json-object-editor 0.10.654 → 0.10.657

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "json-object-editor",
-  "version": "0.10.654",
+  "version": "0.10.657",
   "description": "JOE the Json Object Editor | Platform Edition",
   "main": "app.js",
   "scripts": {

package/readme.md

@@ -5,6 +5,18 @@ JOE is software that allows you to manage data models via JSON objects. There ar
 
 
 
+## What’s new in 0.10.657 (brief)
+- MCP everywhere (prompts, autofill, widget):
+  - **Prompts** (`ai_prompt`): new MCP config block (`mcp_enabled`, `mcp_toolset`, `mcp_selected_tools`, `mcp_instructions_mode`) lets you turn MCP tools on per‑prompt, pick a toolset (`read-only`, `minimal`, `all`, or `custom`), and auto‑generate short tool instructions.
+  - **Autofill fields**: the same MCP keys are now supported under a field’s `ai` config so autofill runs can optionally call MCP tools with the same toolset/playlist model.
+  - **Audit**: `ai_response` records MCP config and actual tool calls (`mcp_tools_used[]`) plus `used_openai_file_ids[]` so you can see which tools and files were used for any run.
+- Uploader file roles + AI‑aware attachments:
+  - Uploader fields can define `file_roles` (e.g. `{ value:'transcript', label:'Intake Transcript', default:true }`) and JOE renders a per‑file role `<select>` that saves to `file_role` on each file object.
+  - `executeJOEAiPrompt` now sends a compact `uploaded_files[]` header (including `itemtype`, `field`, `filename`, `file_role`, and `openai_file_id`) alongside Responses input so prompts can reason about “transcript vs summary” sources while the OpenAI Files integration still handles raw content.
+  - Responses+tools (`runWithTools`) now attaches files on both the initial tool‑planning call and the final answer call, so MCP runs see the same attachments end‑to‑end.
+- History safety:
+  - Hardened `JOE.Storage.save` history diffing to avoid a `craydent-object` edge case where comparing `null`/`undefined` values could throw on `.toString()`. This only affects `_history.changes`, not what is saved.
+
 ## What’s new in 0.10.654 (brief)
 - OpenAI Files mirrored on S3 upload; uploader tiles show the `openai_file_id`. Retry upload is available per file.
 - Responses integration improvements:
@@ -220,6 +232,23 @@ JOE is software that allows you to manage data models via JSON objects. There ar
     fs.writeFileSync('downloaded.bin', buf);
     ```
 
+### File roles on uploader fields
+- **Schema configuration**:
+  - Any uploader field can declare `file_roles` as an array of `{ value, label?, default? }` objects, for example:
+    - `{ value:'transcript', label:'Intake Transcript', default:true }`
+    - `{ value:'summary', label:'Intake Summary' }`
+  - `label` is optional; it falls back to `value`. At most one role should have `default:true`.
+- **Runtime behavior**:
+  - JOE renders a role `<select>` next to each uploaded file with:
+    - A blank option, and one option per configured role.
+    - The select updates the file object’s `file_role` property in the parent object (e.g. `client.files[].file_role`).
+    - Existing uploads show the role selector on first render as long as `file_roles` is configured on the field.
+  - When OpenAI Files are enabled, uploader files still receive `openai_file_id`, `openai_purpose`, `openai_status`, and `openai_error` as before; `file_role` is an additional, JOE‑level label.
+- **AI integration**:
+  - When running an AI prompt via `executeJOEAiPrompt`, JOE inspects referenced objects for uploader fields and builds an `uploaded_files[]` header:
+    - Each entry includes `{ itemtype, field, name, role, openai_file_id }`.
+    - This header is merged into the user input so prompts can explicitly reason about which files are “transcripts”, “summaries”, etc., while the actual file bytes are attached via Responses `input_file` parts / `file_search` tool resources.
+
 ### Related endpoints (server/plugins)
 
 - `POST /API/plugin/awsConnect` – S3 upload (and OpenAI mirror when configured)
@@ -731,4 +760,4 @@ To help you develop and debug the widget + plugin in your instance, JOE exposes
 - Added a Responses‑based tool runner for `<joe-ai-widget>` that wires `ai_assistant.tools` into MCP functions via `chatgpt.runWithTools`.
 - Enhanced widget UX: assistant/user bubble theming (using `assistant_color` and user `color`), inline “tools used this turn” meta messages, and markdown rendering for assistant replies.
 - Expanded the AI widget test page with an assistant picker, live tool JSON viewer, a clickable conversation history list (resume existing `ai_widget_conversation` threads), and safer user handling (widget conversations now store user id/name/color explicitly and OAuth token‑exchange errors from Google are surfaced clearly during login).
-- Added field-level AI autofill support: schemas can declare `ai` config on a field (e.g. `{ name:'ai_summary', type:'rendering', ai:{ prompt:'Summarize the project in a few sentences.' } }`), which renders an inline “AI” button that calls `_joe.Ai.populateField('ai_summary')` and posts to `/API/plugin/chatgpt/autofill` to compute a JSON `patch` and update the UI (with confirmation before overwriting non-empty values).
+- Added field-level AI autofill support: schemas can declare `ai` config on a field (e.g. `{ name:'ai_summary', type:'rendering', ai:{ prompt:'Summarize the project in a few sentences.' } }`), which renders an inline “AI” button that calls `_joe.Ai.populateField('ai_summary')` and posts to `/API/plugin/chatgpt/autofill` to compute a JSON `patch` and update the UI (with confirmation before overwriting non-empty values).