README

Reads and writes. Both directions are covered: readers (decode bytes → values) and writers (encode values → bytes), split under src/readers/ and src/writers/. The reader path is the more mature one — the writers mirror it type-for-type, with a few decode-only paths (Dynamic, JSON, the runtime header/compile path, and the columnar typed-array path) not yet mirrored.

Status

• ✅ Sonnet 4.6: 60% -> 94.0% pass rate
• ✅ Opus 4.8: 71% -> 94.7% pass rate
• ✅ Haiku 4.5: 52% -> 86.0% pass rate
• ✅ Composer 2.5 Fast: 3x parser performance
• ✅ 724/724 tests (readers + writers)
• ✅ type-checked
• ✅ benchmarked

Example

Take a small orders result:

The API-only reader — what you write by composing the library’s combinators. Correct, clear, and a fine default:

The optimized reader the skill generates — same row, monomorphized to straight-line code. The whole row is fixed-width (1 + 16 + 8 + 1 = 26 bytes), so the four separate bounds checks coalesce into one advance(s, 26) and every leaf read happens at a constant offset; the per-field combinators are gone:

Same values, same streaming-safety — ~3.4x faster.

How to use

As a library (comes with the skill):

As a skill only:

Using it with the ClickHouse JS client

This library only decodes the bytes — it doesn’t open connections. Pair it with the official client to fetch a RowBinary response and feed the byte chunks into streamRowBatches(chunks, readRow).

RowBinary isn’t one of the formats the client decodes itself, so don’t use client.query({ format: ... }) for it. Instead use client.exec({ query }) with the FORMAT RowBinary clause written into the SQL yourself — exec hands back the raw, undecoded byte stream of the response, which is exactly what this library consumes. (Use plain RowBinary, not RowBinaryWithNamesAndTypes, unless your reader also skips the leading names/types header.)

The row reader below is the orders example from EXAMPLES.md; swap in the reader the skill generates for your own columns.

Why it’s worth it

Four pillars — speed, correctness, judgment, and lifting smaller models:

• ~2–3x faster code than the straightforward decoder. The skill emits monomorphized, flattened, straight-line code — inlined reads, bounds checks coalesced across adjacent fixed-width columns, the right array layout — measured at ~1.3–3.4x over the same logic written with the plain combinator API (npm run bench). This is why
  • inlined JIT friendly code
  • benchmarked hot paths
  • minimal allocations
  • v8 and Node.js specific optimizations
• Correct on the gotchas that otherwise quietly break. UUID byte order, Variant‘s sort-by-type-name discriminant, DateTime64 sub-second precision, signed-high-word wide integers, faithful decimals, Dynamic/JSON self-description, transparent wrappers, opaque AggregateFunction — each encoded with a live, server-verified test (details below).
• Judgment, not just code. The skill carries the working knowledge to make the right call before writing a line, so the agent neither over- nor under-engineers:
  • Is RowBinary even right? For string-heavy results read as text, a JSON* format + V8’s native JSON.parse (plus gzip/zstd) can beat a JS RowBinary decoder — reach for RowBinary when the data is numeric / wide-integer / binary-blob heavy.
  • Whole buffer or stream? Drop the advance() bounds checks for a complete in-memory buffer (faster); keep them for a chunked HTTP response that must survive rows straddling chunk boundaries.
  • Drop the portability scaffolding. RowBinary is little-endian and the target is x86/ARM, so the skill steers away from big-endian / byte-swap “portability” code a cautious one-shot pass tends to add.
• Improves smaller models’ performance. Because the skill hands over the hard-won answers up front, it lifts a weaker model the most. In a 24-eval with-skill vs no-skill benchmark, the skill raised Sonnet 4.6 from 60.4% to 94.0% (+34pp) — bringing it level with skill-equipped Opus 4.8 (94.7%), which itself gained +23pp (71.5% → 94.7%). Composer 2.5 Fast got a 3x parser performance boost, Haiku 4.5 raised from 52% to 86% — the skill closes most of the model-capability gap on this task.

What it does

Given the columns of a query result — their names and ClickHouse type definitions (as returned by RowBinaryWithNamesAndTypes, or supplied by the user) — the skill generates parser code tailored to exactly those types. Rather than shipping a generic, runtime-driven decoder, it emits straight-line code that reads each column in order, so the parser only contains the logic the specific result shape needs.

Schema only known at runtime? compileRowBinaryWithNamesAndTypes(cursor) reads the RowBinaryWithNamesAndTypes header and folds each column type into a reader on the fly (type strings parsed by @clickhouse/datatype-parser), returning a readRows driver for the rest of the stream — a generic, no-codegen path for dynamic schemas. The specialized codegen above stays the fast path when the types are fixed.

Correctness on the gotcha-heavy types

For a plain UInt64, String, DateTime result a strong model already writes fast, correct code on its own. The skill earns its keep on the long tail of RowBinary traps — the encodings where a from-scratch decoder is quietly wrong — each one captured here with a live, server-verified test:

• UUID — two little-endian UInt64 halves, each byte-reversed vs. the text form (not 16 bytes in order).
• Variant(...) — the 1-byte discriminant indexes the alternatives sorted by type name (ClickHouse globally sorts them), NOT declaration order; 0xFF is NULL.
• DateTime64(P) — returned as [Date, nanoseconds] so the sub-second part isn’t lost to a Date‘s millisecond resolution; Time/Time64 are durations, not instants.
• Wide integers — Int128/Int256 compose from 64-bit words with the high word read signed; 64-bit values stay bigint, never a lossy number.
• Decimals — kept as the exact [unscaled, scale] pair, not a lossy float.
• Dynamic / JSON — self-describing: a per-value binary type encoding, then the value; declared typed JSON paths are written without a tag (need the schema). Wrappers are erased (Nullable/Variant → concrete type).
• Transparent wrappers — LowCardinality(T) / SimpleAggregateFunction(f, T) decode as the inner T (no dictionary layer in RowBinary); Nested(...) is Array(Tuple(...)) with no wire of its own.
• AggregateFunction(...) — opaque, unframed state: not decodable or even skippable; finalize server-side instead.
• FixedString preserves trailing NUL padding; Enum decodes to the underlying int (the name map is metadata); BFloat16 is the top 16 bits of a Float32.

This is also where a raw model is most likely to go wrong. In a clean-room test on a Variant / UUID / DateTime64 / LowCardinality schema, a no-skill Sonnet produced a silently wrong UUID (treated the bytes as plain, missing the two-reversed-halves layout), and a no-skill Opus got it right only after three web searches. The skill hands over these answers up front — correct by construction, no lookups. See baseline/README.md for the full control.

And the failure isn’t a one-off — it’s a coin-flip. Running the same no-skill Sonnet on the orders schema (UInt8, UUID, Decimal64(2), Enum8) 5 times in isolation, only 3 of 5 runs decoded correctly; both failures were the same UUID byte-order scramble. Even the passing runs varied ~1.9x in generated-code throughput. With the skill, every run is correct. So a single A/B undersells the gap: from scratch the model is right roughly 60% of the time and silently wrong the rest, while the skill makes correctness deterministic.

Examples

Six end-to-end examples live in EXAMPLES.md. Each ships both an API-combinator reader and an optimized, monomorphized one, with a runnable round-trip test and a benchmark — so the speedups below are measured, not claimed (Node 24 / V8; npm run bench for your own numbers):

Two axes drive the win. Composite structure is one: monomorphization pays in proportion to how many per-row combinator closures it removes (carts / telemetry / observability). Per-row formatting is the other, independent of composites: orders is all-scalar yet the biggest win (~3.4x), almost entirely from swapping the BigInt UUID formatter for the lookup-table formatUUIDTable. The genuinely flat case — a scalar row with no hot formatter (events) — is on par, so the simpler API reader is the right call there. Measure, don’t assume.

Scope

• In scope (reading): RowBinary, RowBinaryWithNames, and RowBinaryWithNamesAndTypes decoding for Node.js — full-buffer and streaming (chunked) via advance()/NeedMoreData, readRows(), and the async streamRowBatches() (with a built-in small-chunk warning and the optional coalesceChunks() debounce filter).
• In scope (writing): the inverse encode path — a writeX mirroring every readX, appending bytes to a Sink, plus writeRows(). Imported from @clickhouse/rowbinary/writer. A handful of decode-only paths are not yet mirrored: Dynamic, JSON, the runtime header/compile path, and the columnar typed-array path.
• Out of scope (for now): browsers and Edge runtimes, non-RowBinary formats (JSON / CSV / TSV / Parquet), and big-endian hosts.

The spirit

A RowBinary codec generator is a narrow thing. But it’s built as an instance of a broader bet about what libraries become once a capable LLM is part of the toolchain. Three shifts, each already visible in this repo:

• Self-modifiable software. The library deliberately ships several equivalent decoders for the same type — readUUID / readUUIDBigInt / readUUIDHiLo, formatUUID / formatUUIDTable, new Array(n) vs []+push, streaming vs whole-buffer — because the fastest one depends on the workload, not the type. Today the agent picks at generation time from measured benchmarks. The next step is to pair the skill with a tracing layer that runs variant A against variant B on the live workload and keeps whichever wins for this data shape and access pattern — a parser that re-tunes itself as the traffic drifts, instead of freezing one author’s guess into a release.

• Custom software. The value here isn’t a fixed high-level API; it’s the benchmarked building blocks plus the judgment to combine them. So the end user doesn’t bend their code to the authors’ generic surface — they have the agent assemble the high-level API they actually want, shaped to their queries, row shapes, and latency/memory budget. Two teams with different workloads grow two different libraries from the same primitives, and neither inherits a design decision that was only ever right for the original authors’ use case.

• Read-write libraries. For either of the above to be safe, the source has to be legible to an LLM, not merely runnable. So this repo is written read-write: every tradeoff is commented where it’s made — the per-column ClickHouse type annotations, the SAFE TO TOGGLE markers on the fast variants, each reader’s doc comment carrying its exact monomorphized form. An LLM can read why a decision was made and change it in depth with confidence — not just call the public functions, but safely rework the internals.

The through-line: the last mile is glue the LLM writes over stable, benchmarked blocks, so the authors’ job shrinks to exporting good primitives and documenting their tradeoffs honestly — rather than trying to bake the right performance constants for every possible workload into the library ahead of time.