A desktop coding agent with free access to frontier models.
TypeScript
/**
* Kent-style instruction primitives for OmniRush.ai system-prompt composition.
*
* create — build one named section
* combine — merge sections in order, one id wins (first non-empty)
* delete — drop a section by id
* expand — replace a section body with derived text when present
* compose — combine sections and return their prompt bodies
*
* Sections are the unit of overlap control: routing/tools/skills/session each
* own one id so transforms stop stacking contradictory brochure text.
*/
export type AgentInstructionSection = {
id: string;
body: string;
};
type AgentInstructionSectionGroup = AgentInstructionSection | AgentInstructionSection[] | null | undefined;
export function createInstructionSection(id: string, body: string): AgentInstructionSection {
return { id, body: body.trim() };
}
export function combineInstructionSections(
...groups: AgentInstructionSectionGroup[]
): AgentInstructionSection[] {
const seen = new Set<string>();
const combined: AgentInstructionSection[] = [];
for (const group of groups) {
if (!group) continue;
const sections = Array.isArray(group) ? group : [group];
for (const section of sections) {
if (!section.body || seen.has(section.id)) continue;
seen.add(section.id);
combined.push(section);
}
}
return combined;
}
export function deleteInstructionSection(
sections: AgentInstructionSection[],
id: string,
): AgentInstructionSection[] {
return sections.filter((section) => section.id !== id);
}
export function expandInstructionSection(
sections: AgentInstructionSection[],
id: string,
expand: (body: string) => string,
): AgentInstructionSection[] {
return sections.map((section) => (
section.id === id
? { ...section, body: expand(section.body).trim() }
: section
)).filter((section) => section.body.length > 0);
}
export function composeAgentInstructions(...groups: AgentInstructionSectionGroup[]): string[] {
const seen = new Set<string>();
const instructions: string[] = [];
for (const group of groups) {
if (!group) continue;
const sections = Array.isArray(group) ? group : [group];
for (const section of sections) {
if (seen.has(section.id)) continue;
const body = section.body;
if (!body) continue;
seen.add(section.id);
instructions.push(body);
}
}
return instructions;
}
/**
* Extend the engine's system prompt in place from an
* `experimental.chat.system.transform` hook.
*
* OpenCode sends every entry of `system` as its own `role: "system"` message.
* Several chat templates behind OpenAI-compatible endpoints reject any system
* message after the first ("System message must be at the beginning."), so
* OmniRush.ai folds its instructions into the existing entry instead of pushing a
* second one. The engine alone sends a single system message; this keeps that
* shape intact. Sections are separated by a blank line so each heading starts
* its own paragraph instead of trailing the previous sentence.
*/
export function appendAgentInstructions(system: string[], ...groups: AgentInstructionSectionGroup[]): void {
const body = composeAgentInstructions(...groups).join("\n\n");
if (!body) return;
const last = system.length - 1;
if (last < 0) {
system.push(body);
return;
}
system[last] = system[last] ? `${system[last]}\n\n${body}` : body;
}