ExternalConfigs.Catalog (fnord v0.9.29)

View Source

Builds the external-configs section of the coordinator's system prompt.

Emits a single combined "available skills" message that groups:

  • fnord's own enabled skills (via Skills.list_enabled/0)
  • project-enabled Cursor skills
  • project-enabled Claude Code skills

Also emits a compact "available Cursor rules" listing covering modes other than :always (agent-requested and auto-attached rules), plus one full-body message per :always rule.

All messages are system-role messages; they are filtered out when the conversation is persisted to disk (see Services.Conversation).

Summary

Types

skills_section()

Functions

build_messages()

Build all prompt messages for the current project's external configs. Returns an ordered list of system-message strings, each intended to be appended to the conversation as its own system message.

build_messages(project)

Build all prompt messages for the given project. Same shape as build_messages/0, but skips the Store.get_project/0 lookup. Intended for callers that already have the project in hand (tests, the injector's matching path).

render_auto_attached_rule(rule, file_path)

Render the system-message body that is injected when an auto-attached rule matches a file the model has just read or written. Includes the triggering file's path so the model can cite it back to the user.

system_messages()

Convenience wrapper: build_messages/0 with each string already wrapped as a system-role message. Intended for sub-agents that build their own messages list and need to inherit the external-configs catalog so their LLM call knows about the parent session's enabled sources.

Types

skills_section()

@type skills_section() ::
  {String.t(), [%{name: String.t(), description: String.t() | nil}]}

Functions

build_messages()

@spec build_messages() :: [String.t()]

Build all prompt messages for the current project's external configs. Returns an ordered list of system-message strings, each intended to be appended to the conversation as its own system message.

Safe when no project is selected, when the project has source_root = nil, or when no toggles are enabled; in those cases it returns [].

build_messages(project)

@spec build_messages(Store.Project.t()) :: [String.t()]

Build all prompt messages for the given project. Same shape as build_messages/0, but skips the Store.get_project/0 lookup. Intended for callers that already have the project in hand (tests, the injector's matching path).

render_auto_attached_rule(rule, file_path)

@spec render_auto_attached_rule(ExternalConfigs.CursorRule.t(), String.t()) ::
  String.t()

Render the system-message body that is injected when an auto-attached rule matches a file the model has just read or written. Includes the triggering file's path so the model can cite it back to the user.

system_messages()

@spec system_messages() :: [map()]

Convenience wrapper: build_messages/0 with each string already wrapped as a system-role message. Intended for sub-agents that build their own messages list and need to inherit the external-configs catalog so their LLM call knows about the parent session's enabled sources.

Sub-agents get the same skills catalog, cursor rules catalog listing, and always-apply rule bodies the coordinator got at its own bootstrap. They do NOT inherit mid-stream auto-attach injections from the parent conversation; those live in Services.Conversation which sub-agents don't read from. The rule bodies remain reachable via the paths listed in the catalog.