fix(agents): keep state.messages intact across z.ai-style provider turns in embedded runs (#76056)

Merged via squash.

Prepared head SHA: ef305bb339
Co-authored-by: openperf <80630709+openperf@users.noreply.github.com>
Co-authored-by: openperf <80630709+openperf@users.noreply.github.com>
Reviewed-by: @openperf

CHANGELOG.md

@@ -23,6 +23,7 @@ Docs: https://docs.openclaw.ai
 - CLI/update: treat inherited Gateway service markers as origin hints and only block package replacement when the managed Gateway is still live, so self-updates can stop the service and continue safely. (#75729) Thanks @hxy91819.
 - Agents/failover: exempt run-level timeouts that fire during tool execution from model fallback, timeout-triggered compaction, and generic timeout payload synthesis. Long `process(poll)`, browser, or `exec` tool calls that exceed `agents.defaults.timeoutSeconds` previously rotated auth profiles, switched to a fallback model, and surfaced a misleading "LLM request timed out" error even though the primary model had already responded. Mirrors the existing `timedOutDuringCompaction` precedent (#46889). Fixes #52147. (#75873) Thanks @simonusa.
 - Docker: copy Bun 1.3.13 from a digest-pinned image and keep CI on the same version. Fixes #74356. Thanks @fede-kamel and @sallyom.
+- Agents/compaction: keep prior context on consecutive turns against z.ai-style providers (z.ai direct, openrouter z-ai/*, in-house GLM gateways); Pi's internal auto-compaction was misfiring after successful turns and clearing state.messages before the next provider request. (#76056) Thanks @openperf.
 
 ## 2026.5.2
 

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

@@ -75,7 +75,11 @@ import {
   setCompactionSafeguardCancelReason,
 } from "../pi-hooks/compaction-safeguard-runtime.js";
 import { createPreparedEmbeddedPiSettingsManager } from "../pi-project-settings.js";
-import { applyPiCompactionSettingsFromConfig } from "../pi-settings.js";
+import {
+  applyPiAutoCompactionGuard,
+  applyPiCompactionSettingsFromConfig,
+  isSilentOverflowProneModel,
+} from "../pi-settings.js";
 import { createOpenClawCodingTools } from "../pi-tools.js";
 import { wrapStreamFnTextTransforms } from "../plugin-text-transforms.js";
 import { registerProviderStreamForModel } from "../provider-stream.js";
@@ -960,12 +964,26 @@ async function compactEmbeddedPiSessionDirectOnce(
       });
       await resourceLoader.reload();
       // DefaultResourceLoader.reload() rehydrates settings from disk and can drop OpenClaw
-      // compaction overrides applied in createPreparedEmbeddedPiSettingsManager.
+      // compaction overrides applied in createPreparedEmbeddedPiSettingsManager — same
+      // rehydration also restores Pi's auto-compaction (openclaw#75799), so re-apply
+      // both guards. effectiveModel.baseUrl matches the surrounding scope so
+      // auth-profile-injected baseUrls reach the endpoint-class detector.
       applyPiCompactionSettingsFromConfig({
         settingsManager,
         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({
+          provider,
+          modelId,
+          baseUrl: effectiveModel.baseUrl ?? undefined,
+        }),
+      });
 
       const { customTools } = splitSdkTools({
         tools: effectiveTools,

src/agents/pi-embedded-runner/run/attempt.spawn-workspace.test-support.ts

@@ -321,6 +321,7 @@ vi.mock("../../pi-settings.js", () => ({
       keepRecentTokens: 40_000,
     },
   }),
+  isSilentOverflowProneModel: () => false,
 }));
 
 vi.mock("../extensions.js", () => ({

src/agents/pi-embedded-runner/run/attempt.ts

@@ -107,6 +107,7 @@ import { createPreparedEmbeddedPiSettingsManager } from "../../pi-project-settin
 import {
   applyPiAutoCompactionGuard,
   applyPiCompactionSettingsFromConfig,
+  isSilentOverflowProneModel,
 } from "../../pi-settings.js";
 import {
   createClientToolNameConflictError,
@@ -1474,10 +1475,16 @@ export async function runEmbeddedAttempt(
         cfg: params.config,
         contextTokenBudget: params.contextTokenBudget,
       });
-      applyPiAutoCompactionGuard({
+      const piAutoCompactionGuardArgs = {
         settingsManager,
         contextEngineInfo: activeContextEngine?.info,
-      });
+        silentOverflowProneProvider: isSilentOverflowProneModel({
+          provider: params.provider,
+          modelId: params.modelId,
+          baseUrl: params.model.baseUrl ?? undefined,
+        }),
+      };
+      applyPiAutoCompactionGuard(piAutoCompactionGuardArgs);
 
       // Sets compaction/pruning runtime state and returns extension factories
       // that must be passed to the resource loader for the safeguard to be active.
@@ -1496,12 +1503,15 @@ export async function runEmbeddedAttempt(
       });
       await resourceLoader.reload();
       // DefaultResourceLoader.reload() rehydrates settings from disk and can drop OpenClaw
-      // compaction overrides applied in createPreparedEmbeddedPiSettingsManager.
+      // compaction overrides applied in createPreparedEmbeddedPiSettingsManager — same
+      // rehydration also restores Pi's auto-compaction (openclaw#75799), so re-apply
+      // both guards.
       applyPiCompactionSettingsFromConfig({
         settingsManager,
         cfg: params.config,
         contextTokenBudget: params.contextTokenBudget,
       });
+      applyPiAutoCompactionGuard(piAutoCompactionGuardArgs);
       prepStages.mark("session-resource-loader");
 
       // Get hook runner early so it's available when creating tools

src/agents/pi-settings.test.ts

@@ -1,8 +1,10 @@
 import { describe, expect, it, vi } from "vitest";
 import { MIN_PROMPT_BUDGET_RATIO, MIN_PROMPT_BUDGET_TOKENS } from "./pi-compaction-constants.js";
 import {
+  applyPiAutoCompactionGuard,
   applyPiCompactionSettingsFromConfig,
   DEFAULT_PI_COMPACTION_RESERVE_TOKENS_FLOOR,
+  isSilentOverflowProneModel,
   resolveCompactionReserveTokensFloor,
 } from "./pi-settings.js";
 
@@ -345,3 +347,179 @@ describe("resolveCompactionReserveTokensFloor", () => {
     ).toBe(0);
   });
 });
+
+describe("isSilentOverflowProneModel", () => {
+  // Reporter's repro shape: openrouter routing to z-ai/glm. Both the bare
+  // `z-ai/...` form and the `openrouter/z-ai/...` qualified form must hit.
+  it("flags z-ai-prefixed model ids regardless of qualifier", () => {
+    expect(isSilentOverflowProneModel({ provider: "openrouter", modelId: "z-ai/glm-5.1" })).toBe(
+      true,
+    );
+    expect(
+      isSilentOverflowProneModel({ provider: "openrouter", modelId: "openrouter/z-ai/glm-5" }),
+    ).toBe(true);
+  });
+
+  it("flags a config-set z.ai provider regardless of model id", () => {
+    expect(isSilentOverflowProneModel({ provider: "z.ai", modelId: "glm-5.1" })).toBe(true);
+    expect(isSilentOverflowProneModel({ provider: "z-ai", modelId: "glm-5.1" })).toBe(true);
+  });
+
+  it("flags a direct api.z.ai baseUrl via endpointClass", () => {
+    expect(
+      isSilentOverflowProneModel({
+        provider: "openai",
+        modelId: "glm-5.1",
+        baseUrl: "https://api.z.ai/api/coding/paas/v4",
+      }),
+    ).toBe(true);
+  });
+
+  // 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 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
+  // 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(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
+  // intentionally do NOT extend the guard to anthropic/openai/google/openrouter-
+  // anthropic routes — adding them without a reproducible repro would broaden
+  // the kill surface and regress baseline behavior for those providers.
+  it("does not flag anthropic, openai, google or other routes", () => {
+    expect(
+      isSilentOverflowProneModel({ provider: "anthropic", modelId: "claude-sonnet-4.6" }),
+    ).toBe(false);
+    expect(isSilentOverflowProneModel({ provider: "openai", modelId: "gpt-5.5" })).toBe(false);
+    expect(
+      isSilentOverflowProneModel({
+        provider: "openrouter",
+        modelId: "anthropic/claude-sonnet-4.6",
+      }),
+    ).toBe(false);
+    expect(isSilentOverflowProneModel({ provider: "google", modelId: "gemini-2.5-pro" })).toBe(
+      false,
+    );
+  });
+
+  it("treats missing fields as not silent-overflow-prone", () => {
+    expect(isSilentOverflowProneModel({})).toBe(false);
+    expect(
+      isSilentOverflowProneModel({ provider: undefined, modelId: undefined, baseUrl: null }),
+    ).toBe(false);
+  });
+});
+
+describe("applyPiAutoCompactionGuard", () => {
+  // Direct repro of openclaw#75799: pi-ai's silent-overflow detection misfires
+  // on a successful turn against z.ai-style providers, triggering Pi's
+  // _runAutoCompaction from inside Session.prompt() and reassigning
+  // agent.state.messages between the runner's prompt.submitted trajectory
+  // event and the provider request. Disabling Pi auto-compaction here keeps
+  // state.messages intact; OpenClaw's preemptive compaction continues to
+  // handle real overflow on its own path.
+  it("disables Pi auto-compaction for silent-overflow-prone providers", () => {
+    const setCompactionEnabled = vi.fn();
+    const settingsManager = {
+      getCompactionReserveTokens: () => 20_000,
+      getCompactionKeepRecentTokens: () => 4_000,
+      applyOverrides: () => {},
+      setCompactionEnabled,
+    };
+
+    const result = applyPiAutoCompactionGuard({
+      settingsManager,
+      silentOverflowProneProvider: true,
+    });
+
+    expect(result).toEqual({ supported: true, disabled: true });
+    expect(setCompactionEnabled).toHaveBeenCalledWith(false);
+  });
+
+  it("disables Pi auto-compaction when a context engine plugin owns compaction", () => {
+    const setCompactionEnabled = vi.fn();
+    const settingsManager = {
+      getCompactionReserveTokens: () => 20_000,
+      getCompactionKeepRecentTokens: () => 4_000,
+      applyOverrides: () => {},
+      setCompactionEnabled,
+    };
+
+    const result = applyPiAutoCompactionGuard({
+      settingsManager,
+      contextEngineInfo: {
+        id: "third-party",
+        name: "Third-party Context Engine",
+        version: "0.1.0",
+        ownsCompaction: true,
+      },
+    });
+
+    expect(result).toEqual({ supported: true, disabled: true });
+    expect(setCompactionEnabled).toHaveBeenCalledWith(false);
+  });
+
+  // Default-mode runs against ordinary providers must keep Pi's auto-compaction
+  // enabled. Disabling it across the board would silently remove Pi's
+  // overflow-recovery path inside Session.prompt() for users who are not
+  // affected by z.ai's silent-overflow accounting.
+  it("leaves Pi auto-compaction alone for non-z.ai providers without engine ownership", () => {
+    const setCompactionEnabled = vi.fn();
+    const settingsManager = {
+      getCompactionReserveTokens: () => 20_000,
+      getCompactionKeepRecentTokens: () => 4_000,
+      applyOverrides: () => {},
+      setCompactionEnabled,
+    };
+
+    const result = applyPiAutoCompactionGuard({
+      settingsManager,
+      contextEngineInfo: {
+        id: "legacy",
+        name: "Legacy Context Engine",
+        version: "1.0.0",
+      },
+      silentOverflowProneProvider: false,
+    });
+
+    expect(result).toEqual({ supported: true, disabled: false });
+    expect(setCompactionEnabled).not.toHaveBeenCalled();
+  });
+
+  it("reports unsupported when the settings manager has no setCompactionEnabled hook", () => {
+    const settingsManager = {
+      getCompactionReserveTokens: () => 20_000,
+      getCompactionKeepRecentTokens: () => 4_000,
+      applyOverrides: () => {},
+    };
+
+    const result = applyPiAutoCompactionGuard({
+      settingsManager,
+      silentOverflowProneProvider: true,
+    });
+
+    expect(result).toEqual({ supported: false, disabled: false });
+  });
+});

src/agents/pi-settings.ts

@@ -1,6 +1,8 @@
 import type { OpenClawConfig } from "../config/types.openclaw.js";
 import type { ContextEngineInfo } from "../context-engine/types.js";
 import { MIN_PROMPT_BUDGET_RATIO, MIN_PROMPT_BUDGET_TOKENS } from "./pi-compaction-constants.js";
+import { resolveProviderEndpoint } from "./provider-attribution.js";
+import { normalizeProviderId } from "./provider-id.js";
 
 export const DEFAULT_PI_COMPACTION_RESERVE_TOKENS_FLOOR = 20_000;
 
@@ -122,18 +124,81 @@ export function applyPiCompactionSettingsFromConfig(params: {
   };
 }
 
-/** Decide whether Pi's internal auto-compaction should be disabled for this run. */
-function shouldDisablePiAutoCompaction(params: { contextEngineInfo?: ContextEngineInfo }): boolean {
-  return params.contextEngineInfo?.ownsCompaction === true;
+/**
+ * Detect providers whose pi-ai `isContextOverflow` Case 2 (silent overflow)
+ * fires on a successful turn and triggers Pi's `_runAutoCompaction` from
+ * inside `Session.prompt()`, collapsing `agent.state.messages` before the
+ * 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 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;
+  modelId?: string | null;
+  baseUrl?: string | null;
+}): boolean {
+  const provider = normalizeProviderId(typeof model.provider === "string" ? model.provider : "");
+  if (provider === "zai") {
+    return true;
+  }
+  if (typeof model.baseUrl === "string" && model.baseUrl.length > 0) {
+    if (resolveProviderEndpoint(model.baseUrl).endpointClass === "zai-native") {
+      return true;
+    }
+  }
+  if (typeof model.modelId === "string" && model.modelId.length > 0) {
+    const normalized = model.modelId.toLowerCase();
+    if (
+      normalized.startsWith("z-ai/") ||
+      normalized.startsWith("openrouter/z-ai/") ||
+      normalized.startsWith("glm-")
+    ) {
+      return true;
+    }
+  }
+  return false;
 }
 
-/** Disable Pi auto-compaction via settings when a context engine owns compaction. */
+/**
+ * Disable Pi's `_checkCompaction → _runAutoCompaction` (which would otherwise
+ * fire from inside `Session.prompt()` and reassign `agent.state.messages`
+ * before the provider call) when OpenClaw or a plugin owns compaction:
+ * `contextEngineInfo.ownsCompaction === true`, or the active model is
+ * silent-overflow-prone (openclaw#75799). Default-mode runs against ordinary
+ * providers keep Pi's auto-compaction as the existing baseline.
+ */
+function shouldDisablePiAutoCompaction(params: {
+  contextEngineInfo?: ContextEngineInfo;
+  silentOverflowProneProvider?: boolean;
+}): boolean {
+  return (
+    params.contextEngineInfo?.ownsCompaction === true || params.silentOverflowProneProvider === true
+  );
+}
+
+/**
+ * Apply the auto-compaction guard. Callers that reload a `DefaultResourceLoader`
+ * MUST call this AGAIN after each `reload()` — `settingsManager.reload()`
+ * rehydrates `compaction.enabled` from disk and silently restores Pi's
+ * default-on behavior, undoing the guard. Mirrors the existing
+ * `applyPiCompactionSettingsFromConfig` re-call pattern at the same sites.
+ */
 export function applyPiAutoCompactionGuard(params: {
   settingsManager: PiSettingsManagerLike;
   contextEngineInfo?: ContextEngineInfo;
+  silentOverflowProneProvider?: boolean;
 }): { supported: boolean; disabled: boolean } {
   const disable = shouldDisablePiAutoCompaction({
     contextEngineInfo: params.contextEngineInfo,
+    silentOverflowProneProvider: params.silentOverflowProneProvider,
   });
   const hasMethod = typeof params.settingsManager.setCompactionEnabled === "function";
   if (!disable || !hasMethod) {