@cendor/squeeze

Shrink long prompts to hit a token budget and expand them back exactly — the same input always produces the same output, with no LLM and no model download. Compression returns a handle; the original is always restorable, byte-for-byte. The TypeScript port of cendor-squeeze.

The TypeScript port of cendor-squeeze. Satisfies @cendor/core's Compressor protocol by shape, so context tooling can use it without importing it.

Using an AI coding assistant? npx @cendor/init (TS) / uvx cendor-init (Python) wires it up — or point it at cendor.ai/docs/for-ai-assistants.

import { compress, decompress } from '@cendor/squeeze';

const [small, handle] = compress(hugeJson, { kind: 'auto' });                    // detect + route
const [small2] = compress(sourceCode, { kind: 'code', fidelity: 'aggressive' });
const [small3] = compress(logs, { kind: 'logs', targetTokens: 400 });           // compress to a budget
const original = handle.expand();                                               // restore, byte-for-byte
decompress(handle) === handle.expand();                                          // true

Surface

| Symbol | What it does | |---|---| | compress(content, opts?) | [small, handle]. opts: kind ("auto" detects), targetTokens (never exceeded), model ("gpt-4o"), fidelity ("lossless" \| "balanced" \| "aggressive"). Accepts a string or a JSON-serializable object. | | decompress(handle) | handle.expand() — the exact original. | | detect(content) | "json" \| "logs" \| "code" \| "prose". | | Handle | .expand(), .technique, .toDict(), Handle.fromDict(), .id, .kind, .originalRef, .restoreMap. | | SqueezeCompressor | Object form matching core's Compressor protocol. | | useStore(store) | Swap the content-addressed store (CCR) backend; returns the previous one. | | MemoryStore(maxItems?) · SQLiteStore(path) | CCR backends (also at @cendor/squeeze/store). MemoryStore is LRU-capped; SQLiteStore persists (via optional better-sqlite3). |

• Four compressors — JSON (minify + drop nulls; budget-shrink drops keys/elements structurally, staying valid JSON), logs (normalize timestamps/UUIDs/IPs/hex/integers + dedup repeats into (×N), chronological), code (string-aware comment stripping — a // or # inside a literal stays put; keeps preprocessor & shebang), prose (extractive, abbreviation-aware sentence splitting).
• Compress to a budget — targetTokens is never exceeded.
• 100% reversible — a content-addressed store (deduped by sha256) keeps every original; handle.expand() restores it no matter how hard you squeeze. The deterministic handle id is the first 32 hex of sha256(ref:technique).

Parity note

Faithful port of the Python package: same public symbols, defaults, string-literal values (kind, fidelity, technique strings, the × dedup marker U+00D7), algorithms, and error names (KeyError, invalid-fidelity throws). Python snake_case keyword args become a trailing options object; content_ref/restore_map stay snake_case inside toDict() for cross-language persistence.

One deliberate difference: token counts. @cendor/core bundles js-tiktoken, so tokens.count returns real tiktoken numbers for gpt-4o (not Python's forced ceil(len/4) test heuristic). The compressors call the same counter their budget checks do, so every guarantee — targetTokens never exceeded, reversibility, ordering — holds identically; only exact token counts differ from the Python test fixtures.