| 1 | /** |
| 2 | * Token usage from model responses, read by the model proxy as they stream |
| 3 | * through it, so a run's token budget is charged by Diffinery and not trusted |
| 4 | * to the agent. Understands Anthropic Messages, OpenAI Responses and Chat |
| 5 | * Completions, streamed (SSE) or not. |
| 6 | */ |
| 7 | |
| 8 | export interface Usage { |
| 9 | /** Input tokens read without a cache. */ |
| 10 | input: number; |
| 11 | output: number; |
| 12 | cacheRead: number; |
| 13 | cacheWrite: number; |
| 14 | } |
| 15 | |
| 16 | export const NO_USAGE: Usage = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }; |
| 17 | |
| 18 | /** |
| 19 | * Tokens as a budget counts them: cache reads cost about a tenth of other |
| 20 | * input with every provider, so they count as a tenth. |
| 21 | */ |
| 22 | export const weighted = (u: Usage): number => u.input + u.cacheWrite + u.output + Math.ceil(u.cacheRead / 10); |
| 23 | |
| 24 | const n = (v: unknown): number => (typeof v === "number" && Number.isFinite(v) && v > 0 ? Math.floor(v) : 0); |
| 25 | |
| 26 | /** The usage one response object (or stream event) reports, if any. */ |
| 27 | export function usageOf(o: unknown): Usage | null { |
| 28 | if (!o || typeof o !== "object") return null; |
| 29 | const x = o as Record<string, unknown>; |
| 30 | const u = (x.usage ?? (x.message as Record<string, unknown> | undefined)?.usage ?? (x.response as Record<string, unknown> | undefined)?.usage) as Record<string, unknown> | undefined; |
| 31 | if (!u || typeof u !== "object") return null; |
| 32 | if ("prompt_tokens" in u || "completion_tokens" in u) { |
| 33 | // Chat Completions: prompt_tokens includes cached ones. |
| 34 | const cached = n((u.prompt_tokens_details as Record<string, unknown> | undefined)?.cached_tokens); |
| 35 | return { input: Math.max(n(u.prompt_tokens) - cached, 0), output: n(u.completion_tokens), cacheRead: cached, cacheWrite: 0 }; |
| 36 | } |
| 37 | const details = u.input_tokens_details as Record<string, unknown> | undefined; |
| 38 | if (details && "cached_tokens" in details) { |
| 39 | // OpenAI Responses: input_tokens includes cached ones. |
| 40 | const cached = n(details.cached_tokens); |
| 41 | return { input: Math.max(n(u.input_tokens) - cached, 0), output: n(u.output_tokens), cacheRead: cached, cacheWrite: 0 }; |
| 42 | } |
| 43 | // Anthropic: input_tokens excludes cache reads and writes. |
| 44 | return { input: n(u.input_tokens), output: n(u.output_tokens), cacheRead: n(u.cache_read_input_tokens), cacheWrite: n(u.cache_creation_input_tokens) }; |
| 45 | } |
| 46 | |
| 47 | /** Events of one response report running totals (Anthropic's message_start, then message_delta): keep the largest of each. */ |
| 48 | const fold = (a: Usage | null, b: Usage | null): Usage | null => (!a ? b : b ? { input: Math.max(a.input, b.input), output: Math.max(a.output, b.output), cacheRead: Math.max(a.cacheRead, b.cacheRead), cacheWrite: Math.max(a.cacheWrite, b.cacheWrite) } : a); |
| 49 | |
| 50 | const parse = (s: string): unknown => { |
| 51 | try { |
| 52 | return JSON.parse(s); |
| 53 | } catch { |
| 54 | return null; |
| 55 | } |
| 56 | }; |
| 57 | |
| 58 | /** The most of a non-streamed body that is read for its usage. */ |
| 59 | const MAX_JSON = 16 * 1024 * 1024; |
| 60 | |
| 61 | /** An SSE line longer than this can't be a usage event: it is dropped rather than buffered. */ |
| 62 | const MAX_SSE_LINE = 1024 * 1024; |
| 63 | |
| 64 | /** Read a response body (a tee of what goes back to the agent) for the usage it reports; null when it reports none. */ |
| 65 | export async function readUsage(body: ReadableStream<Uint8Array>, contentType: string): Promise<Usage | null> { |
| 66 | const sse = contentType.includes("text/event-stream"); |
| 67 | const reader = body.pipeThrough(new TextDecoderStream()).getReader(); |
| 68 | let usage: Usage | null = null; |
| 69 | let buf = ""; |
| 70 | let size = 0; |
| 71 | try { |
| 72 | for (;;) { |
| 73 | const { done, value } = await reader.read(); |
| 74 | if (done) break; |
| 75 | if (!sse) { |
| 76 | size += value.length; |
| 77 | if (size <= MAX_JSON) buf += value; |
| 78 | continue; |
| 79 | } |
| 80 | buf += value; |
| 81 | let nl: number; |
| 82 | while ((nl = buf.indexOf("\n")) >= 0) { |
| 83 | const line = buf.slice(0, nl).trim(); |
| 84 | buf = buf.slice(nl + 1); |
| 85 | if (line.startsWith("data:")) usage = fold(usage, usageOf(parse(line.slice(5).trim()))); |
| 86 | } |
| 87 | if (buf.length > MAX_SSE_LINE) buf = ""; |
| 88 | } |
| 89 | } finally { |
| 90 | reader.releaseLock(); |
| 91 | } |
| 92 | if (sse) return buf.trim().startsWith("data:") ? fold(usage, usageOf(parse(buf.trim().slice(5).trim()))) : usage; |
| 93 | return size <= MAX_JSON ? fold(usage, usageOf(parse(buf))) : usage; |
| 94 | } |
| 95 | |
| 96 | export const formatTokens = (t: number): string => (t >= 1_000_000 ? `${+(t / 1_000_000).toFixed(1)}M` : t >= 1000 ? `${Math.round(t / 1000)}k` : String(t)); |
| 97 | |
| 98 | /** |
| 99 | * What a CLI reports about its own usage in one line of its output (phase 5a), for runs whose model |
| 100 | * requests don't pass Diffinery: a terminal's signed-in subscription. Claude Code's `result` event |
| 101 | * gives its run's totals and cost (`set`); each Codex `turn.completed` gives one turn's tokens (`add`). |
| 102 | */ |
| 103 | export function reportedUsage(runtime: string, line: string): { usage: Usage; costUsd: number | null; mode: "set" | "add" } | null { |
| 104 | const claude = runtime === "claude-code" && line.startsWith('{"type":"result"'); |
| 105 | const codex = runtime === "codex" && line.startsWith('{"type":"turn.completed"'); |
| 106 | if (!claude && !codex) return null; |
| 107 | const j = parse(line) as Record<string, unknown> | null; |
| 108 | if (!j || typeof j !== "object") return null; |
| 109 | const u = j.usage as Record<string, unknown> | undefined; |
| 110 | if (!u || typeof u !== "object") return null; |
| 111 | if (codex) { |
| 112 | // Codex: input_tokens includes the cached ones. |
| 113 | const cached = n(u.cached_input_tokens); |
| 114 | return { usage: { input: Math.max(n(u.input_tokens) - cached, 0), output: n(u.output_tokens), cacheRead: cached, cacheWrite: 0 }, costUsd: null, mode: "add" }; |
| 115 | } |
| 116 | const cost = typeof j.total_cost_usd === "number" && Number.isFinite(j.total_cost_usd) && j.total_cost_usd >= 0 ? j.total_cost_usd : null; |
| 117 | return { usage: { input: n(u.input_tokens), output: n(u.output_tokens), cacheRead: n(u.cache_read_input_tokens), cacheWrite: n(u.cache_creation_input_tokens) }, costUsd: cost, mode: "set" }; |
| 118 | } |
| 119 | |
| 120 | /** |
| 121 | * A run's usage as people see it (phase 5a). Requests through the model proxy are counted by Diffinery |
| 122 | * (`proxy`); a run with none, in a terminal on a subscription, has only what its CLI reported (`runtime`). |
| 123 | * The cost is always the CLI's own figure (Claude Code's), never computed here. |
| 124 | */ |
| 125 | export interface RunUsage { |
| 126 | source: "proxy" | "runtime"; |
| 127 | requests: number; |
| 128 | input: number; |
| 129 | output: number; |
| 130 | cacheRead: number; |
| 131 | cacheWrite: number; |
| 132 | costUsd: number | null; |
| 133 | } |
| 134 | |
| 135 | export function runUsage(r: { |
| 136 | model_calls: number; |
| 137 | input_tokens: number; |
| 138 | output_tokens: number; |
| 139 | cache_read_tokens: number; |
| 140 | cache_write_tokens: number; |
| 141 | reported_input: number; |
| 142 | reported_output: number; |
| 143 | reported_cache_read: number; |
| 144 | reported_cache_write: number; |
| 145 | cost_usd: number | null; |
| 146 | }): RunUsage | null { |
| 147 | const costUsd = r.cost_usd ?? null; |
| 148 | if (r.model_calls > 0) return { source: "proxy", requests: r.model_calls, input: r.input_tokens, output: r.output_tokens, cacheRead: r.cache_read_tokens, cacheWrite: r.cache_write_tokens, costUsd }; |
| 149 | if (r.reported_input || r.reported_output || r.reported_cache_read || r.reported_cache_write || costUsd !== null) |
| 150 | return { source: "runtime", requests: 0, input: r.reported_input, output: r.reported_output, cacheRead: r.reported_cache_read, cacheWrite: r.reported_cache_write, costUsd }; |
| 151 | return null; |
| 152 | } |
| 153 | |
| 154 | /** Several runs' usage added up; the cost counts the runs that reported one, and is null when none did. */ |
| 155 | export function totalUsage(list: Array<RunUsage | null>): Omit<RunUsage, "source"> | null { |
| 156 | const xs = list.filter((u): u is RunUsage => !!u); |
| 157 | if (!xs.length) return null; |
| 158 | const costs = xs.map((u) => u.costUsd).filter((c): c is number => c !== null); |
| 159 | return { |
| 160 | requests: xs.reduce((a, u) => a + u.requests, 0), |
| 161 | input: xs.reduce((a, u) => a + u.input, 0), |
| 162 | output: xs.reduce((a, u) => a + u.output, 0), |
| 163 | cacheRead: xs.reduce((a, u) => a + u.cacheRead, 0), |
| 164 | cacheWrite: xs.reduce((a, u) => a + u.cacheWrite, 0), |
| 165 | costUsd: costs.length ? costs.reduce((a, c) => a + c, 0) : null, |
| 166 | }; |
| 167 | } |
| 168 | |
| 169 | /** Usage as the REST API and MCP give it, with GitHub-style snake_case names. */ |
| 170 | export const usageJson = (u: Omit<RunUsage, "source"> & { source?: RunUsage["source"] }) => ({ |
| 171 | ...(u.source ? { source: u.source } : {}), |
| 172 | requests: u.requests, |
| 173 | input_tokens: u.input, |
| 174 | output_tokens: u.output, |
| 175 | cache_read_tokens: u.cacheRead, |
| 176 | cache_write_tokens: u.cacheWrite, |
| 177 | cost_usd: u.costUsd === null ? null : Math.round(u.costUsd * 10_000) / 10_000, |
| 178 | }); |
| 179 | |
| 180 | /** A cost in dollars as a person reads it: cents, or less than a cent. */ |
| 181 | export const formatCost = (usd: number): string => (usd > 0 && usd < 0.01 ? "<$0.01" : `$${usd.toFixed(2)}`); |
| 182 | |
| 183 | /** The output an estimate assumes when a request doesn't cap it. */ |
| 184 | export const DEFAULT_MAX_OUTPUT = 32_000; |
| 185 | /** What a compaction may write back, which its request can't cap: the most any current OpenAI model emits. */ |
| 186 | export const COMPACT_MAX_OUTPUT = 128_000; |
| 187 | |
| 188 | /** Endpoints that generate tokens, so each request books an estimate, with the fields that cap their output. */ |
| 189 | const GENERATES: Record<string, string[]> = { |
| 190 | "/v1/messages": ["max_tokens"], |
| 191 | "/v1/responses": ["max_output_tokens"], |
| 192 | "/v1/responses/compact": [], |
| 193 | "/v1/chat/completions": ["max_completion_tokens", "max_tokens"], |
| 194 | // OpenRouter's legacy text completions. |
| 195 | "/v1/completions": ["max_tokens"], |
| 196 | }; |
| 197 | |
| 198 | export interface PreparedRequest { |
| 199 | /** The body to forward (changed only to ask for usage). */ |
| 200 | body: string; |
| 201 | /** What the request may cost at most, booked before it goes upstream; null for requests that generate nothing. */ |
| 202 | estimate: Usage | null; |
| 203 | /** Why a review may not send this request, if it may not. */ |
| 204 | reviewRefuse?: string; |
| 205 | /** The body a review sends instead: every output cap set to the booked one, so the estimate holds. */ |
| 206 | reviewBody?: string; |
| 207 | /** The model the request asks for (null when it names none), for runs pinned to one. */ |
| 208 | model?: string | null; |
| 209 | /** A routing field the request sets, which a pinned run may not. */ |
| 210 | routing?: string; |
| 211 | } |
| 212 | |
| 213 | /** |
| 214 | * Request fields that choose other models, providers or paid extras (OpenRouter's fallbacks and routing, |
| 215 | * plugins, web search, and more answers to pick from): a review or a run pinned to a model may not set them, |
| 216 | * or its code could switch to a far pricier model than the one its budget was set for. |
| 217 | */ |
| 218 | const ROUTING = ["models", "route", "provider", "plugins", "web_search_options", "speed", "best_of"] as const; |
| 219 | |
| 220 | /** |
| 221 | * Why a run may not send a prepared request: a review never sets routing fields, and a run pinned to a model |
| 222 | * (one that goes through OpenRouter with a model chosen in Diffinery) asks only for that model. |
| 223 | */ |
| 224 | export function pinRefusal(p: Pick<PreparedRequest, "estimate" | "model" | "routing">, run: { review: boolean; model: string | null }): string | undefined { |
| 225 | const who = run.review ? "agent reviews and other runs with a token budget" : "this task"; |
| 226 | if ((run.review || run.model) && p.routing) return `${p.routing} isn't available to ${who}: Diffinery pins the model and its routing`; |
| 227 | if (run.model && p.estimate && p.model !== run.model) return `this ${run.review ? "review" : "task"} runs on ${run.model}${p.model ? `, not ${p.model}` : ""}`; |
| 228 | return undefined; |
| 229 | } |
| 230 | |
| 231 | /** |
| 232 | * The tools a review's requests may carry: tools the agent runs itself. Function and custom tools (and Anthropic's, |
| 233 | * which have no type), and the client-run built-ins, versioned or not (bash_20250124, text_editor_20250728, OpenAI's |
| 234 | * local_shell and apply_patch). Every other tool type runs on the provider and may cost extra beyond tokens (web |
| 235 | * search, web fetch, code execution, file search, MCP, and whatever providers add next), so it is dropped, with |
| 236 | * `mcp_servers`: a review's token budget then bounds what it spends. |
| 237 | */ |
| 238 | const CLIENT_TOOL = /^(function|custom|(bash|text_editor|computer|memory|local_shell|shell|apply_patch)(_\d+)?)$/; |
| 239 | |
| 240 | const typeOf = (t: unknown): unknown => (t && typeof t === "object" ? (t as { type?: unknown }).type : undefined); |
| 241 | const clientTool = (t: unknown): boolean => { |
| 242 | const type = typeOf(t); |
| 243 | return !!t && typeof t === "object" && (type === undefined || (typeof type === "string" && CLIENT_TOOL.test(type))); |
| 244 | }; |
| 245 | /** The name a tool or a tool choice gives: Anthropic's and Responses' `name`, or Chat Completions' `function.name`. */ |
| 246 | const nameOf = (t: unknown): unknown => (t && typeof t === "object" ? ((t as { name?: unknown }).name ?? ((t as { function?: { name?: unknown } }).function?.name)) : undefined); |
| 247 | |
| 248 | /** A request's fields with only client-run tools; the same object when it has nothing else. */ |
| 249 | function withoutPaidTools(x: Record<string, unknown>): Record<string, unknown> { |
| 250 | const tools = Array.isArray(x.tools) ? (x.tools as unknown[]) : null; |
| 251 | const kept = tools?.filter(clientTool) ?? []; |
| 252 | const dropTools = !!tools && kept.length !== tools.length; |
| 253 | if (!dropTools && x.mcp_servers === undefined) return x; |
| 254 | const out: Record<string, unknown> = { ...x }; |
| 255 | delete out.mcp_servers; |
| 256 | if (dropTools) { |
| 257 | if (kept.length) out.tools = kept; |
| 258 | else delete out.tools; |
| 259 | // A tool choice that names a tool, or a tool type, that isn't kept would be refused upstream, as would any choice once no tools are left. |
| 260 | const choice = out.tool_choice; |
| 261 | if (!kept.length) delete out.tool_choice; |
| 262 | else if (choice && typeof choice === "object") { |
| 263 | const type = typeOf(choice); |
| 264 | const name = nameOf(choice); |
| 265 | const general = type === "auto" || type === "any" || type === "none"; |
| 266 | const keptName = name !== undefined && kept.some((t) => nameOf(t) === name); |
| 267 | const keptType = typeof type === "string" && type !== "function" && type !== "custom" && type !== "tool" && kept.some((t) => typeOf(t) === type); |
| 268 | if (!general && !keptName && !keptType) delete out.tool_choice; |
| 269 | } |
| 270 | } |
| 271 | return out; |
| 272 | } |
| 273 | |
| 274 | /** Responses API fields that pull in context stored at the provider, which the request's size doesn't show. */ |
| 275 | const STORED_CONTEXT = ["previous_response_id", "conversation", "prompt"] as const; |
| 276 | |
| 277 | /** |
| 278 | * Look at a model request before it goes upstream, so that every answer can |
| 279 | * be charged: its estimate (a token per 3 bytes of request, plus the output |
| 280 | * it allows) is booked up front and settled to the usage the answer reports, |
| 281 | * and streamed Chat Completions are asked to include usage. A review may not |
| 282 | * start a background response, which it could leave unfetched and so |
| 283 | * uncharged, pull in stored context its estimate can't see, or ask for more |
| 284 | * than one answer; a review's request that doesn't cap its output is |
| 285 | * capped at the estimate's DEFAULT_MAX_OUTPUT, and paid provider-side tools |
| 286 | * are stripped from it. A compaction can't be capped, |
| 287 | * so it books COMPACT_MAX_OUTPUT. |
| 288 | */ |
| 289 | export function prepareRequest(method: string, path: string, body: string, size: number): PreparedRequest { |
| 290 | const caps = method === "POST" && Object.hasOwn(GENERATES, path) ? GENERATES[path]! : null; |
| 291 | if (!caps) return { body, estimate: null }; |
| 292 | const json = parse(body); |
| 293 | const x = json && typeof json === "object" && !Array.isArray(json) ? (json as Record<string, unknown>) : null; |
| 294 | // A request can set several caps (Chat Completions takes two): book the largest. |
| 295 | const capped = x ? caps.map((k) => n(x[k])).filter((v) => v > 0) : []; |
| 296 | const cap = capped.length ? Math.max(...capped) : undefined; |
| 297 | const estimate: Usage = { input: Math.ceil(size / 3), output: cap ?? (caps.length ? DEFAULT_MAX_OUTPUT : COMPACT_MAX_OUTPUT), cacheRead: 0, cacheWrite: 0 }; |
| 298 | if (!x) return { body, estimate }; |
| 299 | const stored = path.startsWith("/v1/responses") ? STORED_CONTEXT.find((k) => x[k] != null) : undefined; |
| 300 | const reviewRefuse = |
| 301 | x.background === true |
| 302 | ? "background responses aren't available to agent reviews and other runs with a token budget" |
| 303 | : stored |
| 304 | ? `${stored} isn't available to agent reviews and other runs with a token budget: send the whole input instead` |
| 305 | : x.n != null && x.n !== 1 |
| 306 | ? "more than one answer per request (n) isn't available to agent reviews and other runs with a token budget" |
| 307 | : undefined; |
| 308 | let fields = x; |
| 309 | if (path === "/v1/chat/completions" && x.stream === true) { |
| 310 | const so = x.stream_options && typeof x.stream_options === "object" ? (x.stream_options as Record<string, unknown>) : {}; |
| 311 | if (so.include_usage !== true) fields = { ...x, stream_options: { ...so, include_usage: true } }; |
| 312 | } |
| 313 | const out: PreparedRequest = { body: fields === x ? body : JSON.stringify(fields), estimate, model: typeof x.model === "string" ? x.model : null }; |
| 314 | if (reviewRefuse) out.reviewRefuse = reviewRefuse; |
| 315 | const routing = ROUTING.find((k) => x[k] != null); |
| 316 | if (routing) out.routing = routing; |
| 317 | // A review sends every cap as the one booked (the default when it sets none), and no paid provider-side tools. |
| 318 | const set = caps.filter((k) => x[k] !== undefined); |
| 319 | const recap = caps[0] && (!set.length || set.some((k) => x[k] !== estimate.output)); |
| 320 | const unpaid = withoutPaidTools(fields); |
| 321 | if (recap || unpaid !== fields) out.reviewBody = JSON.stringify(recap ? { ...unpaid, ...Object.fromEntries((set.length ? set : [caps[0]!]).map((k) => [k, estimate.output])) } : unpaid); |
| 322 | return out; |
| 323 | } |
| 324 | |
| 325 | /** Read a request body up to `max` bytes, counting as it reads; null when it is larger. */ |
| 326 | export async function readBody(body: ReadableStream<Uint8Array> | null, max: number): Promise<Uint8Array | null> { |
| 327 | if (!body) return new Uint8Array(); |
| 328 | const chunks: Uint8Array[] = []; |
| 329 | let size = 0; |
| 330 | const reader = body.getReader(); |
| 331 | try { |
| 332 | for (;;) { |
| 333 | const { done, value } = await reader.read(); |
| 334 | if (done) break; |
| 335 | size += value.byteLength; |
| 336 | if (size > max) { |
| 337 | await reader.cancel().catch(() => undefined); |
| 338 | return null; |
| 339 | } |
| 340 | chunks.push(value); |
| 341 | } |
| 342 | } finally { |
| 343 | reader.releaseLock(); |
| 344 | } |
| 345 | const out = new Uint8Array(size); |
| 346 | let at = 0; |
| 347 | for (const c of chunks) { |
| 348 | out.set(c, at); |
| 349 | at += c.byteLength; |
| 350 | } |
| 351 | return out; |
| 352 | } |