Tachikoma 0.2.2

[0.2.2] - 2026-07-15

Fixed

• SwiftPM dependency checkouts no longer mistake another resolved Commander checkout for a sibling development package, avoiding conflicting package identities and a future resolution error.

[0.2.1] - 2026-07-15

Added

• Added first-class Claude Sonnet 5 and OpenAI GPT-5.6 Sol/Terra/Luna support with parsing aliases, capabilities, provider routing, and usage estimates.
• Added MiniMax-M3 with 1M context, image input, provider/model parsing, and Anthropic-compatible routing. Thanks @Tugser.
• Added first-class OpenAI chat-latest support with parsing aliases, Responses API routing, model capabilities, and usage estimates.
• Added first-class MiniMax support with the MiniMax-M2.7 catalog models, MINIMAX_API_KEY / MINIMAX_BASE_URL configuration, bearer-token Anthropic-compatible transport, model parsing shortcuts, usage estimates, and provider tests.
• Added first-class Kimi support for the official K2.6 and K2.7 Code model IDs, Moonshot's OpenAI-compatible endpoint, multimodal/tool capabilities, and provider-bound reasoning_content replay. Thanks @Tugser.
• Added explicit LM Studio model shortcuts such as lmstudio and lmstudio/openai/gpt-oss-120b so local provider selections no longer fall through to Ollama custom IDs.

Changed

• Refreshed the first-class model catalog to current provider IDs: OpenAI GPT-5.5/5.4, Claude Fable 5/Opus 4.8/Opus 4.7/Sonnet 4.6/Haiku 4.5, Gemini 3.1, Mistral latest aliases, Groq current production IDs, and xAI Grok 4.3/4.20.
• Added explicit claude-fable-5 support with 1M context, 128K max output, signed-thinking replay, refusal handling, and non-streaming generation; LanguageModel.default remains claude-opus-4-8, while LanguageModel.defaultStreaming now uses streaming-safe gpt-5.5.
• Removed stale direct model support for retired or non-canonical IDs including GPT-5.1/5.2/pseudo-thinking models, deprecated Claude Sonnet/Opus 4 snapshots, Grok 2/3/4-fast rows, old Groq Llama/Mixtral/Gemma aliases, stale Mistral aliases, and invalid LM Studio current.

Fixed

• Ollama now applies configured bearer auth, requests and exposes thinking, safely preserves same-endpoint history without persisting authentication material, maps terminal reasons without executing truncated tool calls, and streams strict NDJSON incrementally across supported platforms.
• Ollama tool conversations now replay assistant calls and named results, preserve recursive arguments and array schemas, report tool-call finishes, and surface HTTP-200 stream errors instead of silently succeeding.
• Sonnet 5 usage estimates now switch from introductory to standard pricing after August 31, 2026.
• GPT-5.6 models now retain their 372K context and 128K output limits through OpenAI-compatible, OpenRouter, and Together endpoints.
• GPT-5.6 OpenRouter routes with terminal variants such as :online, :nitro, :floor, and :exacto now retain those limits without rewriting the routed model ID.
• OpenAI gpt-5-chat-latest now preserves its distinct model identity, appears in model listings, and applies GPT-5 parameter filtering instead of being rewritten to chat-latest.
• SwiftPM consumers now resolve Commander from the package URL instead of accidentally inheriting a sibling local checkout.
• Ollama model parsing now preserves explicit custom vision model IDs such as qwen2.5vl:3b instead of falling back to llama3.3 (#16).
• Auth resolution now snapshots environment-ignore state consistently, preventing parallel tests and concurrent callers from falling back to stored OpenRouter credentials when an environment override is present.
• SwiftPM consumers can now resolve Commander remotely instead of requiring a local ../Commander checkout. Thanks @malpern.
• Custom OpenAI-compatible and Anthropic-compatible providers now honor per-provider options.apiKey values from profile config. Thanks @381181295.
• Google/Gemini request encoding now sends tool results as documented user functionResponse turns and merges consecutive same-role contents before calling the API. Thanks @hsrvc.
• Google/Gemini tool schemas now drop orphan required entries before request encoding so Gemini accepts simplified MCP tool definitions. Thanks @bcharleson.
• OpenAI Responses API providers now resolve shared OAuth/API-key credentials instead of requiring OPENAI_API_KEY directly.
• Credential loading now only maps exact API-key credential names to provider API keys, so OAuth access/refresh tokens no longer overwrite configured OpenAI or Anthropic keys.
• Custom OpenAI-compatible and Anthropic-compatible providers now forward configured proxy headers to request calls.
• Anthropic base URL environment overrides now survive Swift release builds, so ANTHROPIC_BASE_URL can route requests through local proxies. Thanks @shraderdm.