Three pieces of polish that don't depend on external resources.

  · packages/core/schemas/settings.schema.json (NEW, ~165 lines)
    - Full draft-07 JSON schema for settings.json: top-level
      model/baseURL/apiKeyHelper/effortLevel/permissions/autoMode/sandbox/
      hooks/allowedHttpHookUrls/disableAllHooks/memoryLoadCapKB/
      outputStyle/skillOverrides/mcpServers/disabledPlugins/worktree/voice.
    - Definitions: HookMatcherList, HookHandler, McpServer.
    - Published as a static asset (added to package.json files[]).
    - Users add `"$schema": "https://deepcode.dev/schemas/settings.schema.json"`
      to get IDE autocomplete + validation.

  · packages/core/src/config/schema.ts (NEW)
    - settingsSchemaJson() / settingsSchemaObject() — cached loaders.
    - validateSettingsShallow(obj) — fast checks for model / effortLevel
      / defaultMode / hook event names. 7 unit tests.

  · packages/core/src/vision/index.ts (NEW)
    - VisionProvider interface + ImageContentBlock shape.
    - loadImage() resolves data URL / file path / http(s) URL into
      { contentType, base64, byteSize }.
    - parseDataUrl(), guessContentType() helpers.
    - StubVisionProvider (default; rejects images).
    - OpenAICompatVisionProvider (encodes as data URL → image_url payload,
      enforces maxBytes default 20 MB). Works against Qwen-VL / GPT-4o /
      any OpenAI-compatible multimodal endpoint.
    - 14 unit tests covering data URL + local file + remote URL + size
      cap.

  · docs/MIGRATION_FROM_CLAUDE_CODE.md (NEW)
    - 5-minute switch (rename ~/.claude → ~/.deepcode + CLAUDE.md →
      AGENTS.md).
    - Field-by-field mapping (credentials/settings/skills/agents/plugins).
    - Settings.json model field map.
    - Slash command parity table.
    - Hooks + permissions + sandbox + plugins + MCP + sub-agents:
      identical schemas.
    - Behaviors that DIFFER (DeepSeek-only models, no image input until
      vision provider configured, pipeline-aware sandbox bypass).
    - Behaviors NEW in DeepCode (auto classifier, effort-bench,
      LSP bridge, VS Code extension).

Tests: 522 → 549 passing (+27: 7 schema + 14 vision + adjustments).
Build clean across all 6 packages.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>