fix(agent): retry empty anthropic-compatible replies

steipete · steipete · commit be166b9ae48d · 2026-05-15T08:52:10.000+01:00
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -34,6 +34,7 @@ Docs: https://docs.openclaw.ai
 
 ### Fixes
 
+- Agents: retry empty final turns for generic `anthropic-messages` providers instead of limiting non-visible recovery to Kimi, so custom/proxied Anthropic-compatible routes can recover with a visible answer. Addresses #46080. Thanks @wmgx, @w1tv, and @iFwu.
 - Control UI: rotate browser service-worker caches per build so updated Gateways are less likely to keep serving stale dashboard bundles that trigger protocol mismatch errors.
 - Discord: report unresolved configured bot-token SecretRefs during startup instead of treating the account as unconfigured. (#82009) Thanks @giodl73-repo.
 - CLI/config: preserve numeric-looking object keys such as Discord guild IDs during `config patch` recursive merges. (#81999) Thanks @giodl73-repo.
diff --git a/src/agents/pi-embedded-runner/run.incomplete-turn.test.ts b/src/agents/pi-embedded-runner/run.incomplete-turn.test.ts
@@ -817,6 +817,75 @@ describe("runEmbeddedPiAgent incomplete-turn safety", () => {
     expectWarnMessageWith("empty response detected");
   });
 
+  it("retries empty Anthropic-compatible stop turns even when the provider is not Kimi", async () => {
+    mockedClassifyFailoverReason.mockReturnValue(null);
+    mockedResolveModelAsync.mockResolvedValue({
+      model: {
+        id: "claude-opus-4-7",
+        provider: "sub2api",
+        contextWindow: 200000,
+        api: "anthropic-messages",
+      },
+      error: null,
+      authStorage: {
+        setRuntimeApiKey: vi.fn(),
+      },
+      modelRegistry: {},
+    });
+    mockedRunEmbeddedAttempt.mockResolvedValueOnce(
+      makeAttemptResult({
+        assistantTexts: [],
+        lastAssistant: {
+          role: "assistant",
+          api: "anthropic-messages",
+          stopReason: "stop",
+          provider: "sub2api",
+          model: "claude-opus-4-7",
+          content: [],
+          usage: {
+            input: 2048,
+            output: 3100,
+            cacheRead: 0,
+            cacheWrite: 0,
+            totalTokens: 5148,
+          },
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    );
+    mockedRunEmbeddedAttempt.mockResolvedValueOnce(
+      makeAttemptResult({
+        assistantTexts: ["Visible Anthropic-compatible answer."],
+        lastAssistant: {
+          role: "assistant",
+          api: "anthropic-messages",
+          stopReason: "stop",
+          provider: "sub2api",
+          model: "claude-opus-4-7",
+          content: [{ type: "text", text: "Visible Anthropic-compatible answer." }],
+          usage: {
+            input: 2300,
+            output: 8,
+            cacheRead: 0,
+            cacheWrite: 0,
+            totalTokens: 2308,
+          },
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    );
+
+    await runEmbeddedPiAgent({
+      ...overflowBaseRunParams,
+      provider: "sub2api",
+      model: "claude-opus-4-7",
+      runId: "run-empty-anthropic-compatible-stop-continuation",
+    });
+
+    expect(mockedRunEmbeddedAttempt).toHaveBeenCalledTimes(2);
+    const secondCall = runAttemptCall(1);
+    expect(secondCall.prompt).toContain(EMPTY_RESPONSE_RETRY_INSTRUCTION);
+    expectWarnMessageWith("empty response detected");
+  });
+
   it("surfaces an error after exhausting empty-response retries", async () => {
     mockedClassifyFailoverReason.mockReturnValue(null);
     mockedRunEmbeddedAttempt.mockResolvedValue(
diff --git a/src/agents/pi-embedded-runner/run/incomplete-turn.ts b/src/agents/pi-embedded-runner/run/incomplete-turn.ts
@@ -131,7 +131,6 @@ const GEMINI_INCOMPLETE_TURN_MODEL_ID_PATTERN = /^gemini(?:[.-]|$)/;
 // Ollama native `/api/chat` can finish with only thinking/internal blocks when
 // constrained, but it should not inherit the stricter planning-only/ack prompts.
 const OLLAMA_INCOMPLETE_TURN_PROVIDER_ID_PATTERN = /^ollama(?:-|$)/;
-const KIMI_INCOMPLETE_TURN_PROVIDER_ID_PATTERN = /^kimi(?:-|$)/;
 const DEFAULT_PLANNING_ONLY_RETRY_LIMIT = 1;
 const STRICT_AGENTIC_PLANNING_ONLY_RETRY_LIMIT = 2;
 // Allow one immediate continuation plus one follow-up continuation before
@@ -620,14 +619,9 @@ function shouldApplyNonVisibleTurnRetryGuard(params: {
   if (shouldApplyPlanningOnlyRetryGuard(params)) {
     return true;
   }
-  if (normalizeLowercaseStringOrEmpty(params.modelApi ?? "") === "openai-completions") {
-    return true;
-  }
   if (
-    normalizeLowercaseStringOrEmpty(params.modelApi ?? "") === "anthropic-messages" &&
-    KIMI_INCOMPLETE_TURN_PROVIDER_ID_PATTERN.test(
-      normalizeLowercaseStringOrEmpty(params.provider ?? ""),
-    )
+    normalizeLowercaseStringOrEmpty(params.modelApi ?? "") === "openai-completions" ||
+    normalizeLowercaseStringOrEmpty(params.modelApi ?? "") === "anthropic-messages"
   ) {
     return true;
   }
