2026-03-18 14:59:02 +00:00
|
|
|
import { anthropic } from "@workspace/integrations-anthropic-ai";
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
import type Anthropic from "@anthropic-ai/sdk";
|
2026-03-18 14:59:02 +00:00
|
|
|
|
|
|
|
|
export interface EvalResult {
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
accepted: boolean;
|
2026-03-18 14:59:02 +00:00
|
|
|
reason: string;
|
|
|
|
|
}
|
|
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
export interface WorkResult {
|
|
|
|
|
result: string;
|
|
|
|
|
}
|
2026-03-18 14:59:02 +00:00
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
export interface AgentConfig {
|
|
|
|
|
evalModel?: string;
|
|
|
|
|
workModel?: string;
|
|
|
|
|
}
|
2026-03-18 14:59:02 +00:00
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
export class AgentService {
|
|
|
|
|
private readonly evalModel: string;
|
|
|
|
|
private readonly workModel: string;
|
2026-03-18 14:59:02 +00:00
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
constructor(config?: AgentConfig) {
|
|
|
|
|
this.evalModel = config?.evalModel ?? "claude-haiku-4-5";
|
|
|
|
|
this.workModel = config?.workModel ?? "claude-sonnet-4-6";
|
2026-03-18 14:59:02 +00:00
|
|
|
}
|
|
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
async evaluateRequest(requestText: string): Promise<EvalResult> {
|
|
|
|
|
const message = await anthropic.messages.create({
|
|
|
|
|
model: this.evalModel,
|
|
|
|
|
max_tokens: 8192,
|
|
|
|
|
system: `You are Timmy, an AI agent gatekeeper. Evaluate whether a request is acceptable to act on.
|
|
|
|
|
ACCEPT if the request is: clear enough to act on, ethical, lawful, and within the capability of a general-purpose AI.
|
|
|
|
|
REJECT if the request is: harmful, illegal, unethical, incoherent, or spam.
|
|
|
|
|
Respond ONLY with valid JSON: {"accepted": true, "reason": "..."} or {"accepted": false, "reason": "..."}`,
|
|
|
|
|
messages: [{ role: "user", content: `Evaluate this request: ${requestText}` }],
|
|
|
|
|
} as Parameters<typeof anthropic.messages.create>[0]);
|
2026-03-18 14:59:02 +00:00
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
const block = message.content[0] as Anthropic.TextBlock;
|
|
|
|
|
if (block.type !== "text") {
|
|
|
|
|
throw new Error("Unexpected non-text response from eval model");
|
|
|
|
|
}
|
2026-03-18 14:59:02 +00:00
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
let parsed: { accepted: boolean; reason: string };
|
|
|
|
|
try {
|
|
|
|
|
const raw = block.text.replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/, "").trim();
|
|
|
|
|
parsed = JSON.parse(raw) as { accepted: boolean; reason: string };
|
|
|
|
|
} catch {
|
|
|
|
|
throw new Error(`Failed to parse eval JSON: ${block.text}`);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return { accepted: Boolean(parsed.accepted), reason: parsed.reason ?? "" };
|
2026-03-18 14:59:02 +00:00
|
|
|
}
|
|
|
|
|
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
async executeWork(requestText: string): Promise<WorkResult> {
|
|
|
|
|
const message = await anthropic.messages.create({
|
|
|
|
|
model: this.workModel,
|
|
|
|
|
max_tokens: 8192,
|
|
|
|
|
system: `You are Timmy, a capable AI agent. A user has paid for you to handle their request.
|
|
|
|
|
Fulfill it thoroughly and helpfully. Be concise yet complete.`,
|
|
|
|
|
messages: [{ role: "user", content: requestText }],
|
|
|
|
|
} as Parameters<typeof anthropic.messages.create>[0]);
|
|
|
|
|
|
|
|
|
|
const block = message.content[0] as Anthropic.TextBlock;
|
|
|
|
|
if (block.type !== "text") {
|
|
|
|
|
throw new Error("Unexpected non-text response from work model");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return { result: block.text };
|
|
|
|
|
}
|
2026-03-18 14:59:02 +00:00
|
|
|
}
|
Task #2: MVP Foundation — injectable services, DB schema, smoke test
DB schema
- jobs and invoices tables added to lib/db/src/schema/
- schema barrel updated (jobs, invoices, conversations, messages)
- pnpm --filter @workspace/db run push applied successfully
LNbitsService (artifacts/api-server/src/lib/lnbits.ts)
- Injectable class accepting optional { url, apiKey } config
- Falls back to LNBITS_URL / LNBITS_API_KEY env vars
- Auto-detects stub mode when credentials are absent; logs warning
- createInvoice() -> { paymentHash, paymentRequest }
- checkInvoicePaid() -> boolean
- stubMarkPaid() helper for dev/test flows
- Real LNbits REST v1 calls wired behind the stub guard
AgentService (artifacts/api-server/src/lib/agent.ts)
- Injectable class with configurable evalModel / workModel
- evaluateRequest(text) -> { accepted: boolean, reason: string }
uses claude-haiku-4-5; strips markdown fences before JSON parse
- executeWork(text) -> { result: string } uses claude-sonnet-4-6
- Wired via Replit Anthropic AI Integration (no user API key)
PricingService (artifacts/api-server/src/lib/pricing.ts)
- Injectable class with configurable fee/bucket thresholds
- calculateEvalFeeSats() -> 10 sats (fixed)
- calculateWorkFeeSats(text) -> 50/100/250 by char-length bucket
- Zero LLM involvement; fully deterministic
Smoke test (scripts/src/smoke.ts)
- pnpm --filter @workspace/scripts run smoke
- Verifies LNbits stub: create, check unpaid, mark paid, check paid
- Verifies Anthropic: evaluateRequest round-trip
- Both checks passed
replit.md
- Documented required (LNBITS_URL, LNBITS_API_KEY) and auto-provisioned secrets
- Stub-mode behaviour explained
2026-03-18 15:14:23 +00:00
|
|
|
|
|
|
|
|
export const agentService = new AgentService();
|