fix: carry live overflow token counts

CHANGELOG.md

@@ -236,6 +236,7 @@ Docs: https://docs.openclaw.ai
 - Memory/Gemini: normalize returned Gemini embeddings across direct query, direct batch, and async batch paths so memory search uses consistent vector handling for Gemini too. (#43409) Thanks @gumadeiras.
 - Agents/failover: recognize additional serialized network errno strings plus `EHOSTDOWN` and `EPIPE` structured codes so transient transport failures trigger timeout failover more reliably. (#42830) Thanks @jnMetaCode.
 - Telegram/model picker: make inline model button selections persist the chosen session model correctly, clear overrides when selecting the configured default, and include effective fallback models in `/models` button validation. (#40105) Thanks @avirweb.
+- Agents/embedded runner: carry provider-observed overflow token counts into compaction so overflow retries and diagnostics use the rejected live prompt size instead of only transcript estimates. (#40357) thanks @rabsef-bicrym.
 
 ## 2026.3.7
 

src/agents/pi-embedded-helpers.isbillingerrormessage.test.ts

@@ -469,9 +469,9 @@ describe("extractObservedOverflowTokenCount", () => {
         '400 {"type":"error","error":{"message":"prompt is too long: 277403 tokens > 200000 maximum"}}',
       ),
     ).toBe(277403);
-    expect(extractObservedOverflowTokenCount("Context window exceeded: requested 12000 tokens")).toBe(
-      12000,
-    );
+    expect(
+      extractObservedOverflowTokenCount("Context window exceeded: requested 12000 tokens"),
+    ).toBe(12000);
     expect(
       extractObservedOverflowTokenCount(
         "This model's maximum context length is 128000 tokens. However, your messages resulted in 145000 tokens.",

src/agents/pi-embedded-runner/run.overflow-compaction.mocks.shared.ts

@@ -110,9 +110,7 @@ vi.mock("../pi-embedded-helpers.js", () => ({
   formatBillingErrorMessage: vi.fn(() => ""),
   classifyFailoverReason: vi.fn(() => null),
   extractObservedOverflowTokenCount: vi.fn((msg?: string) => {
-    const match = msg?.match(
-      /prompt is too long:\s*([\d,]+)\s+tokens\s*>\s*[\d,]+\s+maximum/i,
-    );
+    const match = msg?.match(/prompt is too long:\s*([\d,]+)\s+tokens\s*>\s*[\d,]+\s+maximum/i);
     return match?.[1] ? Number(match[1].replaceAll(",", "")) : undefined;
   }),
   formatAssistantErrorText: vi.fn(() => ""),