(CODUX-182) Support custom Hugging Face memory models

[numnx]

🧩 Task: Support custom Hugging Face memory models

Automated task PR opened by Code UX for T05.

📋 Summary

Sprint: Sprint 182 · Speech Input and Memory Browser · Task: T05

Add speech-to-text input across the dashboard chat thread composer with local ONNX transcription and external API fallback, package it safely for Electron/npm runtimes, and modernize the Memory page model browser with grouped embedding/TTS models plus custom Hugging Face model links.

🤖 Model & Provider

Provider	Codex CLI
Model	gpt-5.5

⏱️ Timing

Started	2026-07-07 13:25:20 UTC
Finished	2026-07-07 13:37:08 UTC
Duration	11m 48s

📊 Token Usage

Metric	Value
Input tokens	124,618
Cached input tokens	3,961,472
Output tokens	27,539
Total tokens	4,113,629
Invocations	1
Tool calls	79
Active time	10m 37s

Included usage estimate (subscription/local login): $3.43

✅ QA Review

⏳ QA review runs after this PR is opened. Check the task card in the dashboard for the outcome once it lands.

📝 Full Task Prompt

## Objective
Add backend support for custom Hugging Face model links in the memory model catalog while keeping existing built-in embedding model behavior intact.

## Scope
- Edit `src/contracts/memory-types.ts`
- Edit `src/services/embedding-model-catalog.ts`
- Edit `src/services/embedding-model-manager.ts`
- Edit `src/server/memory-routes.ts`
- Edit memory settings defaults, schema, and sanitizer files touched by T01 as needed after consuming its settings contract
- Add or update tests under `tests/backend/services/embedding-model-catalog.test.ts`, `tests/backend/services/embedding-model-manager.test.ts`, and `tests/backend/server/memory-routes.test.ts`
- Update `docs/dashboard/memory.md` and `docs-web/user/dashboard/memory.md`

## Implementation Requirements
1. Add a typed custom embedding model definition with display name, Hugging Face repo or URL, ONNX model file path, tokenizer files, dimension, approximate size, language, and validation status.
2. Persist custom model definitions in settings or another existing durable store so they survive restart and appear alongside built-in models.
3. Add a route such as `POST /api/embedding-models/custom` that validates Hugging Face URLs/repos, rejects non-Hugging Face or malformed links, normalizes repo/path data, and returns the created catalog entry.
4. Update catalog download URL generation to work for both built-in models and custom Hugging Face sources without weakening built-in model validation.
5. Allow `download`, `select`, `delete`, and `status` routes to operate on custom embedding model IDs while preserving current behavior and errors for unknown models.
6. Add tests for valid custom links, invalid links, URL normalization, status listing, custom download URL construction, and existing built-in catalog compatibility.

## Constraints
- Do not build the Memory page UI refactor in this task.
- Do not permit arbitrary HTTP download hosts; custom model links must be Hugging Face URLs or normalized Hugging Face repo identifiers.
- Do not lower embedding dimension validation or allow selecting a custom model before its required files are downloaded.
- Avoid breaking existing memories embedded with built-in model IDs.

## Verification
- Run `pnpm exec vitest run tests/backend/services/embedding-model-catalog.test.ts tests/backend/services/embedding-model-manager.test.ts tests/backend/server/memory-routes.test.ts`.
- Run `pnpm run lint`.
- Success means built-in model tests still pass and custom Hugging Face embedding models can be added, listed, downloaded, selected, and deleted through typed routes.

🌿 Branch Info

Base: feature/CODUX-182-speech-input-and-memory-browser
Head: task/feature-codux-182-speech-input-and-memory-browse-t05-codex-mraolsfc

---

🤖 Generated by Code UX