Tool Result Compressor

An agent loop accumulates tool results faster than any context window can hold them, so something has to be dropped. The usual approach — truncate the longest thing, or paraphrase — drops exactly what the next step needs: the reason a call failed, the order id it produced, and the fact that one call's outcome was never established at all.

This walkthrough uses the tool's public README and checked-in example files. Run the command from a repository checkout with Node.js 22+; inspect the source before using it on your own files.

Run the checked-in example

cat > results.json <<'JSON'
{
  "schemaVersion": "1",
  "results": [
    {
      "id": "call-0007",
      "tool": "payments.refund",
      "status": "failed",
      "sections": [
        { "name": "error", "kind": "failure", "priority": 0,
          "text": "Refund declined: acquirer code 51.",
          "sourceHash": "sha256:9f2c1a4b7e0d3c5a", "pointer": "results/call-0007/error.txt" },
        { "name": "order", "kind": "identifier", "priority": 0, "text": "ord_9F2KQ4" },
        { "name": "retry", "kind": "next-action", "priority": 0,
          "text": "Fund the source account, then retry with the same key." },
        { "name": "transcript", "kind": "evidence", "priority": 5,
          "text": "REQUEST POST /v2/refunds ...",
          "sourceHash": "sha256:deadbeefdeadbeef", "pointer": "results/call-0007/transcript.log" }
      ]
    }
  ]
}
JSON

node bin/tool-result-compressor.mjs --results results.json --budget-chars 600
node bin/tool-result-compressor.mjs --results results.json --budget-chars 600 --json

Read the result

A consumer that pipes stdout must handle an empty stdout on exit 2. That is documented rather than papered over: emitting a fake report for a run that never started would be worse.

Where this check stops

Every limit is enforced and every one has a CLI flag. Reaching one produces an incomplete report naming the limit — never a silent truncation and never a pass.

Before adapting the command to your own workflow, review the accepted inputs, exit codes and safety boundaries in the README.

Compiled with AI assistance from checked-in public documentation and example scripts. Run the example and review the repository's current documentation before relying on its result.