diff --git a/CHANGELOG.md b/CHANGELOG.md
index a0b57cc1b45..43f30946ed3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -254,6 +254,7 @@ Docs: https://docs.openclaw.ai
 - Models/default provider fallback: when the hardcoded default provider is removed from `models.providers`, resolve defaults from configured providers instead of reporting stale removed-provider defaults in status output. (#38947) Thanks @davidemanuelDEV.
 - Agents/cache-trace stability: guard stable stringify against circular references in trace payloads so near-limit payloads no longer crash with `Maximum call stack size exceeded`; adds regression coverage. (#38935) Thanks @MumuTW.
 - Extensions/diffs CI stability: add `headers` to the `localReq` test helper in `extensions/diffs/index.test.ts` so forwarding-hint checks no longer crash with `req.headers` undefined. (supersedes #39063) Thanks @Shennng.
+- Agents/compaction thresholding: apply `agents.defaults.contextTokens` cap to the model passed into embedded run and `/compact` session creation so auto-compaction thresholds use the effective context window, not native model max context. (#39099) Thanks @MumuTW.
 
 ## 2026.3.2
 
diff --git a/src/agents/pi-embedded-runner/compact.hooks.test.ts b/src/agents/pi-embedded-runner/compact.hooks.test.ts
index ce8b9e0f696..9745071654d 100644
--- a/src/agents/pi-embedded-runner/compact.hooks.test.ts
+++ b/src/agents/pi-embedded-runner/compact.hooks.test.ts
@@ -173,6 +173,7 @@ vi.mock("../date-time.js", () => ({
 vi.mock("../defaults.js", () => ({
   DEFAULT_MODEL: "fake-model",
   DEFAULT_PROVIDER: "openai",
+  DEFAULT_CONTEXT_TOKENS: 128_000,
 }));
 
 vi.mock("../utils.js", () => ({
diff --git a/src/agents/pi-embedded-runner/compact.ts b/src/agents/pi-embedded-runner/compact.ts
index 92bf4b97f23..a3d02596886 100644
--- a/src/agents/pi-embedded-runner/compact.ts
+++ b/src/agents/pi-embedded-runner/compact.ts
@@ -376,6 +376,20 @@ export async function compactEmbeddedPiSessionDirect(
       sessionId: params.sessionId,
       warn: makeBootstrapWarn({ sessionLabel, warn: (message) => log.warn(message) }),
     });
+    // Apply contextTokens cap to model so pi-coding-agent's auto-compaction
+    // threshold uses the effective limit, not the native context window.
+    const ctxInfo = resolveContextWindowInfo({
+      cfg: params.config,
+      provider,
+      modelId,
+      modelContextWindow: model.contextWindow,
+      defaultTokens: DEFAULT_CONTEXT_TOKENS,
+    });
+    const effectiveModel =
+      ctxInfo.tokens < (model.contextWindow ?? Infinity)
+        ? { ...model, contextWindow: ctxInfo.tokens }
+        : model;
+
     const runAbortController = new AbortController();
     const toolsRaw = createOpenClawCodingTools({
       exec: {
@@ -398,7 +412,7 @@ export async function compactEmbeddedPiSessionDirect(
       abortSignal: runAbortController.signal,
       modelProvider: model.provider,
       modelId,
-      modelContextWindowTokens: model.contextWindow,
+      modelContextWindowTokens: ctxInfo.tokens,
       modelAuthMode: resolveModelAuthMode(model.provider, params.config),
     });
     const tools = sanitizeToolsForGoogle({
@@ -594,7 +608,7 @@ export async function compactEmbeddedPiSessionDirect(
         agentDir,
         authStorage,
         modelRegistry,
-        model,
+        model: effectiveModel,
         thinkingLevel: mapThinkingLevel(params.thinkLevel),
         tools: builtInTools,
         customTools,
diff --git a/src/agents/pi-embedded-runner/run.ts b/src/agents/pi-embedded-runner/run.ts
index c1d1d414c49..80ef934d63e 100644
--- a/src/agents/pi-embedded-runner/run.ts
+++ b/src/agents/pi-embedded-runner/run.ts
@@ -375,6 +375,12 @@ export async function runEmbeddedPiAgent(
         modelContextWindow: model.contextWindow,
         defaultTokens: DEFAULT_CONTEXT_TOKENS,
       });
+      // Apply contextTokens cap to model so pi-coding-agent's auto-compaction
+      // threshold uses the effective limit, not the native context window.
+      const effectiveModel =
+        ctxInfo.tokens < (model.contextWindow ?? Infinity)
+          ? { ...model, contextWindow: ctxInfo.tokens }
+          : model;
       const ctxGuard = evaluateContextWindowGuard({
         info: ctxInfo,
         warnBelowTokens: CONTEXT_WINDOW_WARN_BELOW_TOKENS,
@@ -866,7 +872,7 @@ export async function runEmbeddedPiAgent(
             disableTools: params.disableTools,
             provider,
             modelId,
-            model,
+            model: effectiveModel,
             authProfileId: lastProfileId,
             authProfileIdSource: lockedProfileId ? "user" : "auto",
             authStorage,