New:Introducing Socket Scanning for VS Code Marketplace Extensions.Learn more →
Get Started
s

swsgbl

npm

26.02K

total weekly downloads

Packages

Current Co-maintainers

Former Co-maintainers

16 packages

@hmharness/agent

hmharness agent execution layer: base tools, system prompt, sub-agent spawn, and the shared task runner that frontends (cli, web) drive.

@hmharness/browser

hmharness BrowserOS driver — discovers, trust-pins and launches the BrowserOS AI browser (Chromium fork), then drives it over the Chrome DevTools Protocol. Client only, zero dependencies, dedicated profile: the user's daily browser is never touched.

@hmharness/cli

hmharness command line: one-shot tasks, an interactive REPL, a fullscreen TUI, the web frontend, and direct tool invocation.

@hmharness/codexhost-bridge

Machine-readable hmharness bridge for CodexHost and other agent desktop hosts.

@hmharness/cognitive

hmharness Cognitive OS core: Environment protocol, World Model, Goal system, Exploration, RLM runtime, five-layer memory, Skill Compiler, Continual Learning control plane, Evolution 2.0 governance, Multi-Agent topology, Transfer lab and GeneralBench schem

@hmharness/domain-harmony

hmharness HarmonyOS domain core: device, toolchain, build, and project lifecycle capabilities. HarmonyOS is the framework's native domain, not an add-on.

@hmharness/domain-ops

hmharness ops keeper: ecosystem radar (OpenHarmony release tracking, snapshot diff, model brief with template fallback). Rewrite of the old line's hm-keeper per its captured behavior spec.

@hmharness/environments

HMH Cognitive OS environment adapters: Terminal (native), HarmonyOS (via domain-harmony), Browser/Desktop (via host automation bridge), ARC3 (external adapter). Every adapter implements the @hmharness/cognitive Environment protocol and must pass the share

@hmharness/evaluation

hmharness evaluation: the Evaluator/Judge contract (V2 blueprint M2). Hard evidence outranks LLM judgment - build results, exit codes, exact/regex assertions first; the LLM judge is a last resort and is labeled as such. Evaluations attach to trajectories

@hmharness/evolution

hmharness evolution subsystem: persistent memory, insight capture, skill library, and the bench that gives evolution its fitness signal. First-class kernel citizen, not a plugin.

@hmharness/extension

hmharness browser-extension bridge — a loopback server + pairing flow (the ChatGPT desktop-app ↔ extension pattern) that lets the agent see and, with approval, act in the user's REAL browser tabs across Chrome/Edge/Brave (MV3 service worker + side panel),

@hmharness/kernel

hmharness kernel: tool registry, provider adapters, the agent loop, session log, config. Zero runtime dependencies (Node >=22 native fetch).

@hmharness/lsp

hmharness LSP Client/Manager — drives external language servers (LSP 3.18 over stdio) for agent code intelligence. Client only: we never implement language servers.

@hmharness/observability

hmharness observability: typed RunEvent schema + append-only TrajectoryStore. Every agent run becomes a reproducible, comparable, exportable experiment sample (V2 blueprint M1 - flight recorder). Zero runtime dependencies.

@hmharness/sandbox

hmharness sandbox runtime (V2 blueprint M3): isolated workspaces with create/exec/read/write/snapshot/restore/diff/destroy, tiered permissions (READ_ONLY / WORKSPACE_WRITE / FULL_ACCESS), git-backed snapshots. Runtime owns agent turns; the sandbox owns wo

@hmharness/web

hmharness web frontend: local node:http server with SSE streaming, remote approval gate, and an embedded zero-build single page app.