Could the generated AGENTS.md block include the .workflow.ts syntax rules, shared with the skill? #633 concerns skill invocation; this proposal concerns syntax available before invocation.
It would contain format rules only, without workflow-engineering guidance:
- Read the
<workflow-map> header first.
- Use one
@links() method per class.
- Use
out(N).to(in(M)), with 0-based indices.
- AI sub-nodes connect with
.uses() and expose .output, never .out().to().
ai_tool takes an array; ai_languageModel, ai_outputParser and ai_memory take single refs.
The skill must keep these rules. The plugin-installed skill is invokable from any directory, including sessions that never load a project's AGENTS.md. Outside a workspace, it is the only visible copy.
This would duplicate content from one source. AiContextGenerator already emits these artifacts, and getWorkspaceAgentContent() derives from getAgentSkillContent(). A shared fragment interpolated into the skill and getAgentsContent() follows that pattern, leaving nothing to keep in sync manually.
This is purely additive to the always-on block. #229 asked to shrink that block because of context overhead and domain bleed, and the current thin block reflects that request. Limiting the addition to format rules narrows its scope, but still adds unconditional context.
Our experience, not evidence for the proposal: we import the whole SKILL.md into our AGENTS.md so it autoloads. It works, but includes Workspace Readiness, Bootstrap Order, Managed Local Runtime and Environments — all inert once our workspace is configured. That incurs exactly the cost #229 objected to; our workaround does not establish that everyone needs these rules loaded unconditionally.
Fixing invocation first may make this unnecessary. #633 is the higher-value change; nothing here is urgent, and closing this as superseded would be fine.
Would you consider a shared, format-only syntax section compatible with #229?
Could the generated
AGENTS.mdblock include the.workflow.tssyntax rules, shared with the skill? #633 concerns skill invocation; this proposal concerns syntax available before invocation.It would contain format rules only, without workflow-engineering guidance:
<workflow-map>header first.@links()method per class.out(N).to(in(M)), with 0-based indices..uses()and expose.output, never.out().to().ai_tooltakes an array;ai_languageModel,ai_outputParserandai_memorytake single refs.The skill must keep these rules. The plugin-installed skill is invokable from any directory, including sessions that never load a project's
AGENTS.md. Outside a workspace, it is the only visible copy.This would duplicate content from one source.
AiContextGeneratoralready emits these artifacts, andgetWorkspaceAgentContent()derives fromgetAgentSkillContent(). A shared fragment interpolated into the skill andgetAgentsContent()follows that pattern, leaving nothing to keep in sync manually.This is purely additive to the always-on block. #229 asked to shrink that block because of context overhead and domain bleed, and the current thin block reflects that request. Limiting the addition to format rules narrows its scope, but still adds unconditional context.
Our experience, not evidence for the proposal: we import the whole
SKILL.mdinto ourAGENTS.mdso it autoloads. It works, but includes Workspace Readiness, Bootstrap Order, Managed Local Runtime and Environments — all inert once our workspace is configured. That incurs exactly the cost #229 objected to; our workaround does not establish that everyone needs these rules loaded unconditionally.Fixing invocation first may make this unnecessary. #633 is the higher-value change; nothing here is urgent, and closing this as superseded would be fine.
Would you consider a shared, format-only syntax section compatible with #229?