sparklang

SparkLang (repository sparklang, CLI spark, product name SparkLang) is an MIT-licensed programming language and runtime that executes plain .spark files. The README positions dry-run fixtures, compile/bytecode handling, playbooks and the Spark IDE as the core story, with network/web, voice/PSTN and live model calls treated as optional surfaces. It states explicitly that it is not Apache Spark, not AdaCore SPARK, and not a Bifrost plugin. Default behaviour is offline: dry-run needs no keys and no network, and only an opt-in --live flag uses a companion HTTP helper with an OpenAI-compatible gateway configured through AI_GATEWAY_URL and an API key. Environment variables such as SPARK_STT_URL / SPARK_TTS_URL and SPARK_TRAIN_* gate optional speech and training endpoints. Language surface described in the README includes ask/generate model calls with slots, classify with single or multi-label enums and confidence, extract with inline schemas, pipeline chaining, tool declarations, listen/speak/voice sessions, review/builder/implement for static review and codegen into out/, model operations (train, build, status, analyze, compare, improve, plan), head abstain/train/attach/ask for an IDK gate on local models, binary open/elf/disasm/understand, network capture and pcap analysis, os design/specify/generate blueprints, browser/mitm operations, cuda/memory/pcie probes, and crypto/encrypt/gateway for encrypt-to-model envelopes using AES-256-GCM. The repository is organised around a documented authoring plan labelled A+B+C: a long-term self-host destination, a thin C bootstrap VM (bootstrap/), and a Spark-native assembler (sparkasm/). The GAS x86_64 assembly tree under asm/ is described as a disposable running scaffold rather than permanent source of truth. Layout entries include asm/, tools/ask/, tools/engine/, tools/browser/, selfhost/, hdl/classify_score.v, examples/*.spark, docs/*.md and a website/ directory for the marketing site. Verification is offered through concrete commands: make, make machine-proof (claims asm-to-ELF output, verified with file), dry-run examples such as hello.spark, http_get.spark, classify_intent.spark and model_improve.spark, plus make test, make test-e2e-browser, make sparkbc-e2e and make test-ide-paint. Python and Node entry points are shown for run(). Safety and honesty statements are prominent: review never evaluates web JavaScript; remote review URLs require --allow-net and otherwise exit with a clear error; network capture stays behind --allow-net-capture with a documented exit code 4 on missing CAP_NET_RAW; dry-run network probes report claimed:false; os generate writes educational stubs only and never reboots or installs an OS; model training in dry-run never starts jobs. Documented non-goals include a custom CPU ISA, a production bare-metal OS, full ES/CSS rendering, a finished self-host compiler, FPGA bitstream shipping claims, and any silent GPU training. Version stated in the README is 0.6.2, with the installer kit listed as 0.6.0.