@backthread/redact

The one redaction fence used across Backthread. Pure, zero-dependency string transforms that strip source code and tool I/O out of an AI coding-agent session transcript before anything leaves the machine — keeping only the natural-language rationale.

The rule (load-bearing)

DROP every tool-use / tool-result record entirely. Keep only natural-language user prompts and assistant text/thinking. In the kept text, REDACT fenced code blocks (``` … ```) to a placeholder.

No source code and no tool I/O may leave this module — only derived rationale. This is defense-in-depth: redact at the boundary so a downstream bug can't exfiltrate code.

Install

npm install @backthread/redact

Requires Node.js ≥ 22.18. This package ships as source-only TypeScript (exports points at ./src/index.ts) and relies on Node's default type stripping. Zero runtime dependencies.

Usage

import { parseJsonl, redactTranscript, renderTranscript } from '@backthread/redact';

const records = parseJsonl(rawJsonlFromDisk);
const redacted = redactTranscript(records); // only prose survives; code → "[code redacted]"
const text = renderTranscript(redacted);

API

• redactTranscript(records: unknown[]): RedactedTranscript — drop tool records, keep prose, redact code fences.
• redactCodeFences(text: string): { text: string; count: number } — redact ``` fences in a single string.
• renderTranscript(t: RedactedTranscript): string — render the surviving turns back to text.
• parseJsonl(raw: string): unknown[] — parse a .jsonl transcript.
• sessionTimestamp(records: unknown[]): string | null — first timestamp in a transcript.
• CODE_REDACTION — the placeholder string substituted for each fenced code block.

License

MIT © Backthread