hive-claude

Claude Code as a hive backend. It exposes Claude Code through the hive addon ports as a headless backend, and names no editor: it loads, runs and reports healthy with no Emacs, Vim or other vessel present.

Coordinates

;; deps.edn
io.github.hive-agi/hive-claude {:mvn/version "0.1.11"}

Two backends, one addon

hive-claude.init is the IAddon implementation — a reify over a nil-railway pipeline, with zero compile-time hive-mcp dependencies; everything host-side resolves through requiring-resolve.

Process. hive-claude.headless.process-backend registers the :claude-process headless backend: claude driven through a ProcessBuilder.

Terminal sessions belong to the vessel. A Claude Code session inside an editor buffer (the :claude terminal) is contributed by the vessel that owns that editor. hive-emacs registers it from hive-emacs.claude-terminal, with the hive-claude-* elisp package in its own elisp/ tree. Another vessel adds its own terminal the same way; hive-claude does not change.

Agent SDK. hive-claude.sdk.agentic-loop implements IAgenticLoop over the Claude Agent SDK:

• async start/abort lifecycle via sdk.lifecycle spawn/kill
• collect-response! with a CompletableFuture and a timeout
• cost tracking accumulated from SAA completion events
• transcript access from session observations
• mid-session constraints through budget guardrails

Layout

Path	Contents
src/clj/hive_claude/	Addon init, guard projection
src/clj/hive_claude/headless/	:claude-process headless backend
src/clj/hive_claude/sdk/	Agent SDK lifecycle, execution, event loop, options, availability, SAA, phase compression

The Python-side Agent SDK dependencies (libpython-clj) are loaded lazily through requiring-resolve, so the process backend works on a JVM that has no Python environment at all.

License

AGPL-3.0-or-later.