fix(agents): strip stale anthropic assistant prefill

CHANGELOG.md

@@ -18,6 +18,7 @@ Docs: https://docs.openclaw.ai
 - Agents/sessions: acquire the session write lock only after cold bootstrap, plugin, and tool setup so fallback runs are not blocked by stalled pre-model startup work. Thanks @codex.
 - Gateway/models: skip external OpenRouter and LiteLLM pricing refreshes for local/self-hosted model endpoints so startup does not wait on remote pricing catalogs for local-only Ollama, vLLM, and compatible providers. Thanks @codex.
 - CLI/plugins: stop security-blocked plugin installs from retrying as hook packs, so normal plugin packages report the scanner failure without a misleading "not a valid hook pack" follow-up. Fixes #61175; supersedes #64102. Thanks @KonsultDigital and @ziyincody.
+- Agents/Anthropic: strip stale trailing assistant prefill turns from outbound replay so context-engine short circuits cannot send unsupported assistant-prefill payloads to provider APIs. Fixes #72556. Thanks @Veda-openclaw.
 - Control UI/Dreaming: require explicit confirmation before applying restart-impacting Dreaming mode changes, with restart warning copy and loading feedback. Fixes #63804. (#63807) Thanks @bbddbb1.
 - CLI/update: keep the automatic post-update completion refresh on the core-command tree so it no longer stages bundled plugin runtime deps before the Gateway restart path, avoiding `.24` update hangs and 1006 disconnect cascades. Fixes #72665. Thanks @sakalaboator and @He-Pin.
 - Agents/Bedrock: stop heartbeat runs from persisting blank user transcript turns and repair existing blank user text messages before replay, preventing AWS Bedrock `ContentBlock` blank-text validation failures. Fixes #72640 and #72622. Thanks @goldzulu.

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

@@ -1633,6 +1633,44 @@ describe("wrapStreamFnSanitizeMalformedToolCalls", () => {
     expect(seenContext.messages).toBe(messages);
   });
 
+  it("strips trailing assistant prefill turns for Anthropic outbound replay", async () => {
+    const messages = [
+      {
+        role: "user",
+        content: [{ type: "text", text: "earlier question" }],
+      },
+      {
+        role: "assistant",
+        content: [{ type: "text", text: "stale assistant answer" }],
+      },
+    ];
+    const baseFn = vi.fn((_model, _context) =>
+      createFakeStream({ events: [], resultMessage: { role: "assistant", content: [] } }),
+    );
+
+    const wrapped = wrapStreamFnSanitizeMalformedToolCalls(baseFn as never, new Set(["read"]), {
+      validateAnthropicTurns: true,
+      preserveSignatures: true,
+      dropThinkingBlocks: false,
+    } as never);
+    const stream = wrapped(
+      { api: "anthropic-messages" } as never,
+      { messages } as never,
+      {} as never,
+    ) as FakeWrappedStream | Promise<FakeWrappedStream>;
+    await Promise.resolve(stream);
+
+    expect(baseFn).toHaveBeenCalledTimes(1);
+    const seenContext = baseFn.mock.calls[0]?.[1] as { messages: unknown[] };
+    expect(seenContext.messages).toEqual([
+      {
+        role: "user",
+        content: [{ type: "text", text: "earlier question" }],
+      },
+    ]);
+    expect(seenContext.messages).not.toBe(messages);
+  });
+
   it("drops signed thinking turns when sibling replay tool calls are not allowlisted", async () => {
     const messages = [
       {

src/agents/pi-embedded-runner/run/attempt.tool-call-normalization.ts

@@ -525,6 +525,32 @@ function sanitizeAnthropicReplayToolResults(
   return changed ? out : messages;
 }
 
+function assistantTurnHasReplayToolCall(message: AgentMessage): boolean {
+  if (!message || typeof message !== "object" || message.role !== "assistant") {
+    return false;
+  }
+  const content = (message as { content?: unknown }).content;
+  if (!Array.isArray(content)) {
+    return false;
+  }
+  return content.some((block) => isReplayToolCallBlock(block));
+}
+
+function stripTrailingAssistantPrefillTurns(messages: AgentMessage[]): AgentMessage[] {
+  let end = messages.length;
+  while (end > 0) {
+    const message = messages[end - 1];
+    if (!message || typeof message !== "object" || message.role !== "assistant") {
+      break;
+    }
+    if (assistantTurnHasReplayToolCall(message)) {
+      break;
+    }
+    end -= 1;
+  }
+  return end === messages.length ? messages : messages.slice(0, end);
+}
+
 function normalizeToolCallIdsInMessage(message: unknown): void {
   if (!message || typeof message !== "object") {
     return;
@@ -873,6 +899,7 @@ export function wrapStreamFnSanitizeMalformedToolCalls(
       nextMessages = sanitizeAnthropicReplayToolResults(nextMessages, {
         disallowEmbeddedUserToolResultsForSignedThinkingReplay: allowProviderOwnedThinkingReplay,
       });
+      nextMessages = stripTrailingAssistantPrefillTurns(nextMessages);
     }
     if (nextMessages === messages) {
       return baseFn(model, context, options);