Adapters¶

Built-in adapters render the same canonical roles into tool-specific outputs.

Built-in targets¶

Target	Output path	Notes
`claude`	`.claude/agents/*.md`	supports model mapping, tool allow-lists, and delegation
`opencode`	`.opencode/agents/*.md`	supports model mapping, permissions, bash allow-lists, and delegation
`cursor`	`.cursor/agents/*.mdc`	ignores per-role model selection
`windsurf`	`.windsurf/rules/*.md`	emits rule files with `model_decision` trigger

Shared adapter model¶

All adapters now share one render pipeline:

validate hierarchy and delegation
validate output layout uniqueness
derive target delegate ids from the selected layout
compute the final output path
render one file per role

That keeps platform-specific code focused on frontmatter and permissions instead of repeating filesystem and validation logic.

Adapters also consume the same centralized capability expansion step. Built-in capability groups, aliases, bash policies, delegation metadata, and all are normalized before target-specific rendering begins.

Third-party adapters¶

Register adapters through the role_forge.adapters entry point group.

[project.entry-points."role_forge.adapters"]
my-target = "my_package.adapters:MyAdapter"

The adapter class should inherit BaseAdapter from role_forge.adapters.base and implement render_agent.