164 lines
5.4 KiB
TypeScript
164 lines
5.4 KiB
TypeScript
/**
|
|
* llm-model-selector.ts — LLM-powered intelligent model selection.
|
|
*
|
|
* Uses an LLM to understand model capabilities and assign optimal models to DevClaw roles.
|
|
*/
|
|
import { runCommand } from "../run-command.js";
|
|
import { ROLE_REGISTRY } from "../roles/index.js";
|
|
import type { ModelAssignment } from "./smart-model-selector.js";
|
|
|
|
/**
|
|
* Build a ModelAssignment where every role/level maps to the same model.
|
|
*/
|
|
function singleModelAssignment(model: string): ModelAssignment {
|
|
const result: ModelAssignment = {};
|
|
for (const [roleId, config] of Object.entries(ROLE_REGISTRY)) {
|
|
result[roleId] = {};
|
|
for (const level of config.levels) {
|
|
result[roleId][level] = model;
|
|
}
|
|
}
|
|
return result;
|
|
}
|
|
|
|
/**
|
|
* Build the JSON format example for the LLM prompt, derived from registry.
|
|
*/
|
|
function buildJsonExample(): string {
|
|
const obj: Record<string, Record<string, string>> = {};
|
|
for (const [roleId, config] of Object.entries(ROLE_REGISTRY)) {
|
|
obj[roleId] = {};
|
|
for (const level of config.levels) {
|
|
obj[roleId][level] = "provider/model-name";
|
|
}
|
|
}
|
|
return JSON.stringify(obj, null, 2);
|
|
}
|
|
|
|
/**
|
|
* Validate that a parsed assignment has all required roles and levels.
|
|
*/
|
|
function validateAssignment(assignment: Record<string, unknown>, fallbackModel: string): ModelAssignment | null {
|
|
const result: ModelAssignment = {};
|
|
for (const [roleId, config] of Object.entries(ROLE_REGISTRY)) {
|
|
const roleData = assignment[roleId] as Record<string, string> | undefined;
|
|
if (!roleData) {
|
|
// Backfill missing roles from the first available role or fallback
|
|
result[roleId] = {};
|
|
for (const level of config.levels) {
|
|
result[roleId][level] = fallbackModel;
|
|
}
|
|
continue;
|
|
}
|
|
result[roleId] = {};
|
|
for (const level of config.levels) {
|
|
if (!roleData[level]) {
|
|
console.error(`Missing ${roleId}.${level} in LLM assignment`);
|
|
return null;
|
|
}
|
|
result[roleId][level] = roleData[level];
|
|
}
|
|
}
|
|
return result;
|
|
}
|
|
|
|
/**
|
|
* Use an LLM to intelligently select and assign models to DevClaw roles.
|
|
*/
|
|
export async function selectModelsWithLLM(
|
|
availableModels: Array<{ model: string; provider: string }>,
|
|
sessionKey?: string,
|
|
): Promise<ModelAssignment | null> {
|
|
if (availableModels.length === 0) {
|
|
return null;
|
|
}
|
|
|
|
// If only one model, assign it to all roles
|
|
if (availableModels.length === 1) {
|
|
return singleModelAssignment(availableModels[0].model);
|
|
}
|
|
|
|
// Create a prompt for the LLM
|
|
const modelList = availableModels.map((m) => m.model).join("\n");
|
|
const jsonExample = buildJsonExample();
|
|
|
|
const prompt = `You are an AI model expert. Analyze the following authenticated AI models and assign them to DevClaw development roles based on their capabilities.
|
|
|
|
Available models:
|
|
${modelList}
|
|
|
|
All roles use the same level scheme based on task complexity:
|
|
- **senior** (most capable): Complex architecture, refactoring, critical decisions
|
|
- **medior** (balanced): Features, bug fixes, code review, standard tasks
|
|
- **junior** (fast/efficient): Simple fixes, routine tasks
|
|
|
|
Rules:
|
|
1. Prefer same provider for consistency
|
|
2. Assign most capable model to senior
|
|
3. Assign mid-tier model to medior
|
|
4. Assign fastest/cheapest model to junior
|
|
5. Consider model version numbers (higher = newer/better)
|
|
6. Stable versions (no date) > snapshot versions (with date like 20250514)
|
|
|
|
Return ONLY a JSON object in this exact format (no markdown, no explanation):
|
|
${jsonExample}`;
|
|
|
|
try {
|
|
const sessionId = sessionKey ?? "devclaw-model-selection";
|
|
|
|
const result = await runCommand(
|
|
["openclaw", "agent", "--local", "--session-id", sessionId, "--message", prompt, "--json"],
|
|
{ timeoutMs: 30_000 },
|
|
);
|
|
|
|
const output = result.stdout.trim();
|
|
|
|
// Parse the response from openclaw agent --json
|
|
const lines = output.split("\n");
|
|
const jsonStartIndex = lines.findIndex((line) => line.trim().startsWith("{"));
|
|
|
|
if (jsonStartIndex === -1) {
|
|
throw new Error("No JSON found in LLM response");
|
|
}
|
|
|
|
const jsonString = lines.slice(jsonStartIndex).join("\n");
|
|
|
|
// openclaw agent --json returns: { payloads: [{ text: "```json\n{...}\n```" }], meta: {...} }
|
|
const response = JSON.parse(jsonString);
|
|
|
|
if (!response.payloads || !Array.isArray(response.payloads) || response.payloads.length === 0) {
|
|
throw new Error("Invalid openclaw agent response structure - missing payloads");
|
|
}
|
|
|
|
// Extract text from first payload
|
|
const textContent = response.payloads[0].text;
|
|
if (!textContent) {
|
|
throw new Error("Empty text content in openclaw agent payload");
|
|
}
|
|
|
|
// Strip markdown code blocks (```json and ```)
|
|
const cleanJson = textContent
|
|
.replace(/```json\n?/g, '')
|
|
.replace(/```\n?/g, '')
|
|
.trim();
|
|
|
|
// Parse the actual model assignment JSON
|
|
const assignment = JSON.parse(cleanJson);
|
|
|
|
// Log what we got for debugging
|
|
console.log("LLM returned:", JSON.stringify(assignment, null, 2));
|
|
|
|
// Validate and backfill
|
|
const validated = validateAssignment(assignment, availableModels[0].model);
|
|
if (!validated) {
|
|
console.error("Invalid assignment structure. Got:", assignment);
|
|
throw new Error(`Invalid assignment structure from LLM. Missing fields in: ${JSON.stringify(Object.keys(assignment))}`);
|
|
}
|
|
|
|
return validated;
|
|
} catch (err) {
|
|
console.error("LLM model selection failed:", (err as Error).message);
|
|
return null;
|
|
}
|
|
}
|