diff --git a/src/agents/pi-settings.test.ts b/src/agents/pi-settings.test.ts
index 608bc2f80c0..600e2d4385c 100644
--- a/src/agents/pi-settings.test.ts
+++ b/src/agents/pi-settings.test.ts
@@ -377,15 +377,25 @@ 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 GLM family
-  // name with or without a path namespace so deploys that proxy it under
-  // their own provider name still hit the guard.
-  it("flags glm- model ids regardless of path namespace", () => {
+  // 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", () => {
     expect(isSilentOverflowProneModel({ provider: "custom", modelId: "glm-5.1" })).toBe(true);
     expect(isSilentOverflowProneModel({ provider: "custom", modelId: "glm-4.7" })).toBe(true);
+  });
+
+  // Detection is intentionally narrow to z.ai-style accounting. Namespaced GLM
+  // ids that route through providers with their own overflow accounting must
+  // NOT be flagged — those hosts may not exhibit the z.ai silent-overflow
+  // shape, and disabling Pi auto-compaction for them would over-broaden the
+  // kill surface beyond the reproducible repro.
+  it("does not flag namespaced GLM ids routed through non-z.ai hosts", () => {
     expect(
       isSilentOverflowProneModel({ provider: "ollama", modelId: "ollama/glm-5.1:cloud" }),
-    ).toBe(true);
+    ).toBe(false);
+    expect(
+      isSilentOverflowProneModel({ provider: "opencode-go", modelId: "opencode-go/glm-5.1" }),
+    ).toBe(false);
   });
 
   // pi-ai's overflow.ts only documents z.ai as the silent-overflow style. We
diff --git a/src/agents/pi-settings.ts b/src/agents/pi-settings.ts
index 3c64ccf72a7..4daedcad832 100644
--- a/src/agents/pi-settings.ts
+++ b/src/agents/pi-settings.ts
@@ -131,12 +131,14 @@ export function applyPiCompactionSettingsFromConfig(params: {
  * provider call (openclaw#75799).
  *
  * True on any of: `zai-native` endpoint class, normalized provider id `zai`,
- * a `z-ai/` / `openrouter/z-ai/` model-id namespace prefix, or a `glm-` model
- * name (with or without a path namespace) — covering in-house gateways and
- * ollama-style deploys that expose Zhipu's GLM family directly without a
- * `z-ai/` qualifier. Intentionally narrow to z.ai-style accounting; other
- * providers documented as silently truncating are not added without a
- * reproducible repro.
+ * a `z-ai/` / `openrouter/z-ai/` model-id namespace prefix, or a bare `glm-`
+ * model id (no namespace prefix) — the latter covers in-house gateways that
+ * expose Zhipu's GLM family directly without a `z-ai/` qualifier. Intentionally
+ * narrow: namespaced GLM ids that route through other providers (e.g.
+ * `ollama/glm-*`, `opencode-go/glm-*`) are NOT included because their hosts
+ * have their own overflow accounting and may not exhibit the z.ai silent-
+ * overflow shape. Other providers documented as silently truncating are not
+ * added without a reproducible repro.
  */
 export function isSilentOverflowProneModel(model: {
   provider?: string | null;
@@ -157,8 +159,7 @@ export function isSilentOverflowProneModel(model: {
     if (
       normalized.startsWith("z-ai/") ||
       normalized.startsWith("openrouter/z-ai/") ||
-      normalized.startsWith("glm-") ||
-      normalized.includes("/glm-")
+      normalized.startsWith("glm-")
     ) {
       return true;
     }