| 1 | // Run: |
| 2 | // corepack pnpm --dir desktop/frontend exec tsx src/__tests__/history-performance-benchmark.tsx |
| 3 | // |
| 4 | // Synthetic, privacy-safe benchmark for long restored histories. It logs counts, |
| 5 | // byte lengths, and elapsed times only; it never uses real conversation content. |
| 6 | |
| 7 | import { |
| 8 | noteActivationRequested, noteActivationSettled, noteActivationStarted, |
| 9 | noteResumeHistoryPage, noteTranscriptRowCounts, resetSessionDiagnostics, sessionPipelineDiagnostics, |
| 10 | } from "../lib/sessionDiagnostics"; |
| 11 | import { historyMessagesToItems, initialState, reducer, type Item } from "../lib/useController"; |
| 12 | import type { HistoryMessage } from "../lib/types"; |
| 13 | import { ChatSource } from "../lib/chatViewSource"; |
| 14 | |
| 15 | type BenchCase = { |
| 16 | name: string; |
| 17 | turns: number; |
| 18 | toolsPerTurn: number; |
| 19 | outputSize: number; |
| 20 | archived: boolean; |
| 21 | }; |
| 22 | |
| 23 | type BenchResult = { |
| 24 | name: string; |
| 25 | messages: number; |
| 26 | items: number; |
| 27 | jsonBytes: number; |
| 28 | itemStringBytes: number; |
| 29 | convertMs: number; |
| 30 | reducerMs: number; |
| 31 | transcriptComputeMs: number; |
| 32 | turnGroups: number; |
| 33 | projectionMs: number; |
| 34 | readMs: number; |
| 35 | projectedNodes: number; |
| 36 | }; |
| 37 | |
| 38 | const cases: BenchCase[] = [ |
| 39 | { name: "200-turns-full-10KB", turns: 200, toolsPerTurn: 1, outputSize: 10 * 1024, archived: false }, |
| 40 | { name: "200-turns-archived-10KB", turns: 200, toolsPerTurn: 1, outputSize: 10 * 1024, archived: true }, |
| 41 | { name: "1000-turns-full-1KB", turns: 1000, toolsPerTurn: 1, outputSize: 1024, archived: false }, |
| 42 | { name: "1000-turns-archived-1KB", turns: 1000, toolsPerTurn: 1, outputSize: 1024, archived: true }, |
| 43 | { name: "1000-turns-archived-3-tools", turns: 1000, toolsPerTurn: 3, outputSize: 1024, archived: true }, |
| 44 | { name: "5000-turns-archived-1-tool", turns: 5000, toolsPerTurn: 1, outputSize: 1024, archived: true }, |
| 45 | { name: "10000-turns-archived-1-tool", turns: 10000, toolsPerTurn: 1, outputSize: 1024, archived: true }, |
| 46 | ]; |
| 47 | |
| 48 | function syntheticHistory(c: BenchCase): HistoryMessage[] { |
| 49 | const messages: any[] = []; |
| 50 | const output = "x".repeat(c.outputSize); |
| 51 | for (let turn = 0; turn < c.turns; turn += 1) { |
| 52 | messages.push({ role: "user", content: `prompt ${turn}` }); |
| 53 | const toolCalls: any[] = []; |
| 54 | for (let tool = 0; tool < c.toolsPerTurn; tool += 1) { |
| 55 | const id = `call_${turn}_${tool}`; |
| 56 | toolCalls.push({ |
| 57 | id, |
| 58 | name: "bash", |
| 59 | arguments: c.archived ? "" : `{"command":"synthetic ${turn} ${tool}"}`, |
| 60 | argumentsArchived: c.archived || undefined, |
| 61 | subject: c.archived ? `synthetic ${turn} ${tool}` : undefined, |
| 62 | summary: c.archived ? "1 line" : undefined, |
| 63 | }); |
| 64 | } |
| 65 | messages.push({ role: "assistant", content: `answer ${turn}`, toolCalls }); |
| 66 | for (let tool = 0; tool < c.toolsPerTurn; tool += 1) { |
| 67 | const id = `call_${turn}_${tool}`; |
| 68 | messages.push({ |
| 69 | role: "tool", |
| 70 | toolCallId: id, |
| 71 | toolName: "bash", |
| 72 | content: c.archived ? "" : output, |
| 73 | toolResultArchived: c.archived || undefined, |
| 74 | }); |
| 75 | } |
| 76 | } |
| 77 | return messages as HistoryMessage[]; |
| 78 | } |
| 79 | |
| 80 | function itemStringBytes(items: Item[]): number { |
| 81 | let total = 0; |
| 82 | for (const item of items) { |
| 83 | if (item.kind === "user") total += item.text.length; |
| 84 | if (item.kind === "assistant") total += item.text.length + item.reasoning.length; |
| 85 | if (item.kind === "tool") total += item.args.length + (item.output?.length ?? 0) + (item.error?.length ?? 0); |
| 86 | } |
| 87 | return total; |
| 88 | } |
| 89 | |
| 90 | function time<T>(fn: () => T): { value: T; ms: number } { |
| 91 | const start = performance.now(); |
| 92 | const value = fn(); |
| 93 | return { value, ms: performance.now() - start }; |
| 94 | } |
| 95 | |
| 96 | function runCase(c: BenchCase): BenchResult { |
| 97 | const messages = syntheticHistory(c); |
| 98 | const jsonBytes = JSON.stringify(messages).length; |
| 99 | const converted = time(() => historyMessagesToItems(messages, "perf")); |
| 100 | const reduced = time(() => reducer(initialState, { type: "history", messages })); |
| 101 | const items = converted.value.items; |
| 102 | const source = new ChatSource(c.name); |
| 103 | const projection = time(() => source.update({ items, running: false, hydrating: false, hasOlder: false, loadingOlder: false })); |
| 104 | const transcript = time(() => source.getOrderSnapshot().filter(key => source.getNodeSnapshot(key)?.kind === "user")); |
| 105 | const range = time(() => source.getOrderSnapshot().map(key => source.getNodeSnapshot(key))); |
| 106 | const nodeCount = range.value.length; |
| 107 | source.dispose(); |
| 108 | return { |
| 109 | name: c.name, |
| 110 | messages: messages.length, |
| 111 | items: items.length, |
| 112 | jsonBytes, |
| 113 | itemStringBytes: itemStringBytes(reduced.value.items), |
| 114 | convertMs: converted.ms, |
| 115 | reducerMs: reduced.ms, |
| 116 | transcriptComputeMs: transcript.ms, |
| 117 | turnGroups: transcript.value.length, |
| 118 | projectionMs: projection.ms, |
| 119 | readMs: range.ms, |
| 120 | projectedNodes: nodeCount, |
| 121 | }; |
| 122 | } |
| 123 | |
| 124 | function printResult(r: BenchResult): void { |
| 125 | process.stdout.write([ |
| 126 | r.name, |
| 127 | `messages=${r.messages}`, |
| 128 | `items=${r.items}`, |
| 129 | `jsonBytes=${r.jsonBytes}`, |
| 130 | `itemStringBytes=${r.itemStringBytes}`, |
| 131 | `convertMs=${r.convertMs.toFixed(2)}`, |
| 132 | `reducerMs=${r.reducerMs.toFixed(2)}`, |
| 133 | `transcriptComputeMs=${r.transcriptComputeMs.toFixed(2)}`, |
| 134 | `turnGroups=${r.turnGroups}`, |
| 135 | `projectionMs=${r.projectionMs.toFixed(2)}`, |
| 136 | `readMs=${r.readMs.toFixed(2)}`, |
| 137 | `projectedNodes=${r.projectedNodes}`, |
| 138 | ].join(" ") + "\n"); |
| 139 | } |
| 140 | |
| 141 | console.log("\nhistory performance benchmark"); |
| 142 | const results = cases.map(runCase); |
| 143 | for (const result of results) { |
| 144 | printResult(result); |
| 145 | } |
| 146 | |
| 147 | const failures: string[] = []; |
| 148 | for (let index = 0; index < results.length; index += 1) { |
| 149 | const result = results[index]; |
| 150 | const input = cases[index]; |
| 151 | const expectedMessages = input.turns * (2 + input.toolsPerTurn); |
| 152 | if (result.messages !== expectedMessages) failures.push(`${result.name}: unexpected message count`); |
| 153 | if (result.turnGroups !== input.turns) failures.push(`${result.name}: unexpected turn-group count`); |
| 154 | if (result.convertMs > 1_000 || result.reducerMs > 1_000 || result.transcriptComputeMs > 1_000 || result.projectionMs > 1_000 || result.readMs > 1_000) { |
| 155 | failures.push(`${result.name}: exceeded 1s responsiveness ceiling`); |
| 156 | } |
| 157 | if (result.projectedNodes < input.turns * 3) failures.push(`${result.name}: missing loaded history nodes`); |
| 158 | } |
| 159 | |
| 160 | const full10KB = results.find((result) => result.name === "200-turns-full-10KB"); |
| 161 | const archived10KB = results.find((result) => result.name === "200-turns-archived-10KB"); |
| 162 | if (!full10KB || full10KB.itemStringBytes * 10 >= full10KB.jsonBytes) { |
| 163 | failures.push("restored full tool results retained too much source text"); |
| 164 | } |
| 165 | if (!archived10KB || archived10KB.itemStringBytes * 5 >= archived10KB.jsonBytes) { |
| 166 | failures.push("restored archived tool results retained too much source text"); |
| 167 | } |
| 168 | |
| 169 | // ── Session-switch diagnostics ─────────────────────────────────────────────── |
| 170 | // These fixtures verify diagnostic interpretation, not physical disk reads. |
| 171 | // The desktop switch tests exercise the real load and snapshot entry points. |
| 172 | // Missing evidence must remain unknown instead of passing a zero-repeat gate. |
| 173 | const switchMessages = syntheticHistory(cases[1]); |
| 174 | const switchPhases = { |
| 175 | resolveMs: 1, loadMs: 12, rebindMs: 30, historyMs: 9, totalMs: 52, |
| 176 | loadedMessages: switchMessages.length, loadedBytes: 65_536, |
| 177 | historyEntries: switchMessages.length, durableReads: 1, outcome: "ok", |
| 178 | }; |
| 179 | |
| 180 | resetSessionDiagnostics(); |
| 181 | noteActivationRequested("switch-ticket"); |
| 182 | noteActivationStarted("switch-ticket", "tab-switch"); |
| 183 | noteActivationSettled("switch-ticket", "ready"); |
| 184 | noteResumeHistoryPage({ messages: switchMessages, switch: switchPhases }, switchPhases.totalMs); |
| 185 | noteTranscriptRowCounts(40, switchMessages.length); |
| 186 | const pipelined = sessionPipelineDiagnostics(); |
| 187 | process.stdout.write(`\n${JSON.stringify({ |
| 188 | activation: pipelined.activation, |
| 189 | history: pipelined.history, |
| 190 | mountedRows: pipelined.mountedRows, |
| 191 | duplicateLoadCount: pipelined.duplicateLoadCount, |
| 192 | }, null, 2)}\n`); |
| 193 | |
| 194 | if (pipelined.duplicateLoadCount !== 0) { |
| 195 | failures.push(`switch performed duplicate durable loads: ${pipelined.duplicateLoadCount}`); |
| 196 | } |
| 197 | if (pipelined.history?.source !== "resume-loaded") { |
| 198 | failures.push(`switch history source = ${pipelined.history?.source}, want resume-loaded`); |
| 199 | } |
| 200 | if (pipelined.history?.entries !== switchMessages.length) { |
| 201 | failures.push(`switch history entries = ${pipelined.history?.entries}, want ${switchMessages.length}`); |
| 202 | } |
| 203 | if (pipelined.activation?.totalMs === undefined || pipelined.activation.startingToReadyMs === undefined) { |
| 204 | failures.push("activation phases were not derived from the ticket"); |
| 205 | } |
| 206 | if (pipelined.mountedRows?.mounted !== 40 || pipelined.mountedRows.total !== switchMessages.length) { |
| 207 | failures.push("mounted row counts were not reported"); |
| 208 | } |
| 209 | // The gate has to be able to fail, or it proves nothing: a switch that rebuilt |
| 210 | // its first screen from a second read reports two. |
| 211 | noteResumeHistoryPage({ messages: switchMessages, switch: { ...switchPhases, durableReads: 2 } }, switchPhases.totalMs); |
| 212 | if (sessionPipelineDiagnostics().duplicateLoadCount !== 1) { |
| 213 | failures.push("duplicate-load gate did not observe a second durable read"); |
| 214 | } |
| 215 | resetSessionDiagnostics(); |
| 216 | if (sessionPipelineDiagnostics().duplicateLoadCount !== null) { |
| 217 | failures.push("missing switch evidence must remain unknown"); |
| 218 | } |
| 219 | noteResumeHistoryPage({ messages: switchMessages, switch: switchPhases }, 60, 8); |
| 220 | if (sessionPipelineDiagnostics().resumeHistory?.source !== "transcript-snapshot" || sessionPipelineDiagnostics().resumeSnapshotMs !== 8) failures.push("modern snapshot timing is missing"); |
| 221 | noteResumeHistoryPage({ messages: [] }, 1, 1); |
| 222 | if (sessionPipelineDiagnostics().duplicateLoadCount !== null || sessionPipelineDiagnostics().resumeSwitch) failures.push("an uninstrumented response retained old switch evidence"); |
| 223 | |
| 224 | if (failures.length > 0) { |
| 225 | for (const failure of failures) process.stderr.write(`FAIL ${failure}\n`); |
| 226 | process.exit(1); |
| 227 | } |
| 228 | process.stdout.write("PASS long-history performance contracts\n"); |
| 229 |