fix(discord): suppress internal agent failure traces

2026-07-21 10:16:44 +00:00 · 2026-06-03 08:03:57 +05:30
parent a9f014e9df
commit a8bf14da84
8 changed files with 170 additions and 88 deletions
@@ -2504,7 +2504,7 @@ describe("processDiscordMessage draft streaming", () => {
    expect(deliverDiscordReply).not.toHaveBeenCalled();
  });

-  it("delivers tool warning finals when no recovered reply is available", async () => {
+  it("suppresses pure tool warning finals when no recovered reply is available", async () => {
    const draftStream = createMockDraftStreamForTest();
    dispatchInboundMessage.mockImplementationOnce(async (params?: DispatchInboundParams) => {
      await params?.dispatcher.sendFinalReply(createNonTerminalToolWarningPayload());
@@ -2519,18 +2519,10 @@ describe("processDiscordMessage draft streaming", () => {

    expect(editMessageDiscord).not.toHaveBeenCalled();
    expect(draftStream.clear).toHaveBeenCalledTimes(1);
-    expect(deliverDiscordReply).toHaveBeenCalledTimes(1);
-    expect(firstMockArg(deliverDiscordReply, "deliverDiscordReply")).toMatchObject({
-      replies: [
-        {
-          text: "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
-          isError: true,
-        },
-      ],
-    });
+    expect(deliverDiscordReply).not.toHaveBeenCalled();
  });

-  it("delivers tool warning finals when the recovered reply fails to send", async () => {
+  it("suppresses tool warning finals when the recovered reply fails to send", async () => {
    deliverDiscordReply.mockRejectedValueOnce(new Error("send failed"));
    dispatchInboundMessage.mockImplementationOnce(async (params?: DispatchInboundParams) => {
      await params?.dispatcher.sendFinalReply({ text: "delivery failed" });
@@ -2549,21 +2541,13 @@ describe("processDiscordMessage draft streaming", () => {

    await runProcessDiscordMessage(ctx);

-    expect(deliverDiscordReply).toHaveBeenCalledTimes(2);
+    expect(deliverDiscordReply).toHaveBeenCalledTimes(1);
    expect(firstMockArg(deliverDiscordReply, "deliverDiscordReply")).toMatchObject({
      replies: [{ text: "delivery failed" }],
    });
-    expect(deliverDiscordReply.mock.calls[1]?.[0]).toMatchObject({
-      replies: [
-        {
-          text: "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
-          isError: true,
-        },
-      ],
-    });
  });

-  it("keeps mutating tool warning finals after successful-looking replies", async () => {
+  it("suppresses mutating tool warning finals after successful-looking replies", async () => {
    const draftStream = createMockDraftStreamForTest();
    dispatchInboundMessage.mockImplementationOnce(async (params?: DispatchInboundParams) => {
      await params?.dispatcher.sendFinalReply({ text: "Done." });
@@ -2582,15 +2566,7 @@ describe("processDiscordMessage draft streaming", () => {

    expectPreviewEditContent("Done.");
    expect(draftStream.clear).not.toHaveBeenCalled();
-    expect(deliverDiscordReply).toHaveBeenCalledTimes(1);
-    expect(firstMockArg(deliverDiscordReply, "deliverDiscordReply")).toMatchObject({
-      replies: [
-        {
-          text: "⚠️ 🛠️ `write file (agent)` failed",
-          isError: true,
-        },
-      ],
-    });
+    expect(deliverDiscordReply).not.toHaveBeenCalled();
  });

  it("suppresses reasoning payload delivery to Discord", async () => {
@@ -66,6 +66,7 @@ import { createDiscordDraftPreviewController } from "./message-handler.draft-pre
 import type { DiscordMessagePreflightContext } from "./message-handler.preflight.js";
 import { resolveForwardedMediaList, resolveMediaList } from "./message-utils.js";
 import { deliverDiscordReply } from "./reply-delivery.js";
+import { sanitizeDiscordFrontChannelReplyPayloads } from "./reply-safety.js";
 import { createDiscordReplyTypingFeedback } from "./reply-typing-feedback.js";
 import {
  DISCORD_ATTACHMENT_IDLE_TIMEOUT_MS,
@@ -111,7 +112,10 @@ function isFallbackOnlyToolWarningFinal(payload: ReplyPayload): boolean {
  return !resolveSendableOutboundReplyParts(payload).hasMedia;
 }

-type DiscordReplySkipReason = "aborted before delivery" | "reasoning payload";
+type DiscordReplySkipReason =
+  | "aborted before delivery"
+  | "reasoning payload"
+  | "internal-only payload";

 export function formatDiscordReplySkip(params: {
  kind: "tool" | "block" | "final";
@@ -669,10 +673,24 @@ async function processDiscordMessageInner(
          })
        : payload.text;
    const effectivePayload = finalText !== payload.text ? { ...payload, text: finalText } : payload;
+    const [deliverablePayload] = sanitizeDiscordFrontChannelReplyPayloads([effectivePayload], {
+      kind: info.kind,
+    });
+    if (!deliverablePayload) {
+      logVerbose(
+        formatDiscordReplySkip({
+          kind: info.kind,
+          reason: "internal-only payload",
+          target: deliverTarget,
+          sessionKey: ctxPayload.SessionKey,
+        }),
+      );
+      return { visibleReplySent: false };
+    }
    const draftStream = draftPreview.draftStream;
    if (draftStream && draftPreview.isProgressMode && info.kind === "block") {
-      const reply = resolveSendableOutboundReplyParts(effectivePayload);
-      if (!reply.hasMedia && !payload.isError) {
+      const reply = resolveSendableOutboundReplyParts(deliverablePayload);
+      if (!reply.hasMedia && !deliverablePayload.isError) {
        return { visibleReplySent: false };
      }
    }
@@ -680,22 +698,22 @@ async function processDiscordMessageInner(
      draftStream &&
      isFinal &&
      (!draftPreview.isProgressMode || draftPreview.hasProgressDraftStarted) &&
-      !payload.isError;
+      !deliverablePayload.isError;
    if (shouldFinalizeDraftPreview) {
-      const reply = resolveSendableOutboundReplyParts(effectivePayload);
+      const reply = resolveSendableOutboundReplyParts(deliverablePayload);
      const hasMedia = reply.hasMedia;
-      const ttsSupplement = getReplyPayloadTtsSupplement(effectivePayload);
-      const previewSourceText = finalText ?? ttsSupplement?.spokenText;
+      const ttsSupplement = getReplyPayloadTtsSupplement(deliverablePayload);
+      const previewSourceText = deliverablePayload.text ?? ttsSupplement?.spokenText;
      const previewFinalText = draftPreview.resolvePreviewFinalText(previewSourceText);
      const previewReplyToId = replyReference.peek();
      const hasExplicitReplyDirective =
-        Boolean(effectivePayload.replyToTag || effectivePayload.replyToCurrent) ||
+        Boolean(deliverablePayload.replyToTag || deliverablePayload.replyToCurrent) ||
        (typeof previewSourceText === "string" &&
          /\[\[\s*reply_to(?:_current|\s*:)/i.test(previewSourceText));

      const result = await deliverWithFinalizableLivePreviewAdapter({
        kind: info.kind,
-        payload: effectivePayload,
+        payload: deliverablePayload,
        adapter: defineFinalizableLivePreviewAdapter({
          draft: {
            flush: () => draftPreview.flush(),
@@ -710,7 +728,7 @@ async function processDiscordMessageInner(
              (hasMedia && !ttsSupplement) ||
              typeof previewFinalText !== "string" ||
              hasExplicitReplyDirective ||
-              payload.isError
+              deliverablePayload.isError
            ) {
              return undefined;
            }
@@ -747,7 +765,7 @@ async function processDiscordMessageInner(
            replyReference.markSent();
          },
          buildSupplementalPayload: () =>
-            ttsSupplement ? buildTtsSupplementMediaPayload(effectivePayload) : undefined,
+            ttsSupplement ? buildTtsSupplementMediaPayload(deliverablePayload) : undefined,
          deliverSupplemental: async (supplementalPayload) => {
            if (isProcessAborted(abortSignal)) {
              return false;
@@ -794,9 +812,9 @@ async function processDiscordMessageInner(
          const fallbackPayload =
            ttsSupplement &&
            ttsSupplement.visibleTextAlreadyDelivered !== true &&
-            !effectivePayload.text?.trim()
-              ? { ...effectivePayload, text: ttsSupplement.spokenText }
-              : effectivePayload;
+            !deliverablePayload.text?.trim()
+              ? { ...deliverablePayload, text: ttsSupplement.spokenText }
+              : deliverablePayload;
          const replyToId = replyReference.use();
          notifyFinalReplyStart();
          await deliverDiscordReply({
@@ -849,7 +867,7 @@ async function processDiscordMessageInner(
    }
    await deliverDiscordReply({
      cfg,
-      replies: [effectivePayload],
+      replies: [deliverablePayload],
      target: deliverTarget,
      token,
      accountId,
@@ -867,7 +885,7 @@ async function processDiscordMessageInner(
      kind: info.kind,
    });
    replyReference.markSent();
-    if (isFinal && payload.isError !== true) {
+    if (isFinal && deliverablePayload.isError !== true) {
      markUserFacingFinalDelivered();
    }
    return { visibleReplySent: true };
@@ -183,6 +183,7 @@ describe("deliverDiscordReply", () => {
          text: [
            "📊 Session Status: current",
            "🛠️ run git status",
+            "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
            "🛠️ `gh pr view`",
            "🛠️ `docker compose up`",
            "🛠️ elevated · `cd /tmp && pnpm test`",
@@ -204,6 +205,26 @@ describe("deliverDiscordReply", () => {
    expect(firstDeliverParams().payloads).toEqual([{ text: "Visible reply." }]);
  });

+  it("drops pure internal tool failure warnings at the final Discord send boundary", async () => {
+    await deliverDiscordReply({
+      replies: [
+        {
+          text: "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
+          isError: true,
+        },
+      ],
+      target: "channel:101",
+      token: "token",
+      accountId: "default",
+      runtime,
+      cfg,
+      textLimit: 2000,
+      kind: "final",
+    });
+
+    expect(sendDurableMessageBatchMock).not.toHaveBeenCalled();
+  });
+
  it("strips serialized tool call blocks at the final Discord send boundary", async () => {
    await deliverDiscordReply({
      replies: [
@@ -243,7 +264,7 @@ describe("deliverDiscordReply", () => {
    await deliverDiscordReply({
      replies: [
        {
-          text: "commentary: calling tool\nanalysis: inspect private state",
+          text: "tool_call: calling tool\ntool_result: inspect private state",
          mediaUrl: "https://example.com/result.png",
        },
      ],
@@ -283,7 +304,7 @@ describe("deliverDiscordReply", () => {
    await deliverDiscordReply({
      replies: [
        {
-          text: "analysis: internal only",
+          text: "tool_result: internal only",
          channelData,
        },
      ],
@@ -313,7 +334,7 @@ describe("deliverDiscordReply", () => {
    await deliverDiscordReply({
      replies: [
        {
-          text: "commentary: hidden",
+          text: "function_call: hidden",
          presentation,
        },
      ],
@@ -3,13 +3,6 @@ import { resolveSendableOutboundReplyParts } from "openclaw/plugin-sdk/reply-pay
 import { sanitizeAssistantVisibleText } from "openclaw/plugin-sdk/text-chunking";
 import { stripPlainTextToolCallBlocks } from "openclaw/plugin-sdk/tool-payload";

-const DISCORD_INTERNAL_TRACE_LINE_RE =
-  /^(?:>\s*)?(?:📊|🛠️|📖|📝|🔍|🔎|⚙️)\s*(?:Session Status|Exec|Read|Edit|Write|Patch|Search|Open|Click|Find|Screenshot|Update Plan|Tool Call|Tool Result|Function Call|Shell|Command)\s*:/i;
-const DISCORD_INTERNAL_COMPACT_COMMAND_TRACE_LINE_RE =
-  /^(?:>\s*)?🛠️\s*(?:(?:(?:elevated|pty)\b\s*(?:·|,)\s*)+)?(?:`{1,2}\s*\S|(?:run|check|fetch|pull|push|view|show|list|switch|create|merge|rebase|stage|restore|reset|stash|search|find|print|copy|move|remove|install|start|cd|git|pnpm|npm|yarn|bun|node|python|python3|bash|sh)\b)/i;
-const DISCORD_INTERNAL_CHANNEL_LINE_RE =
-  /^(?:>\s*)?(?:analysis|commentary|tool[-_ ]?call|tool[-_ ]?result|function[-_ ]?call|thinking|reasoning)\s*[:=]/i;
-
 function hasNonEmptyRecord(value: unknown): value is Record<string, unknown> {
  return Boolean(
    value && typeof value === "object" && !Array.isArray(value) && Object.keys(value).length > 0,
@@ -36,30 +29,6 @@ function hasNonTextReplyPayloadContent(payload: ReplyPayload): boolean {
  );
 }

-function stripDiscordInternalTraceLines(text: string): string {
-  let inFence = false;
-  const kept: string[] = [];
-  for (const line of text.split(/\r?\n/)) {
-    if (/^\s*```/.test(line)) {
-      inFence = !inFence;
-      kept.push(line);
-      continue;
-    }
-    if (!inFence) {
-      const trimmed = line.trim();
-      if (
-        DISCORD_INTERNAL_TRACE_LINE_RE.test(trimmed) ||
-        DISCORD_INTERNAL_COMPACT_COMMAND_TRACE_LINE_RE.test(trimmed) ||
-        DISCORD_INTERNAL_CHANNEL_LINE_RE.test(trimmed)
-      ) {
-        continue;
-      }
-    }
-    kept.push(line);
-  }
-  return kept.join("\n");
-}
-
 function collapseExcessBlankLines(text: string): string {
  return text.replace(/[ \t]+\n/g, "\n").replace(/\n{3,}/g, "\n\n");
 }
@@ -68,8 +37,7 @@ export function sanitizeDiscordFrontChannelText(text: string): string {
  const withoutToolCallBlocks = stripPlainTextToolCallBlocks(text);
  const withoutAssistantScaffolding = sanitizeAssistantVisibleText(withoutToolCallBlocks);
  const withoutResidualToolCallBlocks = stripPlainTextToolCallBlocks(withoutAssistantScaffolding);
-  const withoutTraceLines = stripDiscordInternalTraceLines(withoutResidualToolCallBlocks);
-  return collapseExcessBlankLines(withoutTraceLines).trim();
+  return collapseExcessBlankLines(withoutResidualToolCallBlocks).trim();
 }

 export function sanitizeDiscordFrontChannelReplyPayloads(
@@ -82,7 +50,7 @@ export function sanitizeDiscordFrontChannelReplyPayloads(
    const safeText =
      typeof payload.text === "string"
        ? preserveVerboseToolProgress
-          ? collapseExcessBlankLines(sanitizeAssistantVisibleText(payload.text)).trim()
+          ? collapseExcessBlankLines(payload.text).trim()
          : sanitizeDiscordFrontChannelText(payload.text)
        : payload.text;
    const nextPayload =
@@ -304,6 +304,37 @@ describe("sanitizeUserFacingText", () => {
    expect(sanitizeUserFacingText(input)).toBe("Before\n\nAfter");
  });

+  it("strips internal tool trace warning lines from error-context delivery text", () => {
+    const input = [
+      "Visible intro.",
+      "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
+      "🛠️ run git status",
+      "📖 Read: lines 1-40 from secret.md",
+      "Visible outro.",
+    ].join("\n");
+
+    expect(sanitizeUserFacingText(input, { errorContext: true })).toBe(
+      "Visible intro.\nVisible outro.",
+    );
+  });
+
+  it("preserves explicit tool progress outside error-context delivery text", () => {
+    const input = "🛠️ Exec: echo queued-progress";
+
+    expect(sanitizeUserFacingText(input)).toBe(input);
+  });
+
+  it("preserves internal trace examples inside fenced code", () => {
+    const input = [
+      "Example:",
+      "```",
+      "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
+      "```",
+    ].join("\n");
+
+    expect(sanitizeUserFacingText(input)).toBe(input);
+  });
+
  it("strips plural XML function-call wrappers before user-facing delivery", () => {
    const input = [
      "Before",
@@ -1,3 +1,7 @@
+import {
+  normalizeLowercaseStringOrEmpty,
+  normalizeOptionalLowercaseString,
+} from "../../../packages/normalization-core/src/string-coerce.js";
 import { stripPlainTextToolCallBlocks } from "../../../packages/tool-call-repair/src/index.js";
 import { stripInboundMetadata } from "../../auto-reply/reply/strip-inbound-meta.js";
 import {
@@ -11,10 +15,7 @@ import {
 } from "../../shared/assistant-error-format.js";
 import { coerceChatContentText } from "../../shared/chat-content.js";
 import {
-  normalizeLowercaseStringOrEmpty,
-  normalizeOptionalLowercaseString,
-} from "../../../packages/normalization-core/src/string-coerce.js";
-import {
+  stripAssistantInternalTraceLines,
  stripLegacyBracketToolCallBlocks,
  stripMinimaxToolCallXml,
  stripToolCallXmlTags,
@@ -414,8 +415,11 @@ export function sanitizeUserFacingText(text: unknown, opts?: { errorContext?: bo
  // It is internal scaffolding, so drop standalone placeholder lines before delivery
  // while preserving ordinary inline mentions a user may be discussing.
  const withoutPlaceholder = stripToolCallsOmittedPlaceholderLines(withoutToolCallXml);
+  const withoutInternalTraceLines = errorContext
+    ? stripAssistantInternalTraceLines(withoutPlaceholder)
+    : withoutPlaceholder;
  const withoutToolCallBlocks = stripPlainTextToolCallBlocks(
-    stripLegacyBracketToolCallBlocks(withoutPlaceholder),
+    stripLegacyBracketToolCallBlocks(withoutInternalTraceLines),
  );
  const trimmed = withoutToolCallBlocks.trim();
  if (!trimmed) {
@@ -830,6 +830,34 @@ describe("sanitizeAssistantVisibleText", () => {
    expect(sanitizeAssistantVisibleText(input)).toBe("Visible answer");
  });

+  it("strips internal tool trace warning lines on the delivery path", () => {
+    const input = [
+      "Visible intro.",
+      "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
+      "🛠️ run git status",
+      "Visible outro.",
+    ].join("\n");
+
+    expect(sanitizeAssistantVisibleText(input)).toBe("Visible intro.\nVisible outro.");
+  });
+
+  it("preserves internal tool trace examples inside fenced code", () => {
+    const input = [
+      "Example:",
+      "```",
+      "⚠️ 🛠️ `run openclaw definitely-not-a-real-subcommand (agent)` failed",
+      "```",
+    ].join("\n");
+
+    expect(sanitizeAssistantVisibleText(input)).toBe(input);
+  });
+
+  it("preserves ordinary analysis headings", () => {
+    const input = ["Analysis:", "This is user-visible reasoning about the result."].join("\n");
+
+    expect(sanitizeAssistantVisibleText(input)).toBe(input);
+  });
+
  it("drops malformed reasoning before orphan close tags when final text follows", () => {
    expect(sanitizeAssistantVisibleText("private chain of thought </think> Visible answer")).toBe(
      "Visible answer",
@@ -11,6 +11,14 @@ import {
 const MEMORY_TAG_RE = /<\s*(\/?)\s*relevant[-_]memories\b[^<>]*>/gi;
 const MEMORY_TAG_QUICK_RE = /<\s*\/?\s*relevant[-_]memories\b/i;
 const LEGACY_BRACKET_TOOL_BLOCK_QUICK_RE = /\[\s*\/?\s*TOOL_(?:CALL|RESULT)\s*\]/i;
+const INTERNAL_TRACE_LINE_QUICK_RE =
+  /(?:📊|🛠️|📖|📝|🔍|🔎|⚙️|tool[-_ ]?call|tool[-_ ]?result|function[-_ ]?call)/i;
+const INTERNAL_TRACE_LINE_RE =
+  /^(?:>\s*)?(?:⚠️\s*)?(?:📊|🛠️|📖|📝|🔍|🔎|⚙️)\s*(?:Session Status|Exec|Read|Edit|Write|Patch|Search|Open|Click|Find|Screenshot|Update Plan|Tool Call|Tool Result|Function Call|Shell|Command)\s*:/i;
+const INTERNAL_COMPACT_COMMAND_TRACE_LINE_RE =
+  /^(?:>\s*)?(?:⚠️\s*)?🛠️\s*(?:(?:(?:elevated|pty)\b\s*(?:·|,)\s*)+)?(?:`{1,2}\s*\S|(?:run|check|fetch|pull|push|view|show|list|switch|create|merge|rebase|stage|restore|reset|stash|search|find|print|copy|move|remove|install|start|cd|git|pnpm|npm|yarn|bun|node|python|python3|bash|sh)\b)/i;
+const INTERNAL_CHANNEL_TRACE_LINE_RE =
+  /^(?:>\s*)?(?:tool[-_ ]?call|tool[-_ ]?result|function[-_ ]?call)\s*[:=]/i;

 /**
 * Strip XML-style tool call tags that models sometimes emit as plain text.
@@ -760,6 +768,33 @@ function stripRelevantMemoriesTags(text: string): string {
  return result;
 }

+export function stripAssistantInternalTraceLines(text: string): string {
+  if (!text || !INTERNAL_TRACE_LINE_QUICK_RE.test(text)) {
+    return text;
+  }
+
+  const codeRegions = findCodeRegions(text);
+  let result = "";
+  let lineStart = 0;
+  while (lineStart < text.length) {
+    const newlineIndex = text.indexOf("\n", lineStart);
+    const lineEnd = newlineIndex === -1 ? text.length : newlineIndex + 1;
+    const rawLine = text.slice(lineStart, lineEnd);
+    const line = rawLine.endsWith("\n") ? rawLine.slice(0, -1).replace(/\r$/, "") : rawLine;
+    const trimmed = line.trim();
+    const shouldStrip =
+      !isInsideCode(lineStart, codeRegions) &&
+      (INTERNAL_TRACE_LINE_RE.test(trimmed) ||
+        INTERNAL_COMPACT_COMMAND_TRACE_LINE_RE.test(trimmed) ||
+        INTERNAL_CHANNEL_TRACE_LINE_RE.test(trimmed));
+    if (!shouldStrip) {
+      result += rawLine;
+    }
+    lineStart = lineEnd;
+  }
+  return result;
+}
+
 export type AssistantVisibleTextSanitizerProfile = "delivery" | "history" | "internal-scaffolding";

 type AssistantVisibleTextPipelineOptions = {
@@ -833,6 +868,7 @@ function applyAssistantVisibleTextStagePipeline(
      stripFunctionCallsXmlPayloads: options.stripFunctionCallsXmlPayloads,
      stripFunctionResponseAfterPluralToolCalls: options.stripFunctionResponseAfterPluralToolCalls,
    });
+    cleaned = stripAssistantInternalTraceLines(cleaned);
    cleaned = stripLegacyBracketToolCallBlocks(cleaned);
    cleaned = stripPlainTextToolCallBlocks(cleaned);
    if (!options.preserveDowngradedToolText) {