fix(agents): complete pi-settings mock in compact.hooks harness

src/agents/pi-embedded-runner/compact.hooks.harness.ts

@@ -301,8 +301,10 @@ export async function loadCompactHooksHarness(): Promise<{
   }));
 
   vi.doMock("../pi-settings.js", () => ({
+    applyPiAutoCompactionGuard: vi.fn(() => ({ supported: true, disabled: false })),
     applyPiCompactionSettingsFromConfig: vi.fn(),
     ensurePiCompactionReserveTokens: vi.fn(),
+    isSilentOverflowProneModel: vi.fn(() => false),
     resolveCompactionReserveTokensFloor: vi.fn(() => 0),
   }));
 

src/agents/pi-embedded-runner/compact.ts

@@ -973,6 +973,9 @@ async function compactEmbeddedPiSessionDirectOnce(
         cfg: params.config,
         contextTokenBudget: ctxInfo.tokens,
       });
+      // contextEngineInfo is intentionally omitted: this guard runs inside the
+      // compaction LLM session, which is not the user-facing agent session and
+      // has no associated context engine.
       applyPiAutoCompactionGuard({
         settingsManager,
         silentOverflowProneProvider: isSilentOverflowProneModel({

src/agents/pi-settings.test.ts

@@ -378,10 +378,16 @@ describe("isSilentOverflowProneModel", () => {
   // openclaw#75799 reporter's setup: an OpenAI-compatible in-house gateway
   // exposing Zhipu's GLM family directly (model id `glm-5.1`, no `z-ai/`
   // qualifier, custom baseUrl that is not api.z.ai). Catch the bare GLM
-  // family name so direct gateway deployments hit the guard.
-  it("flags bare glm- model ids without a namespace prefix", () => {
+  // family name so direct gateway deployments hit the guard regardless of
+  // what `provider` field the user picked — gateways relabel the upstream
+  // identity, so `provider` here can be anything from `openai` to a custom
+  // string. False positives only disable Pi's secondary compaction path;
+  // OpenClaw's preemptive compaction continues to handle real overflow.
+  it("flags bare glm- model ids without a namespace prefix, regardless of provider", () => {
     expect(isSilentOverflowProneModel({ provider: "custom", modelId: "glm-5.1" })).toBe(true);
     expect(isSilentOverflowProneModel({ provider: "custom", modelId: "glm-4.7" })).toBe(true);
+    expect(isSilentOverflowProneModel({ provider: "openai", modelId: "glm-5.1" })).toBe(true);
+    expect(isSilentOverflowProneModel({ provider: "openrouter", modelId: "glm-5.1" })).toBe(true);
   });
 
   // Detection is intentionally narrow to z.ai-style accounting. Namespaced GLM