q-kdb-code-review（q kdb 代码审查）

q-kdb-code-review

AI-powered code review for Q/kdb+ — catch bugs, performance issues, and security vulnerabilities in the most terse language in quantitative finance.

What it does

Reviews Q/kdb+ code with deep understanding of Q idioms, performance patterns, and common pitfalls. Built for quant developers, kdb+ DBAs, and trading infrastructure teams.

Catches:

• Type errors in implicit casts (e.g., mixing longs and floats in comparisons)
• Rank errors from wrong argument counts in function calls
• Unescaped signals in protected evaluation
• Memory-inefficient queries (selecting all columns when only some are needed)
• Missing peach parallelism opportunities for embarrassingly parallel operations
• Unsafe eval/value usage on user-supplied strings (Q injection)
• Unlocked tables during concurrent inserts
• Missing ` g# `` grouped attributes on high-cardinality join columns
• N-squared joins that should be aj (asof joins) or wj (window joins)
• Race conditions in timer callbacks (.z.ts)
• Unprotected IPC handlers (.z.pg, .z.ps) and exposed .z.pw

Strictness modes:

Intelligent routing via Astrai: Complex algorithmic Q (custom signal generation, real-time CEP) routes to powerful models. Simple table operations (selects, inserts, schema definitions) route to cheaper, faster models. You get the best result at the lowest cost.

BYOK (Bring Your Own Keys): Your provider API keys, your billing. Astrai routes to the best model among your configured providers.

Setup

1. Get a free API key at as-trai.com
2. Set your API key:

   export ASTRAI_API_KEY="your_key_here"

1. Optionally add provider keys for BYOK routing:

   export ANTHROPIC_API_KEY="sk-ant-..."
   export OPENAI_API_KEY="sk-..."

1. Run /review-q on any .q file

Usage

/review-q                          Review current Q file
/review-q --strict                 Strict mode: bugs + performance + style
/review-q --focus security         Security mode: eval injection, IPC, .z.pw
/review-q --file tick.q            Review a specific file

Example output

Reviewing tick.q (strict mode)...
Model: claude-opus-4-6 via Astrai

Found 3 issues:

[CRITICAL] Line 12: Missing `s# attribute on time column
  `trade` table uses `aj` but `time` column lacks sorted attribute.
  Without `s#`, asof join scans linearly — O(n) instead of O(log n).
  Fix: trade: `trade upsert update `s#time from trade

[WARNING] Line 34: Using `each` where vector operation suffices
  {x*y} each' (price;qty) can be replaced with price*qty
  Vector multiply is ~100x faster than each-both.

[INFO] Line 45: Consider `peach` for independent symbol processing
  Processing each symbol sequentially. Since operations are independent,
  `peach` would utilize all cores.
  Fix: results: func peach syms

Summary: 1 critical, 1 warning, 1 info. Focus on the missing sorted
attribute — it will cause aj performance to degrade from microseconds
to milliseconds at scale.

Environment Variables

External Endpoints

Security & Privacy

• No code storage: Your Q code is sent to the selected AI provider for inference and is not stored by Astrai.
• BYOK: When you provide your own provider keys, requests go directly through Astrai's router to your provider account. Astrai does not store or log your provider keys beyond the request lifecycle.
• Transport: All communication uses HTTPS/TLS.
• No telemetry: The skill does not send analytics or telemetry data. Only the review request goes to Astrai.
• Local processing: File reading and result formatting happen entirely on your machine.

Why Q needs specialized review

Q is unlike any mainstream programming language:

• Extreme terseness: A single line of Q can express what takes 20 lines in Python. This density makes bugs nearly invisible during manual review.
• Implicit type coercion: Q silently coerces types in many operations. Comparing a long to a float, or joining on mismatched key types, can produce silently wrong results.
• 1000x performance gaps: The difference between idiomatic and naive Q is not 2x or 10x — it is often 1000x. Missing a sorted attribute on a time column turns an O(log n) asof join into O(n). Using each instead of vector operations adds interpreter overhead per element.
• Adverb complexity: Q's adverbs (/, \, ', /:, \:, ':) modify function behavior in powerful but subtle ways. +/ is reduce-add, +\ is scan-add, +' is each-both-add. Confusing these causes wrong results, not errors.
• Most AI models struggle: Without Q-specific prompting, general-purpose AI models treat Q code as line noise. This skill provides detailed system prompts that teach the model Q semantics, kdb+ internals, and finance-domain patterns.

Pricing

Uses Astrai's inference routing. Your costs depend on the models selected:

With BYOK, you pay your provider directly at their rates. Astrai's routing is included in the plan price.