Skip to content

The open-source agent harness - the runtime layer that turns an LLM into a working agent.

feat: enhance large tool response handling with detailed size information (#1131)

Sujai Kumar Gupta731880664491
diff --git a/.changeset/large-tool-response-size.md b/.changeset/large-tool-response-size.md
new file mode 100644
index 000000000..67fdc615c
--- /dev/null
+++ b/.changeset/large-tool-response-size.md
@@ -0,0 +1,5 @@
+---
+'@truefoundry/trueforge-core': patch
+---
+
+The sandbox file-dump message for a blocked large tool response now states its exact size (`Content too large (50,000 characters). Result saved to: …`) and prompts the model to plan reads against that total, instead of guessing with small incremental reads.
diff --git a/packages/trueforge-core/src/core/capabilities/builtins/LargeToolResponse.ts b/packages/trueforge-core/src/core/capabilities/builtins/LargeToolResponse.ts
index bdfcc88ea..fb6cd0774 100644
--- a/packages/trueforge-core/src/core/capabilities/builtins/LargeToolResponse.ts
+++ b/packages/trueforge-core/src/core/capabilities/builtins/LargeToolResponse.ts
@@ -1,6 +1,6 @@
import dedent from 'dedent';
import type { Logger } from 'winston';
-import { estimateTokensForString } from '../../llm/usage';
+import { describeContentSize, estimateTokensForString } from '../../llm/usage';
import type { ToolCallResult } from '../../mcp/executeToolCalls';
import type { HarnessSandbox, SandboxInfo } from '../../sandbox/Sandbox';
import { createSandboxLargeToolResponseGuidance, SANDBOX_SCHEMA_INFER_TAG } from '../../sandbox/Sandbox';
@@ -86,11 +86,12 @@ export class LargeToolResponseProcessor implements ToolResponseProcessor {
.join('\n')}`;
}
- private createFileDumpMessage(params: { filePath: string; preview: string }): string {
- const { filePath, preview } = params;
+ private createFileDumpMessage(params: { filePath: string; preview: string; size: string }): string {
+ const { filePath, preview, size } = params;
const steps: string[] = [
dedent`Use sandbox to read parts of the file or extract data.
- Use ${SANDBOX_SCHEMA_INFER_TAG} to undertand the schema.`,
+ Use ${SANDBOX_SCHEMA_INFER_TAG} to understand the schema.
+ The file's total size is given above. Plan reads against it, for example extract fields with a script or read sized ranges, instead of probing with small incremental reads.`,
];
if (this.dynamicSubAgentsPresent) {
@@ -103,7 +104,7 @@ export class LargeToolResponseProcessor implements ToolResponseProcessor {
.map((step, index) => `${String(index + 1)}. ${step}`)
.join('\n')}`;
- return `Content too large. Result saved to: ${filePath}.\n\n${guidance}\n\nPreview (first and last ${String(this.previewNumberOfCharacters)} chars):\n${preview}`;
+ return `Content too large (${size}). Result saved to: ${filePath}.\n\n${guidance}\n\nPreview (first and last ${String(this.previewNumberOfCharacters)} chars):\n${preview}`;
}
private buildResultsWithInfo(results: ToolCallResult[], sandbox: HarnessSandbox | undefined): ToolResultWithInfo[] {
@@ -161,7 +162,11 @@ export class LargeToolResponseProcessor implements ToolResponseProcessor {
sourceTool: entry.toolCallResult.info,
});
const preview = createContentPreview(entry.originalContent, this.previewNumberOfCharacters);
- entry.toolCallResult.message.content = this.createFileDumpMessage({ filePath: stored.filePath, preview });
+ entry.toolCallResult.message.content = this.createFileDumpMessage({
+ filePath: stored.filePath,
+ preview,
+ size: describeContentSize(entry.originalContent),
+ });
entry.toolCallResult.isStructuredContent = false;
return stored.sandboxCreated;
} catch (error) {
diff --git a/packages/trueforge-core/src/core/llm/usage.ts b/packages/trueforge-core/src/core/llm/usage.ts
index 079845015..86e6072c0 100644
--- a/packages/trueforge-core/src/core/llm/usage.ts
+++ b/packages/trueforge-core/src/core/llm/usage.ts
@@ -5,6 +5,11 @@ export function estimateTokensForString(s: string): number {
return s.length / 4;
}
+/** Size label for prompts replacing full content, so the model can size reads without guessing. */
+export function describeContentSize(content: string): string {
+ return `${content.length.toLocaleString('en-US')} characters`;
+}
+
function sumOptional(a: number | undefined, b: number | undefined): number | undefined {
if (a === undefined && b === undefined) {
return undefined;
diff --git a/packages/trueforge-core/tests/core/usage.test.ts b/packages/trueforge-core/tests/core/usage.test.ts
index 39e911637..f1d542219 100644
--- a/packages/trueforge-core/tests/core/usage.test.ts
+++ b/packages/trueforge-core/tests/core/usage.test.ts
@@ -1,5 +1,5 @@
import { CompletionUsageSchema, getEmptyUsage } from '../../src/core/llm/LLMTypes';
-import { mergeUsage } from '../../src/core/llm/usage';
+import { describeContentSize, mergeUsage } from '../../src/core/llm/usage';
import {
addAgentThreadMetrics,
createEmptyAgentThreadMetrics,
@@ -33,6 +33,13 @@ describe('CompletionUsage', () => {
});
});
+describe('describeContentSize', () => {
+ it('formats characters with thousands separators', () => {
+ expect(describeContentSize('x'.repeat(50_000))).toBe('50,000 characters');
+ expect(describeContentSize('')).toBe('0 characters');
+ });
+});
+
describe('mergeUsage', () => {
it('sums cost across model calls', () => {
const merged = mergeUsage(