pi-free 2.0.10 → 2.0.12

package/CHANGELOG.md

@@ -5,8 +5,48 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+## [2.0.12] - 2026-05-13
+
+### Added
+
+- **Novita AI provider** — OpenAI-compatible API at `api.novita.ai/openai/v1` with 100+ open-source models. Non-standard but rich metadata: per-model pricing (`input_token_price_per_m`), context size, max output tokens, reasoning/vision features, and model descriptions. 3 free models, 99 paid.
+
+- **FastRouter provider** — OpenRouter-compatible API at `api.fastrouter.ai/api/v1` with 170+ models. Always discovered (no auth needed for model listing). Full pricing, context lengths, and feature metadata. 129 text models (6 free, 123 paid) after filtering image/video. Set `FASTROUTER_API_KEY` for chat completions.
+
+- **Dynamic model fetching for OpenCode and OpenRouter** — Pi's built-in providers now get their models fetched dynamically from the API (`opencode.ai/zen/v1/models` and `openrouter.ai/api/v1/models`), same as Mistral, Groq, Cerebras, and xAI. Overwrites Pi's defaults with the full model list. OpenCode uses name-based free detection (API returns no pricing); OpenRouter uses full cost-based detection.
+
+- **API key reading from `~/.pi/agent/auth.json`** — `getOpencodeApiKey()` and `getOpenrouterApiKey()` now fall back to Pi's auth.json when the env var isn't set, matching how Pi's built-in providers read their keys.
+
+### Changed
+
+- **`_pricingKnown` guard in `isFreeModel`** — Providers can now signal whether pricing data is authoritative. When `_pricingKnown` is explicitly `false` (API returned no pricing), `isFreeModel` falls back to name-only detection (checks for "free" in the model name). This eliminates false positives where missing pricing data was treated as $0 cost. All affected providers (ZenMux, Together, CrofAI, dynamic-built-in, fetchOpenAICompatibleModels, deepinfra, sambanova, novita) now set this flag correctly.
+
+- **All providers now use `isFreeModel` consistently** — Together switched from hardcoded `cost===0` check to `isFreeModel`. DeepInfra and SambaNova switched from manual free lists to `isFreeModel` with proper `_pricingKnown` metadata. NVIDIA, Codestral, and Ollama explicitly documented as free-tier providers (`freeModels = allModels`).
+
+- **Unified OpenRouter-based providers** — Kilo, OpenRouter, and Cline now share the same `fetchOpenRouterCompatibleModels` / OpenRouter API logic.
+
+### Removed
+
+- **`DEFAULT_MIN_SIZE_B` (30B minimum model size filter)** — Removed from `model-fetcher.ts` and `cline-models.ts`. All models are now shown regardless of parameter count. NVIDIA still uses its own 70B threshold (`NVIDIA_MIN_SIZE_B`).
+
+### Fixed
+
+- **ZenMux false free classifications** — Models without `pricings` data (DeepSeek Chat V3.1, Kimi K2 0711, Claude 3.7 Sonnet) were incorrectly classified as free because missing pricing defaulted to $0. Fixed to 3 genuinely free models (down from 6 false positives).
+
+- **Together AI, CrofAI, dynamic-built-in missing-pricing false positives** — Same `?? 0` pattern across multiple providers could mark unpriced models as free. All now set `_pricingKnown: false` when pricing is absent from the API response.
+
 ## [2.0.10] - 2026-05-08
 
+### Fixed
+
+- **Config wipe on JSON parse failure** — `saveConfig` used `loadConfigFile()` which returns `{}` on any parse error, causing `{ ...{}, ...updates }` to write a partial config that permanently destroyed all API keys. Now reads the raw file directly and refuses to save if corrupt. `ensureConfigFile` also refuses to overwrite corrupt files.
+
+- **Built-in provider keys removed from pi-free config** — `mistral_api_key`, `groq_api_key`, `cerebras_api_key`, `xai_api_key`, and `hf_token` are no longer in `~/.pi/free.json`. These are pi's own built-in providers; their keys come from environment variables only.
+
+## [2.0.9] - 2026-05-08
+
 ### Added
 
 - **Together AI provider** — Fast inference on 200+ open-source models (Llama, DeepSeek, Qwen, etc.) through an OpenAI-compatible API. $1 trial credit on signup, no credit card required. Set `TOGETHER_AI_API_KEY`.