Merge branch 'main' into backmerge/release-sdk-0.13.3

Author: NamelsKing

packages/cli/CHANGELOG.md

@@ -1,5 +1,55 @@
 # Changelog
 
+# QVAC CLI v0.7.0 Release Notes
+
+📦 **NPM:** https://www.npmjs.com/package/@qvac/cli/v/0.7.0
+
+This release adds image-to-video generation and audio encoding to the OpenAI-compatible HTTP server. It also fixes token accounting and `finish_reason` reporting across all chat-category routes.
+
+## New Features
+
+### Image-to-video via POST /v1/videos
+
+`POST /v1/videos` now supports img2vid in addition to txt2vid. Supply the reference image as a multipart file field (the form the OpenAI SDK sends for `Uploadable`), as a JSON `{ image_url }` (base64 data URI or HTTP(S) URL up to 100 MB), or as a JSON `{ file_id }` referencing a previously uploaded file. Mode is inferred automatically from the presence of `input_reference`.
+
+```typescript
+import OpenAI, { toFile } from 'openai'
+import fs from 'node:fs'
+
+const client = new OpenAI({ baseURL: 'http://localhost:11434/v1' })
+
+// img2vid via local file (multipart)
+const job = await client.videos.create({
+  model: 'wan-i2v',
+  prompt: 'subject slowly turns and smiles',
+  input_reference: await toFile(fs.createReadStream('./frame.png'), 'frame.png'),
+})
+
+// img2vid via data URI (JSON)
+const job2 = await client.videos.create({
+  model: 'wan-i2v',
+  prompt: 'subject slowly turns and smiles',
+  input_reference: { image_url: 'data:image/png;base64,...' },
+})
+```
+
+### Audio encoding — mp3, opus, aac, flac
+
+`POST /v1/audio/speech` now supports `response_format: "mp3"`, `"opus"`, `"aac"`, and `"flac"` in addition to `wav` and `pcm`. Encoding is handled by `ffmpeg` on the server's `PATH`; if ffmpeg is absent, these formats return `503 transcode_unavailable`. Use `qvac doctor` to check availability.
+
+Two new discovery endpoints are also available:
+
+```
+GET /v1/audio/voices  →  list of configured TTS voices
+GET /v1/audio/models  →  list of loaded (READY) TTS models
+```
+
+## Bug Fixes
+
+### Correct finish_reason and token accounting
+
+`finish_reason: "length"` is now returned when generation is truncated by `max_tokens` or `max_completion_tokens` (previously always `"stop"`). The Responses API equivalent is `status: "incomplete"` with `incomplete_details.reason: "max_output_tokens"`. Token counts (`completion_tokens` / `output_tokens`) now consistently use the SDK's `generatedTokens` stats across `/v1/chat/completions`, `/v1/completions`, and `/v1/responses`.
+
 ## [0.6.0]
 
 Release Date: 2026-06-02