MCP response trimming replaces the largest values whole (often the top result) and ignores topN; clickstack_table has no limit, having or default order #3542

Summary

When a tool result is over the size budget, trimToolResponse (packages/api/src/utils/trimToolResponse.ts) keeps the first items of an array and replaces any value that is still too big with {"__hdx_trimmed": true, "originalSize": N}. It never shortens a string. Because the budget is split equally across keys, the biggest values go first, and in pattern and log results the biggest values are usually the most important ones: long stack traces.

For clickstack_table, the tool has no limit, no having and no default ORDER BY, so a large grouped result is cut to its first rows in whatever order ClickHouse returned them.

Environment

HyperDX 2-nightly image (late Oct 2026) with ClickHouse 26.10.3. Code references checked on main at d08b2fa (2026-11-03, 2.43.1).

Steps to reproduce

1. clickstack_event_patterns with topN: 30 on a service whose top errors are multi-line stack traces.
2. The response has 10 patterns, and the top ones look like {"rank":1,"pattern":{"__hdx_trimmed":true,"originalSize":7443},"whereSnippet":{"__hdx_trimmed":true,…},"samples":[{"__hdx_trimmed":true,…}]}.

Self-contained (repro.mts below; runs the real trimToolResponse from main on an event_patterns-shaped object with 3 long and 27 short patterns):

isTrimmed=true; topN requested=30, patterns returned=10
rank 1: pattern={"__hdx_trimmed":true,"originalSize":4736} whereSnippet={"__hdx_trimmed":true,…} samples[0]={"__hdx_trimmed":true,…}
rank 2: … same
rank 3: … same
rank 4: pattern="Short INFO line 0 took <*> ms" (complete)

3. clickstack_table with a groupBy that yields a few thousand groups and no orderBy: the response holds the first rows of an unordered result (one run: 2,638 rows in, about 75 out), plus "Result was trimmed".

Expected vs actual

Impact

In 4 agent investigations on production-size data: the most frequent error pattern of one service (99.5% of its ERROR lines) came back as a marker in 3 of 4 calls; topN 15–30 returned 10 every time; 66 KB and 81 KB responses still overflowed the client; and table calls returned 89–12,194 unordered groups that were then cut.

Proposed fix

1. In trimToolResponse, shorten long strings (e.g. to 500–2,000 chars with a suffix saying how much was cut) before replacing values or dropping array items.
2. For pattern results, drop samples (keep 1) before dropping patterns, and keep the requested topN when it fits after that.
3. Add limit (default about 50) and having to clickstack_table, as dashboard tiles already have (packages/api/src/mcp/tools/dashboards/schemas.ts:839, :947), and default ORDER BY to the first select descending when groupBy is set.
4. Say in the note what was cut (e.g. "20 patterns dropped, 3 strings shortened").

Related

• feat(mcp): improve MCP tool quality — error hints, shared helpers, better messages #2337: shared MCP helpers and error hints

// Repro for trimToolResponse (packages/api/src/utils/trimToolResponse.ts) on an
// event_patterns-shaped result: long values are replaced, not shortened, and the
// array is cut to 10 items regardless of topN.
// Run from <hyperdx>/packages/api:  npx tsx <this file> ./src/utils/trimToolResponse.ts
import { pathToFileURL } from 'node:url';
const { trimToolResponse } = await import(pathToFileURL(process.argv[2]).href);

const stack = (n: number) => `java.lang.NullPointerException: Cannot invoke "Catalog.items()" (case ${n})\n` +
  Array.from({ length: 90 }, (_, i) => `\tat com.example.svc.Layer${i}.call(Layer${i}.java:${i + 10})`).join('\n'); // ~7.4 KB

const pattern = (rank: number, text: string, count: number) => ({
  rank, id: `p${rank}`, pattern: text, estimatedCount: count, sampleCount: count, shareOfTotal: 0,
  whereSnippet: `Body:"${text.slice(0, 2000)}"`, samples: Array.from({ length: 5 }, () => text),
});
const patterns = [
  ...[0, 1, 2].map(i => pattern(i + 1, stack(i), 9950 - i)),               // the 3 most frequent: long stack traces
  ...Array.from({ length: 27 }, (_, i) => pattern(i + 4, `Short INFO line ${i} took <*> ms`, 30 - i)),
];
const output = { summary: { totalCount: 100000, patternsReturned: patterns.length }, patterns, usage: 'shareOfTotal is …' };

const { data, isTrimmed } = trimToolResponse(output);
console.log(`isTrimmed=${isTrimmed}; topN requested=30, patterns returned=${data.patterns.length}`);
for (const p of data.patterns.slice(0, 4)) {
  const show = (v: unknown) => (typeof v === 'string' ? `"${v.slice(0, 40)}…" (${v.length} chars)` : JSON.stringify(v));
  console.log(`rank ${p.rank}: pattern=${show(p.pattern)} whereSnippet=${show(p.whereSnippet)} samples[0]=${show(Array.isArray(p.samples) ? p.samples[0] : p.samples)}`);
}