oxedyne/daimond/dev/probe_provider.mjs
4.6 KiB, 1 run
created by r2519314175:65, which is this file's identity for as long as the history lasts, whatever it is later renamed to
download · who wrote it · its history
| 1 | // Replay a real agentic loop's requests at a REAL provider, and read the clock. |
| 2 | // |
| 3 | // The mock can say what Daimond sends and what Daimond's own overhead costs. It |
| 4 | // cannot say what the provider does with a 57 KB request twelve times in a row -- |
| 5 | // whether the prefix is served from a cache, and what the first token costs when |
| 6 | // it is not. So the bodies captured by dev/latmock.mjs (LAT_KEEP_BODY=1) are |
| 7 | // replayed verbatim, in order, with only the model swapped. |
| 8 | // |
| 9 | // node dev/probe_provider.mjs <captured.log> <model> [--rounds 13] [--pass 2] |
| 10 | // |
| 11 | // The key is read from the environment and from nowhere else: |
| 12 | // |
| 13 | // OPENROUTER_API_KEY=$(cat ~/.config/oxedyne/daimond/openrouter.key) \ |
| 14 | // node dev/probe_provider.mjs dev/latmock-x.log z-ai/glm-5.2 |
| 15 | // |
| 16 | // Every round reports time to first token, total time, the prompt tokens the |
| 17 | // provider counted, how many of them it served from its cache, and what it says |
| 18 | // the call cost. Nothing is written to the log but numbers. |
| 19 | |
| 20 | import fs from 'node:fs'; |
| 21 | |
| 22 | const LOG = process.argv[2]; |
| 23 | const MODEL = process.argv[3]; |
| 24 | const arg = (n, d) => { const i = process.argv.indexOf('--' + n); return i === -1 ? d : process.argv[i + 1]; }; |
| 25 | const ROUNDS = Number(arg('rounds', 99)); |
| 26 | const PASSES = Number(arg('pass', 1)); |
| 27 | const MAXTOK = Number(arg('maxtok', 200)); |
| 28 | const URL_ = arg('url', 'https://openrouter.ai/api/v1/chat/completions'); |
| 29 | |
| 30 | if (!LOG || !MODEL) { |
| 31 | console.error('usage: probe_provider.mjs <captured.log> <model> [--rounds n] [--pass n]'); |
| 32 | process.exit(2); |
| 33 | } |
| 34 | const KEY = process.env.OPENROUTER_API_KEY; |
| 35 | if (!KEY) { |
| 36 | console.error('Set OPENROUTER_API_KEY before running this probe.'); |
| 37 | process.exit(2); |
| 38 | } |
| 39 | |
| 40 | const bodies = fs.readFileSync(LOG, 'utf8').trim().split('\n') |
| 41 | .filter(Boolean).map(l => JSON.parse(l)).filter(l => l.raw).map(l => JSON.parse(l.raw)) |
| 42 | .slice(0, ROUNDS); |
| 43 | if (!bodies.length) { |
| 44 | console.error(`${LOG} holds no captured bodies -- run the probe with LAT_KEEP_BODY=1.`); |
| 45 | process.exit(2); |
| 46 | } |
| 47 | |
| 48 | const now = () => Number(process.hrtime.bigint() / 1000n) / 1000; |
| 49 | |
| 50 | /// One replayed round: the same messages and the same tool schemas, at a real model. |
| 51 | async function round(body) { |
| 52 | const send = { |
| 53 | ...body, |
| 54 | model: MODEL, |
| 55 | max_tokens: MAXTOK, |
| 56 | stream: true, |
| 57 | stream_options: { include_usage: true }, |
| 58 | // OpenRouter reports what the call cost in the final usage chunk when asked. |
| 59 | usage: { include: true }, |
| 60 | }; |
| 61 | const raw = JSON.stringify(send); |
| 62 | const t0 = now(); |
| 63 | const r = await fetch(URL_, { |
| 64 | method: 'POST', |
| 65 | headers: { 'Content-Type': 'application/json', 'Authorization': `Bearer ${KEY}` }, |
| 66 | body: raw, |
| 67 | }); |
| 68 | if (!r.ok) { |
| 69 | const t = await r.text(); |
| 70 | return { err: `${r.status} ${t.slice(0, 200)}` }; |
| 71 | } |
| 72 | let first = 0, usage = null, out = 0; |
| 73 | const reader = r.body.getReader(); |
| 74 | const dec = new TextDecoder(); |
| 75 | let buf = ''; |
| 76 | for (;;) { |
| 77 | const { done, value } = await reader.read(); |
| 78 | if (done) break; |
| 79 | if (!first) first = now(); |
| 80 | buf += dec.decode(value, { stream: true }); |
| 81 | let nl; |
| 82 | while ((nl = buf.indexOf('\n')) !== -1) { |
| 83 | const line = buf.slice(0, nl).trim(); |
| 84 | buf = buf.slice(nl + 1); |
| 85 | if (!line.startsWith('data: ')) continue; |
| 86 | const d = line.slice(6); |
| 87 | if (d === '[DONE]') continue; |
| 88 | let j; try { j = JSON.parse(d); } catch { continue; } |
| 89 | if (j.usage) usage = j.usage; |
| 90 | const ch = (j.choices || [])[0]; |
| 91 | if (ch && ch.delta) out += ((ch.delta.content || '') + (ch.delta.reasoning || '')).length; |
| 92 | } |
| 93 | } |
| 94 | const end = now(); |
| 95 | const u = usage || {}; |
| 96 | const det = u.prompt_tokens_details || {}; |
| 97 | return { |
| 98 | bytes: Buffer.byteLength(raw, 'utf8'), |
| 99 | ttft: +(first - t0).toFixed(0), |
| 100 | total: +(end - t0).toFixed(0), |
| 101 | ptok: u.prompt_tokens || 0, |
| 102 | cached: det.cached_tokens || 0, |
| 103 | ctok: u.completion_tokens || 0, |
| 104 | cost: u.cost || 0, |
| 105 | chars: out, |
| 106 | }; |
| 107 | } |
| 108 | |
| 109 | console.log(`model ${MODEL} rounds ${bodies.length} pass(es) ${PASSES}`); |
| 110 | let spend = 0; |
| 111 | for (let p = 1; p <= PASSES; p++) { |
| 112 | console.log(`\npass ${p}`); |
| 113 | console.log('round reqKB promptTok cached %cached ttft total outTok cost'); |
| 114 | for (let i = 0; i < bodies.length; i++) { |
| 115 | const r = await round(bodies[i]); |
| 116 | if (r.err) { console.log(`${String(i + 1).padStart(5)} ERROR ${r.err}`); continue; } |
| 117 | spend += r.cost; |
| 118 | const pc = r.ptok ? (100 * r.cached / r.ptok) : 0; |
| 119 | console.log([ |
| 120 | String(i + 1).padStart(5), |
| 121 | (r.bytes / 1024).toFixed(1).padStart(7), |
| 122 | String(r.ptok).padStart(11), |
| 123 | String(r.cached).padStart(8), |
| 124 | pc.toFixed(0).padStart(8), |
| 125 | String(r.ttft).padStart(6), |
| 126 | String(r.total).padStart(7), |
| 127 | String(r.ctok).padStart(8), |
| 128 | r.cost.toFixed(5).padStart(8), |
| 129 | ].join('')); |
| 130 | } |
| 131 | } |
| 132 | console.log(`\nspent on ${MODEL}: $${spend.toFixed(4)}`); |