@librechat/agents 3.1.71-dev.0 → 3.1.71

package/src/tools/ToolNode.ts

@@ -36,7 +36,6 @@ import { executeHooks } from '@/hooks';
 import { Constants, GraphEvents, CODE_EXECUTION_TOOLS } from '@/common';
 import {
   buildReferenceKey,
-  annotateToolOutputWithReference,
   ToolOutputReferenceRegistry,
 } from '@/tools/toolOutputReferences';
 
@@ -429,21 +428,17 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
         const isError = toolMsg.status === 'error';
         if (isError) {
           /**
-           * Error ToolMessages bypass registration/annotation but must
-           * still carry the unresolved-refs hint so the LLM can
-           * self-correct when its reference key caused the failure.
+           * Error ToolMessages bypass registration but still stamp the
+           * unresolved-refs hint into `additional_kwargs` so the lazy
+           * annotation transform surfaces it to the LLM, letting the
+           * model self-correct when its reference key caused the
+           * failure. Persisted `content` stays clean.
            */
-          if (
-            unresolvedRefs.length > 0 &&
-            typeof toolMsg.content === 'string'
-          ) {
-            toolMsg.content = this.applyOutputReference(
-              runId,
-              toolMsg.content,
-              toolMsg.content,
-              undefined,
-              unresolvedRefs
-            );
+          if (unresolvedRefs.length > 0) {
+            toolMsg.additional_kwargs = {
+              ...toolMsg.additional_kwargs,
+              _unresolvedRefs: unresolvedRefs,
+            };
           }
           return toolMsg;
         }
@@ -454,35 +449,35 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
               rawContent,
               this.maxToolResultChars
             );
-            toolMsg.content = this.applyOutputReference(
+            toolMsg.content = llmContent;
+            const refMeta = this.recordOutputReference(
               runId,
-              llmContent,
               rawContent,
               refKey,
               unresolvedRefs
             );
+            if (refMeta != null) {
+              toolMsg.additional_kwargs = {
+                ...toolMsg.additional_kwargs,
+                ...refMeta,
+              };
+            }
           } else {
             /**
              * Non-string content (multi-part content blocks — text +
              * image). Known limitation: we cannot register under a
              * reference key because there's no canonical serialized
              * form. Warn once per tool per run when the caller
-             * intended to register.
-             *
-             * Still surface unresolved-ref warnings so the LLM gets
-             * the self-correction signal that the string and error
-             * paths already emit. Prepended as a leading text block
-             * to keep the original content ordering intact.
+             * intended to register. The unresolved-refs hint is still
+             * stamped as metadata; the lazy transform prepends a text
+             * block at request time so the LLM gets the self-correction
+             * signal.
              */
-            if (unresolvedRefs.length > 0 && Array.isArray(toolMsg.content)) {
-              const warningBlock = {
-                type: 'text',
-                text: `[unresolved refs: ${unresolvedRefs.join(', ')}]`,
+            if (unresolvedRefs.length > 0) {
+              toolMsg.additional_kwargs = {
+                ...toolMsg.additional_kwargs,
+                _unresolvedRefs: unresolvedRefs,
               };
-              toolMsg.content = [
-                warningBlock,
-                ...toolMsg.content,
-              ] as typeof toolMsg.content;
             }
             if (
               refKey != null &&
@@ -504,9 +499,8 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
         rawContent,
         this.maxToolResultChars
       );
-      const content = this.applyOutputReference(
+      const refMeta = this.recordOutputReference(
         runId,
-        truncated,
         rawContent,
         refKey,
         unresolvedRefs
@@ -514,8 +508,11 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
       return new ToolMessage({
         status: 'success',
         name: tool.name,
-        content,
+        content: truncated,
         tool_call_id: call.id!,
+        ...(refMeta != null && {
+          additional_kwargs: refMeta as Record<string, unknown>,
+        }),
       });
     } catch (_e: unknown) {
       const e = _e as Error;
@@ -561,64 +558,73 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
           });
         }
       }
-      let errorContent = `Error: ${e.message}\n Please fix your mistakes.`;
-      if (unresolvedRefs.length > 0) {
-        errorContent = this.applyOutputReference(
-          runId,
-          errorContent,
-          errorContent,
-          undefined,
-          unresolvedRefs
-        );
-      }
+      const errorContent = `Error: ${e.message}\n Please fix your mistakes.`;
+      const refMeta =
+        unresolvedRefs.length > 0
+          ? this.recordOutputReference(
+            runId,
+            errorContent,
+            undefined,
+            unresolvedRefs
+          )
+          : undefined;
       return new ToolMessage({
         status: 'error',
         content: errorContent,
         name: call.name,
         tool_call_id: call.id ?? '',
+        ...(refMeta != null && {
+          additional_kwargs: refMeta as Record<string, unknown>,
+        }),
       });
     }
   }
 
   /**
-   * Finalizes the LLM-visible content for a tool call and (when a
-   * `refKey` is provided) registers the full, raw output under that
-   * key.
+   * Registers the full, raw output under `refKey` (when provided) and
+   * builds the per-message ref metadata stamped onto the resulting
+   * `ToolMessage.additional_kwargs`. The metadata is read at LLM-
+   * request time by `annotateMessagesForLLM` to produce a transient
+   * annotated copy of the message — the persisted `content` itself
+   * stays clean.
    *
-   * @param llmContent  The content string the LLM will see. This is
-   *   the already-truncated, post-hook view; the annotation is
-   *   applied on top of it.
    * @param registryContent  The full, untruncated output to store in
    *   the registry so `{{tool<i>turn<n>}}` substitutions deliver the
    *   complete payload. Ignored when `refKey` is undefined.
    * @param refKey  Precomputed `tool<i>turn<n>` key, or undefined when
    *   the output is not to be registered (errors, disabled feature,
    *   unavailable batch/turn).
-   * @param unresolved  Placeholder keys that did not resolve; appended
-   *   as `[unresolved refs: …]` so the LLM can self-correct.
-   *
-   * `refKey` is passed in (rather than built from `this.currentTurn`)
-   * so parallel `invoke()` calls on the same ToolNode cannot race on
-   * the shared turn field.
+   * @param unresolved  Placeholder keys that did not resolve; surfaced
+   *   to the LLM lazily so it can self-correct.
+   * @returns A `ToolMessageRefMetadata` object when there is anything
+   *   to stamp, otherwise `undefined`.
    */
-  private applyOutputReference(
+  private recordOutputReference(
     runId: string | undefined,
-    llmContent: string,
     registryContent: string,
     refKey: string | undefined,
     unresolved: string[]
-  ): string {
+  ): t.ToolMessageRefMetadata | undefined {
     if (this.toolOutputRegistry != null && refKey != null) {
       this.toolOutputRegistry.set(runId, refKey, registryContent);
     }
-    /**
-     * `annotateToolOutputWithReference` handles both the ref-key and
-     * unresolved-refs cases together so JSON-object outputs stay
-     * parseable: unresolved refs land in an `_unresolved_refs` field
-     * instead of as a trailing text line that would break
-     * `JSON.parse` for downstream consumers.
-     */
-    return annotateToolOutputWithReference(llmContent, refKey, unresolved);
+    if (refKey == null && unresolved.length === 0) return undefined;
+    const meta: t.ToolMessageRefMetadata = {};
+    if (refKey != null) {
+      meta._refKey = refKey;
+      /**
+       * Stamp the registry scope alongside the key so the lazy
+       * annotation transform can look up the right bucket. Anonymous
+       * invocations get a synthetic per-batch scope (`\0anon-<n>`)
+       * that `attemptInvoke` cannot derive from
+       * `config.configurable.run_id` — without this, anonymous-run
+       * refs would silently fail registry lookup and the LLM would
+       * never see `[ref: …]` markers for outputs that were registered.
+       */
+      if (runId != null) meta._refScope = runId;
+    }
+    if (unresolved.length > 0) meta._unresolvedRefs = unresolved;
+    return meta;
   }
 
   /**
@@ -1054,25 +1060,30 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
         if (result.status === 'error') {
           contentString = `Error: ${result.errorMessage ?? 'Unknown error'}\n Please fix your mistakes.`;
           /**
-           * Error results bypass registration/annotation but must still
-           * carry the unresolved-refs hint so the LLM can self-correct
-           * when its reference key caused the failure.
+           * Error results bypass registration but stamp the
+           * unresolved-refs hint into `additional_kwargs` so the lazy
+           * annotation transform surfaces it to the LLM at request
+           * time, letting the model self-correct when its reference
+           * key caused the failure. Persisted `content` stays clean.
            */
           const unresolved = unresolvedByCallId.get(result.toolCallId) ?? [];
-          if (unresolved.length > 0) {
-            contentString = this.applyOutputReference(
-              registryRunId,
-              contentString,
-              contentString,
-              undefined,
-              unresolved
-            );
-          }
+          const errorRefMeta =
+            unresolved.length > 0
+              ? this.recordOutputReference(
+                registryRunId,
+                contentString,
+                undefined,
+                unresolved
+              )
+              : undefined;
           toolMessage = new ToolMessage({
             status: 'error',
             content: contentString,
             name: toolName,
             tool_call_id: result.toolCallId,
+            ...(errorRefMeta != null && {
+              additional_kwargs: errorRefMeta as Record<string, unknown>,
+            }),
           });
 
           if (hasFailureHook) {
@@ -1145,9 +1156,8 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
             turn != null
               ? buildReferenceKey(batchIndex, turn)
               : undefined;
-          contentString = this.applyOutputReference(
+          const successRefMeta = this.recordOutputReference(
             registryRunId,
-            contentString,
             registryRaw,
             refKey,
             unresolved
@@ -1159,6 +1169,9 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
             content: contentString,
             artifact: result.artifact,
             tool_call_id: result.toolCallId,
+            ...(successRefMeta != null && {
+              additional_kwargs: successRefMeta as Record<string, unknown>,
+            }),
           });
         }
 

package/src/tools/__tests__/BashExecutor.test.ts

@@ -27,6 +27,19 @@ describe('buildBashExecutionToolDescription', () => {
     expect(composed).toContain('{{tool<idx>turn<turn>}}');
   });
 
+  it('nudges the model toward heredoc when payloads may contain shell metacharacters', () => {
+    /**
+     * Real-world failure observed against ClickHouse + bash piping:
+     * the model emitted `echo '{{ref}}' | wc -c` and the substituted
+     * binary payload contained literal single quotes, breaking the
+     * shell. The model self-corrected to a heredoc on retry. Surface
+     * the heredoc pattern upfront so the round-trip isn't burned to
+     * rediscover it.
+     */
+    expect(BashToolOutputReferencesGuide).toContain('heredoc');
+    expect(BashToolOutputReferencesGuide).toContain('<< \'EOF\'');
+  });
+
   it('separates base and guide with a blank line', () => {
     const composed = buildBashExecutionToolDescription({
       enableToolOutputReferences: true,

package/src/tools/__tests__/ToolNode.outputReferences.test.ts

@@ -7,10 +7,27 @@ import type * as t from '@/types';
 import * as events from '@/utils/events';
 import { HookRegistry } from '@/hooks';
 import { ToolNode } from '../ToolNode';
-import {
-  ToolOutputReferenceRegistry,
-  TOOL_OUTPUT_REF_KEY,
-} from '../toolOutputReferences';
+import { ToolOutputReferenceRegistry } from '../toolOutputReferences';
+
+/**
+ * Reads the lazy ref-metadata stamped onto a `ToolMessage` by ToolNode.
+ * The metadata replaces the durable `[ref: …]` content mutation that the
+ * earlier eager-annotation design used; the LLM-facing annotation is
+ * applied at request time by `annotateMessagesForLLM` instead.
+ */
+function getRefKey(msg: ToolMessage): string | undefined {
+  return (msg.additional_kwargs as { _refKey?: string } | undefined)?._refKey;
+}
+function getRefScope(msg: ToolMessage): string | undefined {
+  return (msg.additional_kwargs as { _refScope?: string } | undefined)
+    ?._refScope;
+}
+function getUnresolvedRefs(msg: ToolMessage): string[] {
+  return (
+    (msg.additional_kwargs as { _unresolvedRefs?: string[] } | undefined)
+      ?._unresolvedRefs ?? []
+  );
+}
 
 /**
  * Captures the `command` arg each time the tool is invoked and returns
@@ -98,7 +115,7 @@ describe('ToolNode tool output references', () => {
   });
 
   describe('enabled', () => {
-    it('annotates string outputs with a [ref: …] prefix line', async () => {
+    it('keeps string outputs clean and stamps the ref key as metadata', async () => {
       const t1 = createEchoTool({
         capturedArgs: [],
         outputs: ['hello world'],
@@ -112,10 +129,19 @@ describe('ToolNode tool output references', () => {
         { id: 'c1', name: 'echo', command: 'run' },
       ]);
 
-      expect(msg.content).toBe('[ref: tool0turn0]\nhello world');
+      expect(msg.content).toBe('hello world');
+      expect(getRefKey(msg)).toBe('tool0turn0');
+      /**
+       * `_refScope` is what lets `annotateMessagesForLLM` recover the
+       * registry bucket at request time without re-deriving it from
+       * `config.configurable.run_id` (which fails for anonymous
+       * batches). For named runs it equals the run_id.
+       */
+      expect(getRefScope(msg)).toBe('test-run');
+      expect(getUnresolvedRefs(msg)).toEqual([]);
     });
 
-    it('injects _ref into JSON-object string outputs', async () => {
+    it('keeps JSON-object string outputs unmodified and stamps ref metadata', async () => {
       const t1 = createEchoTool({
         capturedArgs: [],
         outputs: ['{"a":1,"b":"x"}'],
@@ -130,11 +156,13 @@ describe('ToolNode tool output references', () => {
       ]);
 
       const parsed = JSON.parse(msg.content as string);
-      expect(parsed[TOOL_OUTPUT_REF_KEY]).toBe('tool0turn0');
       expect(parsed.a).toBe(1);
+      expect(parsed.b).toBe('x');
+      expect(parsed._ref).toBeUndefined();
+      expect(getRefKey(msg)).toBe('tool0turn0');
     });
 
-    it('uses the [ref: …] prefix for JSON array outputs', async () => {
+    it('keeps JSON array outputs unmodified and stamps ref metadata', async () => {
       const t1 = createEchoTool({ capturedArgs: [], outputs: ['[1,2,3]'] });
       const node = new ToolNode({
         tools: [t1],
@@ -145,7 +173,8 @@ describe('ToolNode tool output references', () => {
         { id: 'c1', name: 'echo', command: 'run' },
       ]);
 
-      expect(msg.content).toBe('[ref: tool0turn0]\n[1,2,3]');
+      expect(msg.content).toBe('[1,2,3]');
+      expect(getRefKey(msg)).toBe('tool0turn0');
     });
 
     it('registers the un-annotated output for piping into later calls', async () => {
@@ -192,9 +221,9 @@ describe('ToolNode tool output references', () => {
         { id: 'b3c1', name: 'echo', command: '{{tool0turn1}}' },
       ]);
 
-      expect(m0.content).toContain('[ref: tool0turn0]');
-      expect(m1.content).toContain('[ref: tool0turn1]');
-      expect(m2.content).toContain('[ref: tool0turn2]');
+      expect(getRefKey(m0)).toBe('tool0turn0');
+      expect(getRefKey(m1)).toBe('tool0turn1');
+      expect(getRefKey(m2)).toBe('tool0turn2');
       expect(capturedArgs[2]).toBe('two');
     });
 
@@ -221,8 +250,8 @@ describe('ToolNode tool output references', () => {
         { id: 'c2', name: 'beta', command: 'b' },
       ]);
 
-      expect(messages[0].content).toContain('[ref: tool0turn0]');
-      expect(messages[1].content).toContain('[ref: tool1turn0]');
+      expect(getRefKey(messages[0])).toBe('tool0turn0');
+      expect(getRefKey(messages[1])).toBe('tool1turn0');
     });
 
     it('reports unresolved placeholders after the output', async () => {
@@ -242,7 +271,8 @@ describe('ToolNode tool output references', () => {
       ]);
 
       expect(capturedArgs[0]).toBe('see {{tool9turn9}}');
-      expect(msg.content).toContain('[unresolved refs: tool9turn9]');
+      expect(msg.content).toBe('done');
+      expect(getUnresolvedRefs(msg)).toEqual(['tool9turn9']);
     });
 
     it('stores the raw untruncated output in the registry, independent of the LLM-visible truncation', async () => {
@@ -343,10 +373,10 @@ describe('ToolNode tool output references', () => {
         messages: ToolMessage[];
       }>;
 
-      expect(resA.messages[0].content).toContain('[ref: tool0turn0]');
-      expect(resA.messages[0].content).toContain('output-A');
-      expect(resB.messages[0].content).toContain('[ref: tool0turn1]');
-      expect(resB.messages[0].content).toContain('output-B');
+      expect(getRefKey(resA.messages[0])).toBe('tool0turn0');
+      expect(resA.messages[0].content).toBe('output-A');
+      expect(getRefKey(resB.messages[0])).toBe('tool0turn1');
+      expect(resB.messages[0].content).toBe('output-B');
 
       const registry = node._unsafeGetToolOutputRegistry()!;
       expect(registry.get('concurrent-run', 'tool0turn0')).toBe('output-A');
@@ -417,7 +447,7 @@ describe('ToolNode tool output references', () => {
         { id: 'c1', name: 'boom', command: 'x' },
       ]);
 
-      expect((msg.content as string).startsWith('[ref:')).toBe(false);
+      expect(getRefKey(msg)).toBeUndefined();
       expect(
         node._unsafeGetToolOutputRegistry()!.get('test-run', 'tool0turn0')
       ).toBeUndefined();
@@ -445,7 +475,8 @@ describe('ToolNode tool output references', () => {
       ]);
 
       expect(msg.content).toContain('Error: nope');
-      expect(msg.content).toContain('[unresolved refs: tool9turn9]');
+      expect(msg.content as string).not.toContain('[unresolved refs:');
+      expect(getUnresolvedRefs(msg)).toEqual(['tool9turn9']);
     });
 
     it('surfaces unresolved refs on tool-returned error ToolMessages', async () => {
@@ -473,8 +504,8 @@ describe('ToolNode tool output references', () => {
         { id: 'c1', name: 'errReturn', command: 'see {{tool9turn9}}' },
       ]);
 
-      expect(msg.content).toContain('handled failure');
-      expect(msg.content).toContain('[unresolved refs: tool9turn9]');
+      expect(msg.content).toBe('handled failure');
+      expect(getUnresolvedRefs(msg)).toEqual(['tool9turn9']);
     });
 
     it('isolates state between overlapping runs on the same ToolNode', async () => {
@@ -584,13 +615,26 @@ describe('ToolNode tool output references', () => {
         messages: ToolMessage[];
       }>;
 
-      // Each invocation produces its own annotated output — neither's
+      // Each invocation stamps its own ref metadata — neither's
       // registered tool0turn0 was clobbered by the other's sync-prefix
       // reset.
-      expect(resA.messages[0].content).toContain('[ref: tool0turn0]');
-      expect(resA.messages[0].content).toContain('out-A');
-      expect(resB.messages[0].content).toContain('[ref: tool0turn0]');
-      expect(resB.messages[0].content).toContain('out-B');
+      expect(getRefKey(resA.messages[0])).toBe('tool0turn0');
+      expect(resA.messages[0].content).toBe('out-A');
+      expect(getRefKey(resB.messages[0])).toBe('tool0turn0');
+      expect(resB.messages[0].content).toBe('out-B');
+
+      /**
+       * Each anonymous invocation stamps a distinct synthetic
+       * `_refScope` so the lazy annotation transform can later look
+       * up the right registry bucket — `config.configurable.run_id`
+       * is undefined for both calls and would collapse them to the
+       * same `\0anon` bucket without this stamping.
+       */
+      const scopeA = getRefScope(resA.messages[0]);
+      const scopeB = getRefScope(resB.messages[0]);
+      expect(scopeA).toMatch(/^\0anon-\d+$/);
+      expect(scopeB).toMatch(/^\0anon-\d+$/);
+      expect(scopeA).not.toBe(scopeB);
     });
 
     it('clears state on every batch when run_id is absent (anonymous caller)', async () => {
@@ -616,9 +660,7 @@ describe('ToolNode tool output references', () => {
       })) as { messages: ToolMessage[] };
 
       expect(capturedArgs[1]).toBe('echo {{tool0turn0}}');
-      expect(result.messages[0].content).toContain(
-        '[unresolved refs: tool0turn0]'
-      );
+      expect(getUnresolvedRefs(result.messages[0])).toEqual(['tool0turn0']);
     });
 
     it('lets two ToolNodes sharing a registry resolve each other\'s refs', async () => {
@@ -728,7 +770,7 @@ describe('ToolNode tool output references', () => {
       expect(JSON.parse(stepCompletedArgs[1]).command).toBe('echo STORED');
     });
 
-    it('prepends unresolved-refs warning to non-string ToolMessage content', async () => {
+    it('records unresolved refs as metadata on non-string ToolMessage content (content untouched)', async () => {
       const complexTool = tool(
         async () =>
           new ToolMessage({
@@ -760,12 +802,13 @@ describe('ToolNode tool output references', () => {
 
       expect(Array.isArray(msg.content)).toBe(true);
       const blocks = msg.content as Array<{ type: string; text?: string }>;
+      // Multi-part content is untouched at storage time — the lazy
+      // transform handles the unresolved-refs warning at request time.
+      expect(blocks).toHaveLength(2);
       expect(blocks[0].type).toBe('text');
-      expect(blocks[0].text).toContain('[unresolved refs: tool9turn9]');
-      // Original blocks follow the warning.
-      expect(blocks[1].type).toBe('text');
-      expect(blocks[1].text).toBe('data');
-      expect(blocks[2].type).toBe('image_url');
+      expect(blocks[0].text).toBe('data');
+      expect(blocks[1].type).toBe('image_url');
+      expect(getUnresolvedRefs(msg)).toEqual(['tool9turn9']);
     });
 
     it('resets the registry and turn counter when the runId changes', async () => {
@@ -800,10 +843,9 @@ describe('ToolNode tool output references', () => {
       )) as { messages: ToolMessage[] };
 
       expect(capturedArgs[1]).toBe('echo {{tool0turn0}}');
-      expect(resultB.messages[0].content).toContain('[ref: tool0turn0]');
-      expect(resultB.messages[0].content).toContain(
-        '[unresolved refs: tool0turn0]'
-      );
+      expect(resultB.messages[0].content).toBe('from-run-B');
+      expect(getRefKey(resultB.messages[0])).toBe('tool0turn0');
+      expect(getUnresolvedRefs(resultB.messages[0])).toEqual(['tool0turn0']);
     });
   });
 
@@ -836,7 +878,7 @@ describe('ToolNode tool output references', () => {
       }) as unknown as StructuredToolInterface;
     }
 
-    it('annotates the output the host returns', async () => {
+    it('keeps host-returned output clean and stamps the ref key as metadata', async () => {
       const node = new ToolNode({
         tools: [createSchemaStub('echo')],
         eventDrivenMode: true,
@@ -858,7 +900,8 @@ describe('ToolNode tool output references', () => {
         { configurable: { run_id: 'run-host' } }
       )) as { messages: ToolMessage[] };
 
-      expect(result.messages[0].content).toBe('[ref: tool0turn0]\nhost-output');
+      expect(result.messages[0].content).toBe('host-output');
+      expect(getRefKey(result.messages[0])).toBe('tool0turn0');
       expect(
         node._unsafeGetToolOutputRegistry()!.get('run-host', 'tool0turn0')
       ).toBe('host-output');
@@ -963,9 +1006,10 @@ describe('ToolNode tool output references', () => {
       })) as { messages: ToolMessage[] };
 
       expect(result.messages[0].content).toContain('Error: host failure');
-      expect(result.messages[0].content).toContain(
-        '[unresolved refs: tool9turn9]'
+      expect(result.messages[0].content as string).not.toContain(
+        '[unresolved refs:'
       );
+      expect(getUnresolvedRefs(result.messages[0])).toEqual(['tool9turn9']);
     });
 
     it('reports unresolved refs even when the host succeeds', async () => {
@@ -995,9 +1039,8 @@ describe('ToolNode tool output references', () => {
         ],
       })) as { messages: ToolMessage[] };
 
-      expect(result.messages[0].content).toContain(
-        '[unresolved refs: tool9turn9]'
-      );
+      expect(result.messages[0].content).toBe('done');
+      expect(getUnresolvedRefs(result.messages[0])).toEqual(['tool9turn9']);
     });
 
     it('registers the post-hook output when PostToolUse replaces it', async () => {
@@ -1035,9 +1078,8 @@ describe('ToolNode tool output references', () => {
         { configurable: { run_id: 'run-posthook' } }
       )) as { messages: ToolMessage[] };
 
-      expect(result.messages[0].content).toBe(
-        '[ref: tool0turn0]\nhooked-output'
-      );
+      expect(result.messages[0].content).toBe('hooked-output');
+      expect(getRefKey(result.messages[0])).toBe('tool0turn0');
       expect(
         node._unsafeGetToolOutputRegistry()!.get('run-posthook', 'tool0turn0')
       ).toBe('hooked-output');
@@ -1252,9 +1294,10 @@ describe('ToolNode tool output references', () => {
       // call attempts `{{tool0turn0}}` — which points at the direct
       // call running *in the same batch*. Correct behavior: the
       // placeholder stays unresolved (cross-batch only), and the
-      // event args received by the host must carry the literal
-      // template string plus the LLM-visible `[unresolved refs:…]`
-      // trailer.
+      // event args received by the host carry the literal template
+      // string. The unresolved-refs hint is stamped into the resulting
+      // ToolMessage's `additional_kwargs._unresolvedRefs` so the lazy
+      // annotation transform surfaces it to the LLM at request time.
       await node.invoke(
         {
           messages: [