All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

Unreleased

2.0.0 - 2026-11-16

Providers and models become data, and Omni gains a second kind of request: decisions. Most apps need a config change and little else. See Upgrading to v2.

Added

  • Decisions — Omni.decide/3 asks decision models typed questions and returns probabilities, over two new dialects (TypeSafe System One and OpenAI Decisions).
  • Catalog providers — any models.dev provider loads by id, with no module to write. Custom providers need only config.
  • New built-in providers — Mistral (:mistral), Together AI (:togetherai), xAI (:xai) and TypeSafe (:typesafe).
  • String model references — "anthropic:claude-sonnet-5-5" works wherever a {provider_id, model_id} tuple does.
  • Compat flags — a provider's quirks are settings on its endpoint, changeable in config.
  • Model loading controls — filter: and transform: under config :omni, :models, and per-provider models: to add or correct models.
  • Content constructors — Omni.attachment/1,2 and Omni.tool_result/1, and a string in a content list becomes a text block.
  • mix omni.verify — runs live requests against one model to check it works with your config.
  • Updated model catalog — refreshed across all providers.

Changed

  • Provider config is keyed by id (breaking) — it moves under config :omni, :providers, and the providers: list takes ids, not modules.
  • A URL per dialect (breaking) — a provider's base_url is replaced by host:, or a whole URL on the endpoint of each dialect it serves.
  • %Omni.Model{} (breaking) — a model embeds its provider as a struct, names a dialect per capability, and reasoning becomes thinking. Omni.put_model/2 is now put_model/1.
  • Thinking follows each model (breaking) — effort levels are translated from catalog data, and a token budget is an integer (thinking: 20_000).
  • Provider modules, dialects and sources (breaking) — all three behaviours changed shape.
  • GoogleGemini dialect renamed (breaking) — it is now Omni.Dialects.GoogleGenerateContent.
  • Thinking goes back only to the provider and dialect that wrote it — switching model mid-conversation no longer fails.
  • Cached tokens are counted once — usage.input_tokens excludes them on every dialect, so totals and costs are no longer too high.

Fixed

  • Truncated responses — a response cut off by the token limit reports :length, including inside a tool use.
  • Reasoning across turns — OpenAI Responses and more Chat Completions providers have earlier reasoning sent back.
  • Ollama Cloud streaming — every request ended in :incomplete_stream.
  • OpenRouter prompt caching — :cache now works for Claude models.
  • Chat Completions parsing — several tool uses in one chunk, and content sent as a list, are read in full.
  • Provider fixes — Z.ai honours :max_tokens, and Groq no longer puts <think> tags in the answer.

1.6.5 - 2026-10-09

Added

  • Updated model catalog — refreshed across all providers. Include latest Claude and GPT 6.1 models.

1.6.4 - 2026-09-02

Added

  • Updated model catalog — refreshed across all providers. Include GLM 5.3 and GLM 5.3 Flash.

Fixed

  • Packed SSE chunks in OpenAI Completions dialect — providers that bundle multiple payload types in a single chunk (e.g. tool_call arguments alongside finish_reason) no longer silently drop fields after the first match. All fields are now extracted in a single pass.

1.6.3 - 2026-08-26

Added

  • Updated model catalog — refreshed across all providers.

1.6.2 - 2026-07-17

Added

  • Updated model catalog — refreshed across all providers. Includes Kimi K3.

Fixed

  • Sequential block closure in StreamingResponse — block *_end events (e.g. thinking_end, text_end) are now emitted inline when the next block starts, rather than bunched at stream finalization. In multi-block responses (thinking+text, text+tool_use), a block's end event now always precedes the next block's start event.

1.6.1 - 2026-07-10

Added

  • Updated model catalog — refreshed across all providers. Captures GPT-5.6 family (Sol, Terra, Luna).

1.6.0 - 2026-07-06

Added

  • Pluggable model sources — model catalog data now loads through the new Omni.Source behaviour, configurable globally or per provider.
  • Omni.Sources.ModelsDev — the default source: a bundled models.dev snapshot, with an optional live mode (live: true) that fetches fresh catalog data at boot and degrades gracefully to the snapshot when offline.
  • Omni.Sources.LLMDB — alternative source backed by the optional llm_db package.
  • Omni.Providers.OllamaCloud — new provider for Ollama's hosted cloud service (:ollama_cloud, reads OLLAMA_API_KEY).

Changed

  • Providers declare a canonical id (breaking) — use Omni.Provider, id: :mistral, named after the provider's models.dev catalog key (snake_cased). The id drives registration, Omni.get_model/2, and catalog lookup, and models/0 now defaults to loading from the configured source — a catalog-backed custom provider needs only id: and config/0.
  • Provider config moved to config :omni, :models (breaking) — registration goes under providers: (a list of provider modules; :builtins, the default, names the built-ins and may appear in the list), the model source under source:. The legacy :providers key raises at boot with migration instructions.
  • :moonshot is now :moonshotai (breaking), matching the catalog id.
  • Ollama split into local and cloud providers (breaking) — Omni.Providers.Ollama (:ollama) now targets local instances only, with models listed in its models: config. Cloud users switch to Omni.Providers.OllamaCloud (:ollama_cloud).
  • Dialect resolution is now data-first — a model's dialect from catalog data wins over the provider module's declared dialect, which is now a fallback. No behavior change for built-in providers.
  • Omni.Provider.load_models/2 (breaking) — takes options instead of a JSON file path. The curated priv/models/*.json files are replaced by the verbatim snapshot, captured with the new mix omni.snapshot task.

Removed

  • Legacy mix tasks — mix models.update and mix models.update_llmdb.

1.5.4 - 2026-06-25

Added

  • Updated model catalog — refreshed across all providers.

1.5.3 - 2026-06-16

Added

  • Updated model catalog — refreshed across all providers.

1.5.2 - 2026-06-02

Added

  • Updated model catalog — refreshed across all providers. Notable additions: Claude Opus 4.8, Minimax M3.

1.5.1 - 2026-05-22

Added

  • NEAR AI Cloud provider — built-in provider for NEAR AI Cloud's OpenAI-compatible API, including TEE-backed inference models.
  • Updated model catalog — refreshed across all providers.

1.5.0 - 2026-05-18

Added

  • Updated model catalog — refreshed across all providers.

Changed

  • All built-in providers loaded by default — all 11 built-in providers are now available out of the box without configuration. The config :omni, :providers option can still be used to restrict which providers load at startup.

Fixed

  • Prompt cache usage tokens — cache_read_tokens and cache_write_tokens on %Usage{} were always zero for non-Anthropic providers. All dialects now correctly report prompt cache hits from their respective APIs.

1.4.1 - 2026-05-11

Added

  • Venice AI provider — opt-in built-in provider for Venice AI's hosted models.
  • Updated model catalog — refreshed across all providers.

1.4.0 - 2026-05-08

Added

  • Omni.Schema.Adapter behaviour — pluggable schema validators. Pass a {module, state} tuple anywhere a JSON Schema map is accepted (:output option, Tool input_schema, tool schema/0 callback) to delegate wire encoding and validation to a custom adapter. Enables JSV, Ecto, or other validators with zero new dependencies in Omni itself.
  • Omni.Schema.to_schema/1 — public dispatcher that returns the wire-form JSON Schema map for a raw schema or adapter tuple.
  • Omni.Tool schema/1 callback — state-aware variant of schema/0 for tools whose input schema depends on init state. Mirrors the existing description/0,1 pattern; the default schema/1 delegates to schema/0, so existing tools are unaffected.
  • Updated model catalog — refreshed across all providers.

Changed

  • Omni.Tool description/0 and schema/0 are no longer required when implementing the 1-arity variant — the macro now generates raising defaults (matching call/1,2), so a tool can implement just description/1 or schema/1 without compiler warnings.
  • Omni.Schema is now the default Omni.Schema.Adapter — to_schema/1 and validate/2 accept either a raw map or an adapter tuple and dispatch accordingly.
  • Omni.Schema.validate/2 return shape — now returns {:ok, term()} | {:error, String.t()} consistently. Errors come back as a pre-formatted human-readable string instead of a list of Peri error structs.

  • Built-in validator — replaced internal Peri schema converter with Peri.from_json_schema/1 (peri bumped to ~> 0.8).

Fixed

  • JSON Schema constraints now enforced — multipleOf, minItems, maxItems, uniqueItems, const, null, oneOf, allOf, union types, and string format are validated locally where previously they were silently skipped.
  • number type accepts integers — schemas with {"type": "number"} now accept integer inputs as required by the JSON Schema spec (previously rejected).

1.3.2 - 2026-05-01

Added

  • Updated model catalog — refreshed across all providers, including GPT-5.5 Pro and Deepseek v4 Pro.

1.3.1 - 2026-04-27

Added

  • Alibaba provider — opt-in built-in provider for Alibaba Cloud's hosted models.
  • Updated model catalog — refreshed across all providers, including GPT-5.5 and Deepseek v4 (OpenRouter).

1.3.0 - 2026-04-21

Added

  • Groq provider — opt-in built-in provider for Groq's hosted models.
  • Moonshot AI provider — opt-in built-in provider for Moonshot's Kimi models.
  • Z.ai provider — opt-in built-in provider for Z.ai.
  • Omni.Codec — lossless encode/decode of messages, content blocks, and usage to JSON-safe maps for persistence.
  • :xhigh thinking level — new level between :high and :max.
  • Claude Opus 4.7 support — full support including adaptive thinking.
  • Updated model catalog — refreshed across all providers, including Claude Opus 4.7 and Kimi K2.6.

Changed

  • Google Gemini API version — now defaults to v1beta unconditionally, removing the previous v1/v1beta option.

Fixed

  • OpenAI Completions tool use parsing — arguments were dropped when name and arguments arrived in a single streaming event.
  • OpenAI structured output — additionalProperties: false now applied on object schemas for strict mode compatibility.
  • OpenAI Responses PDF attachments — added missing filename field on file content blocks.
  • Google Gemini API version — hardcoded v1beta path prefix, as many features require the beta API.

1.2.1 - 2026-04-02

Added

  • Dynamic tool descriptions — Override description/1 to incorporate init/1 state into the tool description at construction time.

Fixed

  • Google Gemini structured output — auto-upgrade to v1beta API when :output is set, fixing 400 errors caused by responseMimeType/responseSchema fields not existing in v1.

1.2.0 - 2026-03-23

Added

  • release_date on %Model{} — optional Date.t() field populated from models.dev release date data. Enables filtering models by release date.
  • New models — GPT-5.4 Mini and GPT-5.4 Nano added to model data.

Removed

  • Agent extracted — Omni.Agent, Omni.Agent.*, and Omni.MessageTree have been extracted into the standalone omni_agent package. This package is now purely the stateless LLM API layer.
  • %Turn{} struct removed - Loop :turn_id and :turn_parent options removed.

Changed

  • %Attachment{} — removed unused description field. Renamed opts to meta — an application-layer map that dialects do not read or send to providers (e.g. for filenames or display labels).
  • %Response{} — turn field replaced by messages (list), usage (%Usage{}), and node_ids (list of tree node IDs, used by omni_agent). :message is now optional. Added :cancelled stop_reason.

1.1.0 - 2026-03-06

Added

  • %Turn{} struct — a conversation turn carrying messages, usage, and tree position (id, parent). Used as tree nodes in MessageTree and on Response for accumulated generation data.
  • MessageTree — tree-structured conversation history with branching, navigation, and Enumerable support.
  • Agent conversation tree — agents manage a %MessageTree{} internally. State restructured around system, tools, tree, meta, private fields. assigns replaced by meta (user data) + private (runtime state).
  • Agent.set_state/2,3 — replace agent configuration fields (model, system, tools, tree, opts, meta). Always replaces, atomic, idle-only. /3 accepts a value or updater function.
  • Agent.navigate/2 — set the active conversation path to any turn in the tree.
  • Agent.usage/1 — cumulative token usage across all turns in the tree.
  • Turn data on agent events — :turn and :done events carry a %Turn{} with tree position, enabling external persistence without built-in storage.
  • tree: start option — hydrate an agent with a pre-built %MessageTree{}.
  • Model.to_ref/1 — convert a resolved %Model{} back to its {provider_id, model_id} lookup reference.

Changed

  • %Response{} — messages and usage fields replaced by turn field containing a %Turn{}. Access via response.turn.messages and response.turn.usage.

Fixed

  • OpenAI Completions dialect — tool use name lost when provider sends id on every SSE chunk (affected Kimi models via OpenCode/Fireworks).

1.0.0 - 2026-03-06

Complete rewrite of Omni as a production-ready, multi-provider LLM client for Elixir.

Added

  • Text generation — generate_text/3 and stream_text/3 top-level API.
  • Streaming — StreamingResponse implementing Enumerable with composable event handlers, text stream extraction, cancel support, and incomplete stream detection.
  • Tool use — define tools with JSON schemas and handlers; automatic execution loop runs tools in parallel and feeds results back to the model.
  • Agents — Omni.Agent GenServer for stateful multi-turn conversations with lifecycle callbacks, tool approval flow, pause/resume, and prompt queuing/steering.
  • Structured output — constrained decoding via JSON Schema with per-dialect wire format handling and automatic validation/retry.
  • Providers — behaviour-based provider system with six built-in providers
    • Anthropic, OpenAI, Google Gemini, OpenRouter, Ollama, OpenCode Zen
    • Multi-dialect provider support for gateways that serve models with different wire formats
  • Dialects — wire format translation separated from provider identity
    • Anthropic Messages, OpenAI Chat Completions, OpenAI Responses, Google Gemini, Ollama Chat
  • Model catalog — hundreds of models loaded from bundled JSON data (sourced from models.dev) at startup.
  • Messages and content — two-role message model (:user, :assistant) with typed content blocks: Text, Thinking, Attachment, ToolUse, ToolResult

Versions 0.1.0 and 0.1.1, released in 2024, were early prototypes with a different architecture. Version 1.0 is a complete rewrite and is not compatible with 0.1.x.