model can declare its own generation settings and still validate without warnings: sql_dialect, sql_table, backstage_owner, backstage_lifecycle, backstage_type, confluence_space, confluence_parent, event_catalog_version, and protocol.

Also in this release

  • Release automation now publishes every platform that builds successfully, so one failing build leg no longer blocks the whole release.

0.5.0

17 July 2026

Breaking change: environment variables were renamed

Every OSSUM_GEN_* environment variable is now RIDDLG_* — for example OSSUM_GEN_MODEL_URL becomes RIDDLG_MODEL_URL, and OSSUM_GEN_MAX_TOKENS becomes RIDDLG_MAX_TOKENS. The old names are no longer read. See Configuration for the full list. Config file keys and CLI flags are unaffected.

Documentation, rebuilt on a shared model

Every documentation format is now rendered from one shared document model, so they are genuinely equivalent rather than each format being its own partial implementation. All formats now include:

  • Pages and in-page sections mirroring the model hierarchy
  • Every definition's briefly and described by text
  • Per-definition and overview diagrams, and attached images, as numbered figures
  • A home page with a table of contents and a table of figures

Four documentation formats

gen docs -f gained hugo-book and hugo-geekdoc alongside asciidoc and mkdocs, with hugo as an alias for hugo-geekdoc. Hugo output had been on the roadmap; this is where it landed.

AsciiDoc output builds a website and a PDF

The asciidoc output now includes a pom.xml, a Makefile, and a README.adoc, so the generated directory is a working build:

riddlg gen docs model.riddl -o docs/
cd docs/ && make site   # or: make pdf

Streaming AI responses

POST /ai/messages accepts "stream": true and responds with an Anthropic-shaped text/event-stream. Anthropic profiles forward the upstream stream verbatim, including tool_use blocks; other providers get an equivalent synthesized event sequence, so clients need only one code path.

Fixes

  • riddlg serve no longer holds its listening port after a model download. The download subprocess was inheriting the server's socket.

0.4.0

15 July 2026

Bring your own key

riddlg is no longer limited to the local model. With a Pro subscription you can point it at Anthropic, Google Gemini, OpenAI, the OpenAI Responses API, or any OpenAI-compatible service — Ollama, vLLM, LM Studio, Groq, OpenRouter, Azure, and others — using your own API key. That path needs no GPU and no 23 GB model download.

Keys can come from an environment variable, stdin, or the OS keychain, and riddlg ai manages named profiles:

riddlg ai use anthropic --api-key-stdin
riddlg ai add ollama --type openai --base-url http://localhost:11434/v1
riddlg ai test ollama

See AI Providers.

Pro comes from your Ossum subscription

Breaking change: license files were removed

Earlier versions used an offline license token (OSSUM_GEN_LICENSE, or a file at ~/.ossum-gen/license). That mechanism is gone. Your tier now comes from your Ossum account — the same subscription as Synapify — via riddlg login. Existing license tokens and files are ignored and can be deleted.

riddlg login    # device flow: prints a URL and a code
riddlg whoami   # account and tier

The session keeps working offline for up to 7 days between refreshes.

MCP over HTTP, and more MCP tools

riddlg serve gained POST /mcp — the full MCP tool set over Streamable HTTP, on the same port as everything else, identical to what riddlg mcp serves over stdio. Eleven derivation tools joined the two validation tools: check-completeness, suggest-next, map-domain-to-riddl, validate-partial, expand-pattern, get-template, generate-test-cases, check-simulability, explain-error, list-patterns, and list-templates. See MCP Tools.

Also in this release

  • Model auto-download. The first command that needs the local model fetches it, with progress reported at GET /model/status so a UI can draw a banner.
  • POST /ai/messages — one chat turn in the Anthropic Messages shape, so a client can hold a conversation through riddlg without ever handling API keys.

0.3.1

13 July 2026

  • Linux GPU builds. cuda and vulkan variants join the CPU build. The CUDA variant vendors the CUDA runtime, so you need only the NVIDIA driver; the Vulkan variant runs on AMD, Intel, and NVIDIA alike. See Installation.
  • llama.cpp is now built from source on every release platform, which is what makes the GPU backends actually present in the shipped binary.
  • A latest.json manifest is published alongside each release so download pages can resolve the current version.

0.3.0

10 July 2026

  • Generator correctness. Five bugs that produced invalid RIDDL were fixed; the fidelity baseline reached 79 of 79 models generating valid output.
  • gen code --fill produces compiling output reliably. The AI-filled method bodies are compile-verified with retries.
  • Public distribution. Release artifacts moved to a public bucket, which is what makes brew install ossuminc/tap/riddlg work from a private source repository.
  • Upgraded to riddl-lib 1.29.0.
  • The simulator module was removed — it is outside riddlg's mission. RIDDL simulation is on the roadmap elsewhere.

0.2.0

6 July 2026

The first release under the name riddlg, following riddlc's naming convention.

  • Compositional generation. Large systems are generated by layered decomposition rather than in one shot, with shared types hoisted to domain scope. --multi-file emits an include tree with a file per context, and briefs longer than 240 characters select this strategy automatically.
  • Story layer. Generated models include the epic / user / use-case layer.
  • HOCON configuration at ~/.riddlg/config.conf, with riddlg config to print the effective settings. See Configuration.
  • GPU pre-flight. Model-loading commands stop with a clear message rather than crawling on CPU; --allow-cpu overrides.
  • version and info commands, the latter listing detected compute devices.
  • Homebrew, .deb, and .rpm packaging.

See Also