Files
fusion/packages/dashboard/src/agent-generation.ts
Fusion ed2df90df6 feat(FN-2615): merge fusion/fn-2615 (auto-resolved)
- fix(FN-2615): restore workspace lint and test stability
2026-04-26 15:16:22 -07:00

602 lines
19 KiB
TypeScript

/**
* Agent Generation Session Management
*
* Manages AI-guided sessions for generating agent specifications from role descriptions.
* Sessions are stored in-memory with TTL cleanup.
*
* Pattern follows planning.ts for consistency.
*
* Features:
* - AI agent integration with streaming via callbacks
* - Rate limiting per IP
* - Session expiration and cleanup
*/
import { randomUUID } from "node:crypto";
import { createSessionDiagnostics, nonfatal } from "./ai-session-diagnostics.js";
// Dynamic import for @fusion/core to get prompt override resolution
type PromptOverrideMap = Record<string, string | null>;
type ResolvePromptFn = (key: string, overrides?: PromptOverrideMap) => string;
let resolvePrompt: ResolvePromptFn = () => "";
let promptCatalogReady = false;
async function initPromptCatalog() {
if (promptCatalogReady) return;
try {
const core = await import("@fusion/core");
resolvePrompt = (key: string, overrides?: PromptOverrideMap) =>
core.resolvePrompt(key as keyof typeof core.PROMPT_KEY_CATALOG, overrides);
promptCatalogReady = true;
} catch {
// Use fallback resolution when core is unavailable
resolvePrompt = () => "";
promptCatalogReady = true;
}
}
// Initialize prompt catalog (will be awaited in actual usage)
const promptCatalogReadyPromise = initPromptCatalog();
import { createFnAgent as engineCreateFnAgent } from "@fusion/engine";
// eslint-disable-next-line @typescript-eslint/no-explicit-any
const createFnAgent: any = engineCreateFnAgent;
function ensureEngineReady(): Promise<void> {
return Promise.resolve();
}
// ── Constants ───────────────────────────────────────────────────────────────
/** System prompt for the AI agent that generates agent specifications */
export const AGENT_GENERATION_SYSTEM_PROMPT = `You are an agent specification generator for the fn task board system.
Your job: given a user-provided role description, generate a complete agent specification suitable for creating an AI agent.
## Input
The user will provide a role description like:
- "Senior frontend code reviewer who specializes in React accessibility"
- "Security-focused DevOps engineer"
- "Performance optimization specialist for Node.js applications"
## Output
You MUST respond with ONLY valid JSON (no markdown, no explanation):
{
"title": "A concise display name (max 60 chars)",
"icon": "A single emoji representing the agent",
"role": "The most appropriate capability: triage | executor | reviewer | merger | scheduler | engineer | custom",
"description": "A brief 1-2 sentence description of the agent's purpose and expertise",
"systemPrompt": "A detailed markdown system prompt for the agent. This should be comprehensive and include:\\n- Role definition\\n- Core responsibilities\\n- Specific areas of expertise\\n- Behavioral guidelines\\n- Output format expectations\\n- Edge case handling instructions",
"thinkingLevel": "off | minimal | low | medium | high",
"maxTurns": 1000
}
## Guidelines for System Prompt Generation
- Be specific about the agent's domain expertise
- Include concrete behavioral rules and constraints
- Define the expected output format clearly
- Add error handling and edge case guidance
- Keep the prompt focused and actionable (aim for 200-800 words)
- Use markdown formatting for readability
## Thinking Level Guidelines
- "off": For simple, well-defined tasks (basic CRUD, simple checks)
- "minimal": For straightforward tasks requiring some reasoning
- "low": For moderate complexity tasks
- "medium": For complex analysis, code review, architecture decisions
- "high": For critical decisions, security analysis, complex debugging
## Max Turns Guidelines
- 5-10: Simple, focused tasks (quick reviews, status checks)
- 10-25: Standard tasks (code review, feature planning)
- 25-50: Complex tasks (multi-file changes, architecture analysis)
- 50+: Extended tasks (large refactors, comprehensive audits)
## Role Selection Guidelines
- "reviewer": Agents focused on reviewing, auditing, analyzing
- "executor": Agents that perform implementation work
- "engineer": Agents that do engineering work with broader scope
- "triage": Agents focused on classification and routing
- "custom": Any agent that doesn't fit standard roles
- Default to "custom" if unclear`;
/** Session TTL in milliseconds (30 minutes) */
const SESSION_TTL_MS = 30 * 60 * 1000;
/** Cleanup interval in milliseconds (5 minutes) */
const CLEANUP_INTERVAL_MS = 5 * 60 * 1000;
/** Max agent generation sessions per IP per hour */
const MAX_SESSIONS_PER_IP_PER_HOUR = 10;
/** Rate limiting window in milliseconds (1 hour) */
const RATE_LIMIT_WINDOW_MS = 60 * 60 * 1000;
// ── Types ───────────────────────────────────────────────────────────────────
/** Generated agent specification returned by the AI */
export interface AgentGenerationSpec {
/** Display name for the agent */
title: string;
/** Single emoji icon */
icon: string;
/** Agent capability/role */
role: string;
/** Brief description of the agent's purpose */
description: string;
/** Detailed system prompt in markdown */
systemPrompt: string;
/** Suggested thinking level */
thinkingLevel: "off" | "minimal" | "low" | "medium" | "high";
/** Suggested max turns (1-500) */
maxTurns: number;
}
/** Public state of an agent generation session (no sensitive fields like IP) */
export interface AgentGenerationSession {
id: string;
roleDescription: string;
spec?: AgentGenerationSpec;
createdAt: Date;
updatedAt: Date;
}
// ── Internal Types ──────────────────────────────────────────────────────────
interface Session {
id: string;
ip: string;
roleDescription: string;
spec?: AgentGenerationSpec;
createdAt: Date;
updatedAt: Date;
}
interface RateLimitEntry {
count: number;
firstRequestAt: Date;
}
// ── In-Memory Storage ───────────────────────────────────────────────────────
/** Active agent generation sessions indexed by session ID */
const sessions = new Map<string, Session>();
/** Rate limiting state indexed by IP */
const rateLimits = new Map<string, RateLimitEntry>();
/** Shared diagnostics helper for structured agent-generation telemetry. */
const diagnostics = createSessionDiagnostics("agent-generation");
// ── Cleanup Interval ────────────────────────────────────────────────────────
/**
* Remove expired sessions and stale rate limit entries.
*/
function cleanupExpiredSessions(): void {
const now = Date.now();
let cleanedSessions = 0;
let cleanedRateLimits = 0;
for (const [id, session] of sessions) {
if (now - session.updatedAt.getTime() > SESSION_TTL_MS) {
sessions.delete(id);
cleanedSessions++;
}
}
for (const [ip, entry] of rateLimits) {
if (now - entry.firstRequestAt.getTime() > RATE_LIMIT_WINDOW_MS) {
rateLimits.delete(ip);
cleanedRateLimits++;
}
}
if (cleanedSessions > 0 || cleanedRateLimits > 0) {
diagnostics.info("Cleanup completed", {
cleanedSessions,
cleanedRateLimits,
ttlMs: SESSION_TTL_MS,
rateLimitWindowMs: RATE_LIMIT_WINDOW_MS,
operation: "cleanup-expired",
});
}
}
/**
* Run expiry cleanup immediately. Test-only helper for deterministic assertions.
* @internal
*/
export function __runAgentGenerationCleanupForTests(): void {
cleanupExpiredSessions();
}
const cleanupInterval = setInterval(cleanupExpiredSessions, CLEANUP_INTERVAL_MS);
cleanupInterval.unref?.();
process.on("beforeExit", () => {
clearInterval(cleanupInterval);
});
// ── Rate Limiting ───────────────────────────────────────────────────────────
/**
* Check if IP can create a new agent generation session.
* Returns true if allowed, false if rate limited.
*/
export function checkRateLimit(ip: string): boolean {
const now = Date.now();
const entry = rateLimits.get(ip);
if (!entry) {
rateLimits.set(ip, { count: 1, firstRequestAt: new Date() });
return true;
}
if (now - entry.firstRequestAt.getTime() > RATE_LIMIT_WINDOW_MS) {
rateLimits.set(ip, { count: 1, firstRequestAt: new Date() });
return true;
}
if (entry.count >= MAX_SESSIONS_PER_IP_PER_HOUR) {
return false;
}
entry.count++;
return true;
}
/**
* Get rate limit reset time for an IP.
* Returns null if no rate limit entry exists.
*/
export function getRateLimitResetTime(ip: string): Date | null {
const entry = rateLimits.get(ip);
if (!entry) return null;
return new Date(entry.firstRequestAt.getTime() + RATE_LIMIT_WINDOW_MS);
}
// ── JSON Extraction ─────────────────────────────────────────────────────────
/**
* Extract JSON candidate from AI response text.
* Handles markdown code blocks and embedded JSON.
*/
function extractJsonCandidate(text: string): string | null {
if (!text || !text.trim()) return null;
// 1. Try markdown code blocks
const codeBlockMatch = text.match(/```(?:json)?\s*([\s\S]*?)\s*```/);
if (codeBlockMatch?.[1]) {
const candidate = codeBlockMatch[1].trim();
if (candidate.startsWith("{")) return candidate;
}
// 2. Find balanced brace-delimited objects
const candidates: Array<{ text: string }> = [];
for (let i = 0; i < text.length; i++) {
if (text[i] === "{") {
let depth = 0;
let inString = false;
let escape = false;
for (let j = i; j < text.length; j++) {
const ch = text[j];
if (escape) { escape = false; continue; }
if (ch === "\\") { escape = true; continue; }
if (ch === '"') { inString = !inString; continue; }
if (inString) continue;
if (ch === "{") depth++;
if (ch === "}") depth--;
if (depth === 0) {
const candidate = text.slice(i, j + 1).trim();
try {
JSON.parse(candidate);
candidates.push({ text: candidate });
} catch {
// Not valid JSON, skip
}
break;
}
}
}
}
if (candidates.length > 0) {
candidates.sort((a, b) => b.text.length - a.text.length);
return candidates[0].text;
}
// 3. Last resort: try the full trimmed text
const trimmed = text.trim();
if (trimmed.startsWith("{")) return trimmed;
return null;
}
/**
* Attempt to repair common JSON issues (truncated, trailing commas, etc.).
*/
function repairJson(text: string): string {
let repaired = text;
repaired = repaired.replace(/,\s*([}\]])/g, "$1");
let openBraces = 0;
let openBrackets = 0;
let inString = false;
let escape = false;
for (const ch of repaired) {
if (escape) { escape = false; continue; }
if (ch === "\\") { escape = true; continue; }
if (ch === '"') { inString = !inString; continue; }
if (inString) continue;
if (ch === "{") openBraces++;
if (ch === "}") openBraces--;
if (ch === "[") openBrackets++;
if (ch === "]") openBrackets--;
}
if (inString) repaired += '"';
// Re-count after potential string fix
openBraces = 0;
openBrackets = 0;
inString = false;
escape = false;
for (const ch of repaired) {
if (escape) { escape = false; continue; }
if (ch === "\\") { escape = true; continue; }
if (ch === '"') { inString = !inString; continue; }
if (inString) continue;
if (ch === "{") openBraces++;
if (ch === "}") openBraces--;
if (ch === "[") openBrackets++;
if (ch === "]") openBrackets--;
}
repaired += "]".repeat(Math.max(0, openBrackets));
repaired += "}".repeat(Math.max(0, openBraces));
return repaired;
}
/**
* Parse the AI response text into an AgentGenerationSpec.
*/
export function parseGenerationResponse(text: string): AgentGenerationSpec {
const candidate = extractJsonCandidate(text);
if (!candidate) {
throw new Error("AI returned no valid JSON. Please try again.");
}
let parsed: unknown;
try {
parsed = JSON.parse(candidate);
} catch {
try {
const repaired = repairJson(candidate);
parsed = JSON.parse(repaired);
} catch (repairErr) {
throw new Error(
`Failed to parse AI response: ${repairErr instanceof Error ? repairErr.message : "Unknown error"}. Please try again.`
);
}
}
if (typeof parsed !== "object" || parsed === null) {
throw new Error("AI returned an invalid response structure. Please try again.");
}
const obj = parsed as Record<string, unknown>;
// Validate required fields with defaults
return {
title: typeof obj.title === "string" ? obj.title.slice(0, 60) : "Custom Agent",
icon: typeof obj.icon === "string" ? obj.icon : "🤖",
role: typeof obj.role === "string" ? obj.role : "custom",
description: typeof obj.description === "string" ? obj.description : "",
systemPrompt: typeof obj.systemPrompt === "string" ? obj.systemPrompt : "",
thinkingLevel: ["off", "minimal", "low", "medium", "high"].includes(obj.thinkingLevel as string)
? (obj.thinkingLevel as AgentGenerationSpec["thinkingLevel"])
: "off",
maxTurns: typeof obj.maxTurns === "number"
? Math.max(1, Math.min(2000, Math.round(obj.maxTurns)))
: 1000,
};
}
// ── Session Management ──────────────────────────────────────────────────────
/**
* Start a new agent generation session.
* Creates the session in memory but does not yet generate the spec.
* Call `generateAgentSpec()` to trigger AI generation.
*
* @param ip - Client IP for rate limiting
* @param roleDescription - The user's description of the desired agent role
* @returns Session object (without spec - call generateAgentSpec to populate)
*/
export async function startAgentGeneration(
ip: string,
roleDescription: string,
): Promise<AgentGenerationSession> {
if (!checkRateLimit(ip)) {
const resetTime = getRateLimitResetTime(ip);
throw new RateLimitError(
`Rate limit exceeded. Maximum ${MAX_SESSIONS_PER_IP_PER_HOUR} generation sessions per hour. ` +
`Reset at ${resetTime?.toISOString() || "unknown"}`
);
}
const sessionId = randomUUID();
const session: Session = {
id: sessionId,
ip,
roleDescription,
createdAt: new Date(),
updatedAt: new Date(),
};
sessions.set(sessionId, session);
return toPublicSession(session);
}
/**
* Generate the agent specification for an existing session using AI.
* This calls the AI model with the session's role description and populates
* the session's spec field.
*
* @param sessionId - The session identifier
* @param rootDir - Project root directory for AI agent context
* @param promptOverrides - Optional prompt overrides from project settings
* @returns The generated agent specification
*/
export async function generateAgentSpec(
sessionId: string,
rootDir: string,
promptOverrides?: PromptOverrideMap
): Promise<AgentGenerationSpec> {
const session = sessions.get(sessionId);
if (!session) {
throw new SessionNotFoundError(`Agent generation session ${sessionId} not found or expired`);
}
try {
await ensureEngineReady();
await promptCatalogReadyPromise;
const spec = await generateSpecWithAI(session, rootDir, promptOverrides);
session.spec = spec;
session.updatedAt = new Date();
return spec;
} catch (err) {
diagnostics.errorFromException("AI generation failed for session", err, {
sessionId,
operation: "generate-agent-spec",
});
throw err;
}
}
/**
* Generate an agent specification using the AI agent.
*/
async function generateSpecWithAI(
session: Session,
rootDir: string,
promptOverrides?: PromptOverrideMap
): Promise<AgentGenerationSpec> {
if (!createFnAgent) {
throw new Error("AI agent not available. Ensure the engine is properly configured.");
}
// Resolve the system prompt using prompt overrides (with fallback to default)
const effectiveSystemPrompt = resolvePrompt("agent-generation-system", promptOverrides) || AGENT_GENERATION_SYSTEM_PROMPT;
const agent = await createFnAgent({
cwd: rootDir,
systemPrompt: effectiveSystemPrompt,
tools: "none",
});
try {
await agent.session.prompt(
`Generate an agent specification for the following role:\n\n${session.roleDescription}`
);
// Extract response text
interface AgentMessage {
role: string;
content?: string | Array<{ type: string; text: string }>;
}
const lastMessage = (agent.session.state.messages as AgentMessage[])
.filter((m: AgentMessage) => m.role === "assistant")
.pop();
let responseText = "";
if (lastMessage?.content) {
if (typeof lastMessage.content === "string") {
responseText = lastMessage.content;
} else if (Array.isArray(lastMessage.content)) {
responseText = lastMessage.content
.filter((c: { type: string; text: string }): c is { type: "text"; text: string } => c.type === "text")
.map((c: { type: string; text: string }) => c.text)
.join("");
}
}
return parseGenerationResponse(responseText);
} finally {
nonfatal(
() => agent.session.dispose?.(),
diagnostics,
"Failed to dispose agent-generation session",
{
sessionId: session.id,
operation: "dispose-agent-session",
}
);
}
}
/**
* Get a session by ID.
*
* @param sessionId - The session identifier
* @returns The session, or undefined if not found
*/
export function getAgentGenerationSession(sessionId: string): AgentGenerationSession | undefined {
const session = sessions.get(sessionId);
if (!session) return undefined;
return toPublicSession(session);
}
/**
* Clean up and remove a session.
*
* @param sessionId - The session identifier
*/
export function cleanupAgentGenerationSession(sessionId: string): void {
sessions.delete(sessionId);
}
/**
* Convert internal session to public session type.
*/
function toPublicSession(session: Session): AgentGenerationSession {
return {
id: session.id,
roleDescription: session.roleDescription,
spec: session.spec,
createdAt: session.createdAt,
updatedAt: session.updatedAt,
};
}
/**
* Reset all agent generation state. Used for testing only.
*/
export function __resetAgentGenerationState(): void {
sessions.clear();
rateLimits.clear();
}
// ── Custom Errors ───────────────────────────────────────────────────────────
export class RateLimitError extends Error {
constructor(message: string) {
super(message);
this.name = "RateLimitError";
}
}
export class SessionNotFoundError extends Error {
constructor(message: string) {
super(message);
this.name = "SessionNotFoundError";
}
}