pr/0ff80446a fix whisper mode hiding rich final answe

.github/skills/coc-knowledge/references/admin-config.md

@@ -37,7 +37,7 @@ The admin page is structured as a two-column shell that fills the available vert
 - `<div className="ar-shell">` is a CSS grid (`var(--ar-sidebar-w)` + `1fr`) with `height: 100%; min-height: 0; overflow: hidden`.
 - `<aside className="ar-sidebar">` fills the grid row (`height: 100%; min-height: 0`) and only scrolls internally if its own brand/nav/stats stack ever exceeds the viewport. It is **not** sticky and must not use `100vh` — both would break inside the `AdminDialog` and any nested pane whose container height differs from the viewport.
 - `<main className="ar-main">` is the **single scroll region** of the admin route: `min-height: 0; height: 100%; overflow-y: auto`. The sticky topbar (`.ar-topbar` with the `.ar-breadcrumb`) pins to the top of this scroller, and the page body (`.ar-page` with `.ar-page-header` + cards) flows underneath.
-- The tabs live in the sidebar as `.ar-nav-item` buttons grouped by user task: Configure, Connections, Operations, and Developer / Internals. The grouped nav can mix admin sections (`admin-tab-*` data-testids), promoted settings sections (`settings-subtab-*` data-testids), and embedded tool routes (`skills-toggle`, `logs-toggle`, `stats-toggle`, `models-toggle`, `servers-toggle`). A grouped `.ar-mobile-tab-select` appears only under the responsive `@media (max-width: 600px)` rule, which hides the sidebar and falls back to a `<select>` with `<optgroup>` labels — the main pane still scrolls internally on mobile.
+- The tabs live in the sidebar as `.ar-nav-item` buttons grouped by user task: Configure, Knowledge, Connections (container-only), Operations, and Developer / Internals. The Configure group contains the settings entry, AI Provider, Providers, and Servers. The Connections group only appears in container mode (for Messaging and container Agents). The grouped nav can mix admin sections (`admin-tab-*` data-testids), promoted settings sections (`settings-subtab-*` data-testids), and embedded tool routes (`skills-toggle`, `logs-toggle`, `stats-toggle`, `models-toggle`, `servers-toggle`). A grouped `.ar-mobile-tab-select` appears only under the responsive `@media (max-width: 600px)` rule, which hides the sidebar and falls back to a `<select>` with `<optgroup>` labels — the main pane still scrolls internally on mobile.
 
 ### Settings Sub-Tabs
 

.github/skills/coc-knowledge/references/dashboard-spa.md

@@ -91,6 +91,15 @@ Inside `WhisperCollapsedGroup`, tool calls render as compact "whisper-row" varia
 - Single flat row: kind pill + truncated summary + duration + chevron
 - Color-coded pills: Read/blue, Grep/Glob/green, Edit/Write/amber, Shell/PS/SQL/purple, Skill/grey
 
+In whisper mode (`toolCompactness === 3`), `filterWhisperChunks` keeps a tail of
+the final assistant message plus any `task_complete`/visible `ask_user` chunks,
+collapsing everything else into one summary group. The final message is the last
+`content` chunk plus earlier content chunks separated from it only by
+non-breaking trailing tools (`suggest_follow_ups`, `report_intent`,
+`task_complete`, `ask_user`); the walk-back stops at the first substantive
+tool/tool-group. This keeps a rich answer visible even when a hidden
+`suggest_follow_ups` call splits it from a trivial closing line.
+
 Chat commit strips are detected from real shell output on `powershell`, `shell`,
 and `bash` tool calls. The detector only treats commit-creating commands
 (`git commit`, `git merge`, `git cherry-pick`, `git revert`) with native git

.github/skills/coc-knowledge/references/ralph.md

@@ -105,6 +105,20 @@ The SPA shows a **"Promote to Ralph"** pill in the follow-up area for eligible
 chats and calls this endpoint via `coc-client`'s `processes.promoteToRalph`
 helper.
 
+### Grilling-Phase Prompt Injection
+
+During the `grilling` phase, `chat-base-executor` prepends a directive to the
+**user message** (never the system message) via `buildRalphGrillSuffix(...)`
+(`packages/coc/src/server/executors/chat-base-executor.ts`). It carries the
+`ultra-ralph` grill-section pointer, the `## Goal` machine contract, and — when
+an `AutoFolderContext` resolves — an explicit goal-file save-location directive
+pointing at the repo's `notes/Plans` root (`~/.coc/repos/<workspaceId>/notes/Plans/`)
+with a `*.goal.md` filename. This keeps the goal file out of the repository
+working tree and lets the Notes/scratchpad UI open and edit it (`isGoalFile`
+detects `*.goal.md`). The generic bundled `grill-me` skill stays host-agnostic:
+it defers to whatever save location the host supplies and only falls back to a
+working-directory-relative `Plans/<area>/<feature>/` when none is given.
+
 ## Resume Routes
 
 Session resume endpoints share infrastructure in

.github/skills/coc-knowledge/references/rest-api.md

@@ -221,6 +221,17 @@ See [mcp-settings.md](mcp-settings.md).
 | PATCH | `/api/workspaces/:id/work-items/:itemId` | Update work item |
 | DELETE | `/api/workspaces/:id/work-items/:itemId` | Delete work item |
 
+### AI Authoring (gated by `workItems.aiAuthoring` flag, default `false`)
+
+Draft generation is ephemeral — no data is persisted until the caller explicitly applies it via the standard create/update/plan endpoints.
+
+Response shape: `{ kind: 'clarification', questions: string[], clarificationCount: number }` or `{ kind: 'draft', workItem: {...}, goal?: string, childTasks?: [...] }`.
+
+| Method | Path | Description |
+|--------|------|-------------|
+| POST | `/api/workspaces/:id/work-items/ai-draft` | Generate a draft for a **new** work item from a prompt. Body: `{ prompt, type?, parentId?, clarificationAnswers?, clarificationCount? }`. Returns clarification (up to 3 rounds) or a draft. |
+| POST | `/api/workspaces/:id/work-items/:itemId/ai-draft` | Generate an **improvement** draft for an existing work item. Body: `{ prompt, targets?: ['fields','goal','childTasks'], clarificationAnswers?, clarificationCount? }`. Returns clarification or a draft. |
+
 ## Seen State
 
 | Method | Path | Description |

packages/coc-agent-sdk/src/session-telemetry.ts

@@ -33,6 +33,9 @@ export class SessionTelemetry {
     private usageTurnCount = 0;
     private usageTokenLimit: number | undefined;
     private usageCurrentTokens: number | undefined;
+    private usageSystemTokens: number | undefined;
+    private usageToolDefinitionsTokens: number | undefined;
+    private usageConversationTokens: number | undefined;
 
     // ── Tool call tracking ──────────────────────────────────────────────────
     readonly activeToolCalls = new Map<string, ActiveToolCall>();
@@ -69,9 +72,15 @@ export class SessionTelemetry {
     recordUsageInfo(data: {
         tokenLimit?: number;
         currentTokens?: number;
+        systemTokens?: number;
+        toolDefinitionsTokens?: number;
+        conversationTokens?: number;
     }): void {
         if (data.tokenLimit    != null) { this.usageTokenLimit    = data.tokenLimit; }
         if (data.currentTokens != null) { this.usageCurrentTokens = data.currentTokens; }
+        if (data.systemTokens           != null) { this.usageSystemTokens           = data.systemTokens; }
+        if (data.toolDefinitionsTokens  != null) { this.usageToolDefinitionsTokens  = data.toolDefinitionsTokens; }
+        if (data.conversationTokens     != null) { this.usageConversationTokens     = data.conversationTokens; }
     }
 
     buildTokenUsage(): TokenUsage | undefined {
@@ -87,6 +96,9 @@ export class SessionTelemetry {
             turnCount: this.usageTurnCount,
             tokenLimit: this.usageTokenLimit,
             currentTokens: this.usageCurrentTokens,
+            ...(this.usageSystemTokens          != null ? { systemTokens:           this.usageSystemTokens }          : {}),
+            ...(this.usageToolDefinitionsTokens != null ? { toolDefinitionsTokens:  this.usageToolDefinitionsTokens } : {}),
+            ...(this.usageConversationTokens    != null ? { conversationTokens:     this.usageConversationTokens }    : {}),
         };
     }
 

packages/coc-agent-sdk/src/streaming-session.ts

@@ -79,6 +79,9 @@ export interface ISessionEvent {
         // Session quota (session.usage_info)
         tokenLimit?: number;
         currentTokens?: number;
+        systemTokens?: number;
+        toolDefinitionsTokens?: number;
+        conversationTokens?: number;
         // Tool execution (tool.execution_start / tool.execution_complete)
         toolCallId?: string;
         toolName?: string;
@@ -471,8 +474,11 @@ export class StreamingSession {
 
     private handleUsageInfo(event: ISessionEvent): void {
         this.telemetry.recordUsageInfo({
-            tokenLimit:    event.data?.tokenLimit,
-            currentTokens: event.data?.currentTokens,
+            tokenLimit:            event.data?.tokenLimit,
+            currentTokens:         event.data?.currentTokens,
+            systemTokens:          event.data?.systemTokens,
+            toolDefinitionsTokens: event.data?.toolDefinitionsTokens,
+            conversationTokens:    event.data?.conversationTokens,
         });
         const usage = this.telemetry.buildTokenUsage();
         this.sessionLog.debug(

packages/coc-agent-sdk/src/types.ts

@@ -361,6 +361,12 @@ export interface TokenUsage {
     tokenLimit?: number;
     /** Session-level current token count (last seen from session.usage_info) */
     currentTokens?: number;
+    /** Tokens consumed by the system prompt (from session.usage_info breakdown) */
+    systemTokens?: number;
+    /** Tokens consumed by tool definitions (from session.usage_info breakdown) */
+    toolDefinitionsTokens?: number;
+    /** Tokens consumed by conversation history (from session.usage_info breakdown) */
+    conversationTokens?: number;
 }
 
 // ============================================================================

packages/coc-agent-sdk/test/copilot-sdk-wrapper/session-telemetry.test.ts

@@ -47,6 +47,48 @@ describe('SessionTelemetry — token usage', () => {
         expect(usage!.tokenLimit).toBe(10000);
         expect(usage!.currentTokens).toBe(500);
     });
+
+    it('records usage info breakdown (system, tool, conversation tokens)', () => {
+        const t = new SessionTelemetry();
+        t.recordUsage({ inputTokens: 1, outputTokens: 1 });
+        t.recordUsageInfo({
+            tokenLimit: 200000,
+            currentTokens: 70000,
+            systemTokens: 12400,
+            toolDefinitionsTokens: 8100,
+            conversationTokens: 47200,
+        });
+
+        const usage = t.buildTokenUsage();
+        expect(usage!.tokenLimit).toBe(200000);
+        expect(usage!.currentTokens).toBe(70000);
+        expect(usage!.systemTokens).toBe(12400);
+        expect(usage!.toolDefinitionsTokens).toBe(8100);
+        expect(usage!.conversationTokens).toBe(47200);
+    });
+
+    it('omits breakdown fields when not provided', () => {
+        const t = new SessionTelemetry();
+        t.recordUsage({ inputTokens: 1, outputTokens: 1 });
+        t.recordUsageInfo({ tokenLimit: 100000, currentTokens: 5000 });
+
+        const usage = t.buildTokenUsage();
+        expect(usage!.systemTokens).toBeUndefined();
+        expect(usage!.toolDefinitionsTokens).toBeUndefined();
+        expect(usage!.conversationTokens).toBeUndefined();
+    });
+
+    it('keeps last seen breakdown values across multiple recordUsageInfo calls', () => {
+        const t = new SessionTelemetry();
+        t.recordUsage({ inputTokens: 1, outputTokens: 1 });
+        t.recordUsageInfo({ tokenLimit: 100000, systemTokens: 5000 });
+        t.recordUsageInfo({ currentTokens: 20000, toolDefinitionsTokens: 3000 });
+
+        const usage = t.buildTokenUsage();
+        expect(usage!.systemTokens).toBe(5000);
+        expect(usage!.toolDefinitionsTokens).toBe(3000);
+        expect(usage!.conversationTokens).toBeUndefined();
+    });
 });
 
 describe('SessionTelemetry — tool call tracking', () => {

packages/coc-agent-sdk/test/copilot-sdk-wrapper/streaming-session.test.ts

@@ -409,6 +409,50 @@ describe('StreamingSession — token usage', () => {
         const result = await promise;
         expect(result.tokenUsage).toBeUndefined();
     });
+
+    it('captures session.usage_info breakdown (system, tool, conversation tokens)', async () => {
+        const { session, emit } = makeMockSession();
+        const ss = new StreamingSession();
+        const promise = ss.run(session, baseOptions());
+
+        emit({ type: 'assistant.usage', data: { inputTokens: 1, outputTokens: 1 } });
+        emit({
+            type: 'session.usage_info',
+            data: {
+                tokenLimit: 200000,
+                currentTokens: 70000,
+                systemTokens: 12400,
+                toolDefinitionsTokens: 8100,
+                conversationTokens: 47200,
+            },
+        });
+        emit({ type: 'session.idle' });
+
+        const result = await promise;
+        expect(result.tokenUsage).toBeDefined();
+        expect(result.tokenUsage!.tokenLimit).toBe(200000);
+        expect(result.tokenUsage!.currentTokens).toBe(70000);
+        expect(result.tokenUsage!.systemTokens).toBe(12400);
+        expect(result.tokenUsage!.toolDefinitionsTokens).toBe(8100);
+        expect(result.tokenUsage!.conversationTokens).toBe(47200);
+    });
+
+    it('handles session.usage_info without breakdown fields', async () => {
+        const { session, emit } = makeMockSession();
+        const ss = new StreamingSession();
+        const promise = ss.run(session, baseOptions());
+
+        emit({ type: 'assistant.usage', data: { inputTokens: 1, outputTokens: 1 } });
+        emit({ type: 'session.usage_info', data: { tokenLimit: 100000, currentTokens: 5000 } });
+        emit({ type: 'session.idle' });
+
+        const result = await promise;
+        expect(result.tokenUsage!.tokenLimit).toBe(100000);
+        expect(result.tokenUsage!.currentTokens).toBe(5000);
+        expect(result.tokenUsage!.systemTokens).toBeUndefined();
+        expect(result.tokenUsage!.toolDefinitionsTokens).toBeUndefined();
+        expect(result.tokenUsage!.conversationTokens).toBeUndefined();
+    });
 });
 
 // ─────────────────────────────────────────────────────────────────────────────

packages/coc-client/src/contracts/admin.ts

@@ -160,6 +160,7 @@ export interface RuntimeDashboardConfig {
     claudeEnabled: boolean;
     defaultProvider: 'copilot' | 'codex' | 'claude';
     workItemsHierarchyEnabled: boolean;
+    workItemsAiAuthoringEnabled: boolean;
     gitCommitLookupEnabled: boolean;
     effortLevelsEnabled: boolean;
   };

packages/coc-client/src/contracts/work-items.ts

@@ -290,3 +290,78 @@ export interface WorkItemTreeFilter {
   /** When true, items with status "done" are included. Defaults to false. */
   includeDone?: boolean;
 }
+
+// ============================================================================
+// AI Authoring types
+// ============================================================================
+
+/** Fields that an AI draft can carry for a work item (all optional). */
+export interface WorkItemAiDraftFields {
+  title?: string;
+  description?: string;
+  priority?: WorkItemPriority;
+  tags?: string[];
+  /** Markdown plan / goal content (maps to plan.content). */
+  plan?: string;
+  /** For goal type items: success criteria. */
+  successCriteria?: string;
+  /** Work item type suggested by the AI. */
+  type?: WorkItemType;
+}
+
+/** A drafted child task (leaf work item) for a hierarchy breakdown. */
+export interface WorkItemChildTaskDraft {
+  title: string;
+  description?: string;
+  /** 'work-item' or 'bug' — child leaf types. */
+  type?: 'work-item' | 'bug';
+}
+
+/** Response when the AI needs more information before generating a draft. */
+export interface WorkItemAiClarificationResponse {
+  kind: 'clarification';
+  /** Up to 3 concise clarification questions. */
+  questions: string[];
+  /** Total number of clarification rounds completed so far (0-based). */
+  clarificationCount: number;
+}
+
+/** Response when the AI has produced a complete draft. */
+export interface WorkItemAiDraftResult {
+  kind: 'draft';
+  /** The generated work item fields. */
+  workItem: WorkItemAiDraftFields;
+  /** Optional goal/plan markdown stored as plan.content. */
+  goal?: string;
+  /** Optional child task breakdown (only when hierarchy is applicable). */
+  childTasks?: WorkItemChildTaskDraft[];
+}
+
+/** Union of all possible AI draft API responses. */
+export type WorkItemAiGenerationResponse = WorkItemAiClarificationResponse | WorkItemAiDraftResult;
+
+/** Request body for POST /api/workspaces/:id/work-items/ai-draft */
+export interface NewWorkItemAiDraftRequest extends JsonObject {
+  /** Free-text user prompt describing the feature / problem. Required. */
+  prompt: string;
+  /** Hint for the type to generate (defaults to 'work-item'). */
+  type?: WorkItemType;
+  /** Parent work item ID for hierarchy context. */
+  parentId?: string;
+  /** Answers to previous clarification questions. */
+  clarificationAnswers?: string[];
+  /** Number of clarification rounds already completed (0 = first request). */
+  clarificationCount?: number;
+}
+
+/** Request body for POST /api/workspaces/:id/work-items/:workItemId/ai-draft */
+export interface ImproveWorkItemAiDraftRequest extends JsonObject {
+  /** Instruction for what to improve. Required. */
+  prompt: string;
+  /** Which aspects to draft ('fields', 'goal', 'childTasks'). Defaults to ['fields', 'goal']. */
+  targets?: Array<'fields' | 'goal' | 'childTasks'>;
+  /** Answers to previous clarification questions. */
+  clarificationAnswers?: string[];
+  /** Number of clarification rounds already completed. */
+  clarificationCount?: number;
+}

packages/coc-client/src/domains/work-items.ts

@@ -3,11 +3,14 @@ import type {
   CreateWorkItemRequest,
   ExecuteWorkItemRequest,
   ExecuteWorkItemResponse,
+  ImproveWorkItemAiDraftRequest,
+  NewWorkItemAiDraftRequest,
   RequestWorkItemChangesRequest,
   RequestWorkItemChangesResponse,
   ResolveWorkItemCommentsRequest,
   UpdateWorkItemRequest,
   WorkItem,
+  WorkItemAiGenerationResponse,
   WorkItemFilter,
   WorkItemGroupedResponse,
   WorkItemListResponse,
@@ -143,4 +146,18 @@ export class WorkItemsClient {
     if (filter?.includeDone !== undefined) query.includeDone = filter.includeDone;
     return this.transport.request<WorkItemTreeResponse>(path(workspaceId, '/tree'), { query });
   }
+
+  aiDraft(workspaceId: string, request: NewWorkItemAiDraftRequest): Promise<WorkItemAiGenerationResponse> {
+    return this.transport.request<WorkItemAiGenerationResponse>(path(workspaceId, '/ai-draft'), {
+      method: 'POST',
+      body: { ...request },
+    });
+  }
+
+  aiImprove(workspaceId: string, workItemId: string, request: ImproveWorkItemAiDraftRequest): Promise<WorkItemAiGenerationResponse> {
+    return this.transport.request<WorkItemAiGenerationResponse>(
+      path(workspaceId, `/${encodePathSegment(workItemId)}/ai-draft`),
+      { method: 'POST', body: { ...request } },
+    );
+  }
 }