Add prompt_progress stream chunk + prefill progress on the status line

packages/agent/src/run-agent-loop.ts

@@ -913,6 +913,11 @@ export async function runAgentLoop(opts: AgentLoopOptions): Promise<void> {
           pendingToolCalls.push(chunk.toolCall)
           onChunk(chunk)
         }
+        if (chunk.type === 'prompt_progress') {
+          // Prefill progress arrives before any text/tool delta; forward it so the
+          // UI can show a processing indicator on the current step.
+          onChunk(chunk)
+        }
         if (chunk.type === 'usage') {
           streamUsage = {
             inputTokens: chunk.inputTokens,

packages/llm/src/wire-types.ts

@@ -126,6 +126,13 @@ export type ProviderStreamChunk =
       /** Token counts are a local ~4 chars/token estimate, not agent-reported. */
       estimated?: boolean
     }
+  /**
+   * Prefill progress while the model ingests the prompt (the time-to-first-token
+   * phase). `fraction` is 0–1 when the provider reports it (e.g. LM Studio's
+   * native `prompt_processing.progress` events); omit it for an indeterminate
+   * "processing…" state. Cloud OpenAI/Anthropic do not expose this.
+   */
+  | { type: 'prompt_progress'; fraction?: number }
   | { type: 'done'; stopReason?: string }
 
 // ── The provider contract ────────────────────────────────────────────────────

src/renderer/agent-activity.test.ts

@@ -0,0 +1,21 @@
+import { describe, it } from 'node:test'
+import assert from 'node:assert/strict'
+import { promptProgressLabel } from './agent-activity.ts'
+
+describe('promptProgressLabel', () => {
+  it('is indeterminate when no fraction is reported', () => {
+    assert.equal(promptProgressLabel(undefined), 'Processing prompt…')
+  })
+
+  it('renders a rounded percentage when a fraction is reported', () => {
+    assert.equal(promptProgressLabel(0), 'Processing prompt… 0%')
+    assert.equal(promptProgressLabel(0.47), 'Processing prompt… 47%')
+    assert.equal(promptProgressLabel(1), 'Processing prompt… 100%')
+  })
+
+  it('clamps out-of-range and ignores non-finite values', () => {
+    assert.equal(promptProgressLabel(1.5), 'Processing prompt… 100%')
+    assert.equal(promptProgressLabel(-0.2), 'Processing prompt… 0%')
+    assert.equal(promptProgressLabel(Number.NaN), 'Processing prompt…')
+  })
+})

src/renderer/agent-activity.ts

@@ -6,6 +6,18 @@ import { formatTodoProgress } from '@shared/todos/todo-logic.ts'
 
 export const CONTEXT_TRIM_ACTIVITY = 'Shortened earlier messages'
 
+/**
+ * Label for the prefill phase shown on the live status line. `fraction` (0–1)
+ * is rendered as a percentage when the provider reports it (LM Studio's native
+ * progress events); without it the label is the indeterminate "Processing
+ * prompt…".
+ */
+export function promptProgressLabel(fraction: number | undefined): string {
+  if (fraction === undefined || !Number.isFinite(fraction)) return 'Processing prompt…'
+  const pct = Math.round(Math.min(1, Math.max(0, fraction)) * 100)
+  return `Processing prompt… ${pct.toString()}%`
+}
+
 export function runningToolName(thread: Thread): string | null {
   for (let i = thread.messages.length - 1; i >= 0; i--) {
     const m = thread.messages[i]

src/renderer/controller/agent.ts

@@ -31,7 +31,7 @@ import {
   finishSubagent,
 } from '@shared/store/subagent-helpers.ts'
 import { planAgentTextChunk } from '@copse/agent/agent-text-chunk.ts'
-import { syncAgentActivity, CONTEXT_TRIM_ACTIVITY } from '../agent-activity.ts'
+import { syncAgentActivity, CONTEXT_TRIM_ACTIVITY, promptProgressLabel } from '../agent-activity.ts'
 import { drainMessageQueue, enqueueHookMessage, foldBackContinuationUsed } from './message-queue.ts'
 import { usageRecordFromAgentDelta } from '@shared/usage/usage-record-input.ts'
 import type { UsageDelta } from '@shared/types'
@@ -227,6 +227,13 @@ export function startAgentController(store: AppStore, api: ApiClient): () => voi
         addUsageDelta(store, threadId, delta)
         break
       }
+      case 'prompt_progress': {
+        // Prefill happens before the first token; surface it on the live status
+        // line. The next text/tool chunk recomputes the label (Writing…/Running…)
+        // so this naturally clears once generation starts.
+        store.emit('agent_activity', threadId, promptProgressLabel(chunk.fraction))
+        break
+      }
       case 'context_pressure': {
         updateContextSnapshot(store, threadId, {
           contextWindow: chunk.contextWindow,