synsema

Synsema is a programming language specifically designed for AI agents, distinguishing itself from frameworks or libraries by making observability, security, multi-agent coordination, human interaction, and LLM integration native language primitives. It compiles to a single native binary with no runtime, no GIL, and true multi-core support. Key features include: - **Performance**: Benchmarks show it matches or beats Go on HTTP throughput (e.g., 47.2k req/s vs 42.8k for plaintext endpoints). - **Security**: Deny-by-default capability security at the language level—no network, file, or DB access without explicit `require` declarations. Includes declarative route auth, input validation, and automatic audit logs. Optional information-flow labels track sensitive values through operations and branches, with `declassify` as the only escape. Attestation support binds running code to its source in confidential deployments. - **LLM Integration**: Built-in `analyze`, `decide`, and `generate` operations with swappable providers (Anthropic, OpenAI, Ollama). Responses are validated automatically with retries. - **Human Interaction**: `approve`, `ask`, `confirm`, and `show` primitives for approval gates and questions. - **Multi-Agent Coordination**: Agents communicate via blackboard (shared state with versioning), signals, and resource locks. `spawn` creates agents with OS threads. - **Concurrency**: Real multi-core parallelism via `parallel_map` and `chunk`, with bounded fan-out and fail-fast semantics. - **Web Server**: Native HTTP server with routing, auth, validation, pagination, rate limiting, SSE streaming, TLS/auto-HTTPS, HTTP/2, virtual hosts, and reverse proxy—no external proxy needed. - **Observability**: `trace`, `log`, `measure`, and `checkpoint` primitives, plus rich error diagnostics with variable snapshots and recovery suggestions. - **Agent Memory**: Progress tracking, persistent memory (`remember`/`recall`), and owner rules with levels (must/should/avoid/prefer). - **Language Design**: English-like syntax (e.g., `let name be "World"`, `task greet(person)`), pipe operator, intentional operations (`where`, `collect`, `apply`, `reduce`), custom types, and flat document-style syntax for `.fsyn` files. - **Tooling**: CLI with `run`, `serve`, `check`, `repl`, `ast`, `tokens`, and `testgen` commands. Auto-generates edge-case tests. Includes a VS Code extension and an AI skill for coding assistants like Claude Code. Installation is a single self-contained binary via curl or npm, with source builds via Cargo. The engine is modular, organized into crates for core, capabilities, stdlib, agents, runtime, LLM, and CLI.