Changelog

v0.1.8

Version bump; pin Burrito OTP build version to 28.5.0 to fix macOS binary builds.

Drop :custom_openai, :ollama, and :llama_cpp providers — public provider atoms are now :anthropic | :openai | :google only
:openai with base_url set routes to the OpenAI-compatible adapter (previously required :custom_openai); without base_url routes to the standard OpenAI path
identifier defaults to "OPENAI" on :openai models when nil — resolves OPENAI_API_KEY for local endpoints that don't need a key
Planck.AI.Models.OpenAI.all/1 now accepts a base_url: opt to query a custom server's /models endpoint; without it returns the LLMDB catalog as before
Planck.AI.Models.CustomOpenAI, Planck.AI.Models.Ollama, and Planck.AI.Models.LlamaCpp removed — superseded by :openai + base_url
Planck.AI.Model gains model field — the actual provider model identifier (e.g. "claude-sonnet-4-6"); id becomes the user-facing alias. The adapter uses model || id so structs without the new field keep working
Planck.AI.Model gains has_api_key: boolean() (default true); when false, the adapter passes "not-needed" directly and skips env-var lookup — for local servers like Ollama that require no authentication
Planck.AI.Config functions from_map/1, from_list/1, and load/1 removed — superseded by from_config/2
Planck.AI.Config.from_config/2 added — builds [Model.t()] from a providers map (user-keyed, each entry has a "type" field) and a models list (each entry references a provider key); this is the v0.1.6 config format

New :custom_openai provider for OpenAI-compatible endpoints (NVIDIA, Together, vLLM, etc.)
Planck.AI.Model gains an identifier field — a short uppercase tag (e.g. "NVIDIA") used to derive the env var <IDENTIFIER>_API_KEY at request time
Planck.AI.Models.CustomOpenAI — factory (model/2) and runtime discovery (all/1) via GET {base_url}/models
Planck.AI.Config from_map/1 validates and upcases identifier; rejects values that don't match [A-Z][A-Z0-9]*
API keys for :custom_openai are resolved lazily from the environment at request time, never cached

First release.

Provider-agnostic streaming and completion API over req_llm
Lazy event-tuple stream ({:text_delta, _}, {:thinking_delta, _}, {:tool_call_complete, _}, {:done, _}, {:error, _})
Tool calling with streaming argument assembly
Model catalog for Anthropic, OpenAI, and Google via LLMDB
Local server support for Ollama and llama.cpp with runtime model discovery
JSON config loader (Planck.AI.Config)
Multimodal input: text, image, image_url, file, video_url
parse_provider/1 (private) — String.to_existing_atom → String.to_atom; @valid_providers derived from Planck.AI.list_providers() at compile time (single source of truth). Private function specs added.