@metaharness/host-prime-agent
v0.1.0
Published
Prime Agent (PrimeIntellect-ai/prime-agent) host adapter for agent-harness-generator. Emits project-scoped Python-backed skills (ADR-247).
Maintainers
Readme
@metaharness/host-prime-agent
Prime Agent (PrimeIntellect-ai/prime-agent) host adapter — the 11th harness host, per ADR-247.
What it emits
Integration lands as project-scoped skills under .prime/agent/skills/:
spec.tools[]→ instruction-onlySKILL.mdcontracts.ToolSpecdoes not contain an executable handler, so the adapter never fabricates one.- Remote HTTP
spec.mcpServers[]→.prime/agent/settings.jsonplus real Python-backedrlm.McpIntegrationskill packages. Local stdio MCP remains explicitly unsupported by Prime Agent's current kernel integration. spec.systemPrompt→.prime/agent/APPEND_SYSTEM.mdspec.agents[]→ skills that delegate through Prime Agent's nativerlmrecursive-agent callable- Always:
install-prime-agent.md(install runbook, skill precedence notes, MCP availability, and the ADR-246 autonomous invocation snippet when present)
Sandbox caveat
Prime Agent is not sandboxed. It executes model-written Python at user
permission level with no native allow/deny enforcement. When
spec.permissions.deny is non-empty, this adapter fails closed: it emits a
prominent SANDBOX-REQUIRED.md (and the same warning at the top of
install-prime-agent.md) enumerating the denied capabilities that require an
external sandbox (container, RVM per ADR-018, or equivalent). The deny-list is
never silently dropped (ADR-247 §2.2).
