@agentxm/agent-integration
v0.30.2
Published
AXM agent integration: effectful detection of installed AI coding agents plus the native-format mechanics AXM writes into them — coding-agent adapters, subagent rendering, MCP entry projection and config writing, hook-group editing, and the ownership mark
Readme
@agentxm/agent-integration
Everything AXM knows about a coding agent's native surfaces:
- Detection — project- and user-scope marker evidence and executable resolution for installed AI coding agents.
- Path primitives — home and config directories, catalog-derived agent install paths.
- Coding-agent adapters — the per-agent
CodingAgentcontract: skills directory resolution, MCP entry add/remove, subagent publication. - Native formats — subagent rendering (Markdown+YAML, TOML, JSON, Roo modes), MCP entry projection and config writing across JSON/JSONC/YAML/TOML, hook-group JSON editing, agent override merge patches, and the versioned ownership marker grammar those writers stamp into managed files.
The package depends only on @agentxm/extension-model, so every input crosses
as plain data: a rendered MCP entry, an ownership metadata record, a generation
token, banner text. Native writes go through the NativeWriteAuthority port
(protect, record); a core capability supplies the layer that joins it to
the enclosing workspace transaction, so a target that cannot be snapshotted is
refused before it is mutated. Environment-backed layers live behind ./live
and test doubles behind ./testing.
Unstable and unsupported — use the axm.sh CLI.
