| 1 | // Run: tsx src/__tests__/tool-card-shell-execution.test.tsx |
| 2 | // |
| 3 | // Desktop ToolCard shell-execution presentation for the three host paths: |
| 4 | // 1) live tool_result event (reducer → card) |
| 5 | // 2) history recovery (historyMessagesToItems) |
| 6 | // 3) archived expand (ToolResultForTab injects execution back onto fullData) |
| 7 | |
| 8 | import { JSDOM } from "jsdom"; |
| 9 | import React from "react"; |
| 10 | import { act } from "react"; |
| 11 | import { createRoot } from "react-dom/client"; |
| 12 | import { ToolCard } from "../components/ToolCard"; |
| 13 | import { LocaleProvider } from "../lib/i18n"; |
| 14 | import { historyMessagesToItems, initialState, reducer, type Item } from "../lib/useController"; |
| 15 | import type { HistoryMessage, WireShellExecution } from "../lib/types"; |
| 16 | import { installDesktopHostStub } from "./desktopHostStub"; |
| 17 | |
| 18 | type ToolItem = Extract<Item, { kind: "tool" }>; |
| 19 | |
| 20 | let passed = 0; |
| 21 | let failed = 0; |
| 22 | |
| 23 | function ok(value: unknown, label: string) { |
| 24 | if (value) { |
| 25 | process.stdout.write(` PASS ${label}\n`); |
| 26 | passed += 1; |
| 27 | } else { |
| 28 | process.stdout.write(` FAIL ${label}\n`); |
| 29 | failed += 1; |
| 30 | } |
| 31 | } |
| 32 | |
| 33 | function eq(actual: unknown, expected: unknown, label: string) { |
| 34 | if (actual === expected) ok(true, label); |
| 35 | else ok(false, `${label}: expected ${JSON.stringify(expected)}, got ${JSON.stringify(actual)}`); |
| 36 | } |
| 37 | |
| 38 | function flushTimers(): Promise<void> { |
| 39 | return new Promise((resolve) => setTimeout(resolve, 0)); |
| 40 | } |
| 41 | |
| 42 | function installDom() { |
| 43 | const dom = new JSDOM("<!doctype html><html><body><div id=\"root\"></div></body></html>", { |
| 44 | pretendToBeVisual: true, |
| 45 | url: "http://localhost/", |
| 46 | }); |
| 47 | (globalThis as typeof globalThis & { IS_REACT_ACT_ENVIRONMENT: boolean }).IS_REACT_ACT_ENVIRONMENT = true; |
| 48 | globalThis.window = dom.window as unknown as Window & typeof globalThis; |
| 49 | globalThis.document = dom.window.document; |
| 50 | Object.defineProperty(globalThis, "navigator", { configurable: true, value: dom.window.navigator }); |
| 51 | globalThis.Node = dom.window.Node; |
| 52 | globalThis.Element = dom.window.Element; |
| 53 | globalThis.HTMLElement = dom.window.HTMLElement; |
| 54 | globalThis.Event = dom.window.Event; |
| 55 | globalThis.MouseEvent = dom.window.MouseEvent; |
| 56 | globalThis.requestAnimationFrame = dom.window.requestAnimationFrame.bind(dom.window); |
| 57 | globalThis.cancelAnimationFrame = dom.window.cancelAnimationFrame.bind(dom.window); |
| 58 | dom.window.matchMedia = () => ({ |
| 59 | matches: true, |
| 60 | media: "(prefers-reduced-motion: reduce)", |
| 61 | onchange: null, |
| 62 | addListener: () => undefined, |
| 63 | removeListener: () => undefined, |
| 64 | addEventListener: () => undefined, |
| 65 | removeEventListener: () => undefined, |
| 66 | dispatchEvent: () => false, |
| 67 | }); |
| 68 | return dom; |
| 69 | } |
| 70 | |
| 71 | async function renderCard(item: ToolItem, tabId?: string) { |
| 72 | const dom = installDom(); |
| 73 | const rootEl = document.getElementById("root"); |
| 74 | if (!rootEl) throw new Error("missing root"); |
| 75 | const root = createRoot(rootEl); |
| 76 | await act(async () => { |
| 77 | root.render( |
| 78 | React.createElement(LocaleProvider, null, |
| 79 | React.createElement(ToolCard, { item, tabId }), |
| 80 | ), |
| 81 | ); |
| 82 | await flushTimers(); |
| 83 | }); |
| 84 | return { |
| 85 | root, |
| 86 | dom, |
| 87 | async expand() { |
| 88 | const head = document.querySelector(".tool__head") as HTMLButtonElement | null; |
| 89 | if (!head) throw new Error("tool head missing"); |
| 90 | await act(async () => { |
| 91 | head.click(); |
| 92 | await flushTimers(); |
| 93 | }); |
| 94 | }, |
| 95 | async cleanup() { |
| 96 | await act(async () => { |
| 97 | root.unmount(); |
| 98 | }); |
| 99 | dom.window.close(); |
| 100 | }, |
| 101 | }; |
| 102 | } |
| 103 | |
| 104 | const failedPSExecution: WireShellExecution = { |
| 105 | kind: "shell", |
| 106 | shell: "powershell", |
| 107 | shellVersion: "5.1", |
| 108 | platform: "windows", |
| 109 | supportsAndAnd: false, |
| 110 | state: "failed", |
| 111 | failurePhase: "execution", |
| 112 | exitCode: 1, |
| 113 | outputTail: "Select-String : 找不到路径“C:\\中文\\app.ps1”。\nAt line:1 char:1", |
| 114 | mutationRisk: "may_be_partial", |
| 115 | verification: "not_verification", |
| 116 | durationMs: 42, |
| 117 | }; |
| 118 | |
| 119 | const preflightExecution: WireShellExecution = { |
| 120 | kind: "shell", |
| 121 | shell: "bash", |
| 122 | state: "not_run", |
| 123 | failurePhase: "preflight", |
| 124 | mutationRisk: "not_started", |
| 125 | verification: "not_run", |
| 126 | durationMs: 0, |
| 127 | }; |
| 128 | |
| 129 | console.log("\ntool card shell execution"); |
| 130 | |
| 131 | // Canonical pwsh calls become terminal cards before execution metadata arrives. |
| 132 | { |
| 133 | const livePwsh: ToolItem = { |
| 134 | kind: "tool", id: "pwsh-live", name: "pwsh", |
| 135 | args: `{"command":"Write-Output ready","description":"Display readiness"}`, |
| 136 | readOnly: false, status: "running", |
| 137 | }; |
| 138 | const view = await renderCard(livePwsh); |
| 139 | try { |
| 140 | eq(document.querySelector(".tool__name")?.textContent, "PowerShell 7+", "pwsh dispatch renders PowerShell terminal identity"); |
| 141 | ok((document.body.textContent ?? "").includes("Write-Output ready"), "pwsh dispatch renders command summary"); |
| 142 | } finally { |
| 143 | await view.cleanup(); |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | // ── Path 1: live tool_result event ── |
| 148 | { |
| 149 | let s = reducer(initialState, { type: "event", e: { kind: "turn_started" } }); |
| 150 | s = reducer(s, { |
| 151 | type: "event", |
| 152 | e: { |
| 153 | kind: "tool_dispatch", |
| 154 | tool: { |
| 155 | id: "live-ps", |
| 156 | name: "bash", |
| 157 | args: `{"command":"Get-Content .\\\\中文\\\\app.ps1"}`, |
| 158 | readOnly: false, |
| 159 | }, |
| 160 | }, |
| 161 | }); |
| 162 | s = reducer(s, { |
| 163 | type: "event", |
| 164 | e: { |
| 165 | kind: "tool_result", |
| 166 | tool: { |
| 167 | id: "live-ps", |
| 168 | name: "bash", |
| 169 | readOnly: false, |
| 170 | output: "error: command exited: exit status 1\nSelect-String failed", |
| 171 | err: "command exited: exit status 1", |
| 172 | durationMs: 42, |
| 173 | execution: failedPSExecution, |
| 174 | }, |
| 175 | }, |
| 176 | }); |
| 177 | const item = s.items.find((it): it is ToolItem => it.kind === "tool" && it.id === "live-ps"); |
| 178 | ok(!!item, "live path: tool item created"); |
| 179 | eq(item?.isShell, true, "live path: bash result marked isShell"); |
| 180 | eq(item?.execution?.shell, "powershell", "live path: execution.shell preserved"); |
| 181 | eq(item?.execution?.exitCode, 1, "live path: exitCode preserved"); |
| 182 | eq(item?.execution?.failurePhase, "execution", "live path: failurePhase preserved"); |
| 183 | |
| 184 | const ui = await renderCard(item!); |
| 185 | const name = document.querySelector(".tool__name")?.textContent ?? ""; |
| 186 | ok(name.includes("Windows PowerShell"), `live path: card shows Windows PowerShell (got ${JSON.stringify(name)})`); |
| 187 | const duration = document.querySelector(".tool__duration")?.textContent ?? ""; |
| 188 | ok(duration.includes("exit 1") || duration.includes("execution"), `live path: summary shows exit/phase (got ${JSON.stringify(duration)})`); |
| 189 | const risk = document.body.textContent ?? ""; |
| 190 | ok(risk.includes("partially modified") || risk.includes("部分"), "live path: partial mutation risk visible"); |
| 191 | // Execution failure must not claim the command never ran. |
| 192 | ok( |
| 193 | !(risk.includes("did not run") || risk.includes("command did not run") || risk.includes("命令未执行") || risk.includes("未执行")), |
| 194 | "live path: execution failure must not show not-run label", |
| 195 | ); |
| 196 | |
| 197 | await ui.expand(); |
| 198 | const details = document.querySelector("details.tool__error-details, .tool__error-details, details"); |
| 199 | ok(!!details, "live path: stderr details element present after expand affordance"); |
| 200 | if (details) { |
| 201 | await act(async () => { |
| 202 | details.setAttribute("open", ""); |
| 203 | details.dispatchEvent(new Event("toggle")); |
| 204 | await flushTimers(); |
| 205 | }); |
| 206 | } |
| 207 | const afterExpand = document.body.textContent ?? ""; |
| 208 | ok( |
| 209 | afterExpand.includes("中文") || afterExpand.includes("找不到路径"), |
| 210 | "live path: Chinese stderr from execution.outputTail is visible in the DOM", |
| 211 | ); |
| 212 | |
| 213 | await ui.cleanup(); |
| 214 | } |
| 215 | |
| 216 | // timed_out / cancelled must surface partial-write risk (backend may_be_partial). |
| 217 | { |
| 218 | for (const state of ["timed_out", "cancelled"] as const) { |
| 219 | const item: ToolItem = { |
| 220 | kind: "tool", |
| 221 | id: `risk-${state}`, |
| 222 | name: "bash", |
| 223 | args: `{"command":"long-run"}`, |
| 224 | readOnly: false, |
| 225 | status: "error", |
| 226 | error: state === "timed_out" ? "command timed out" : "context canceled", |
| 227 | isShell: true, |
| 228 | execution: { |
| 229 | kind: "shell", |
| 230 | shell: "bash", |
| 231 | state, |
| 232 | failurePhase: state === "timed_out" ? "timeout" : "cancellation", |
| 233 | mutationRisk: "may_be_partial", |
| 234 | verification: "not_verification", |
| 235 | durationMs: 100, |
| 236 | }, |
| 237 | }; |
| 238 | const ui = await renderCard(item); |
| 239 | const body = document.body.textContent ?? ""; |
| 240 | ok( |
| 241 | body.includes("partially modified") || body.includes("部分"), |
| 242 | `${state}: shows partial mutation risk`, |
| 243 | ); |
| 244 | ok( |
| 245 | !(body.includes("did not run") || body.includes("命令未执行")), |
| 246 | `${state}: does not claim command never ran`, |
| 247 | ); |
| 248 | await ui.cleanup(); |
| 249 | } |
| 250 | } |
| 251 | |
| 252 | // ── Path 2: history recovery ── |
| 253 | { |
| 254 | const messages: HistoryMessage[] = [ |
| 255 | { |
| 256 | role: "assistant", |
| 257 | content: "", |
| 258 | toolCalls: [{ id: "hist-bash", name: "bash", arguments: "{\"command\":\"go test ./...\"}" }], |
| 259 | }, |
| 260 | { |
| 261 | role: "tool", |
| 262 | content: "blocked: mixed mutation and verification", |
| 263 | toolCallId: "hist-bash", |
| 264 | toolName: "bash", |
| 265 | toolResultError: "blocked: mixed mutation and verification command", |
| 266 | execution: preflightExecution, |
| 267 | }, |
| 268 | ]; |
| 269 | const items = historyMessagesToItems(messages, "h").items.filter((it): it is ToolItem => it.kind === "tool"); |
| 270 | eq(items.length, 1, "history path: one tool item"); |
| 271 | eq(items[0]?.isShell, true, "history path: bash isShell"); |
| 272 | eq(items[0]?.execution?.failurePhase, "preflight", "history path: execution restored"); |
| 273 | eq(items[0]?.execution?.mutationRisk, "not_started", "history path: not_started risk"); |
| 274 | |
| 275 | const ui = await renderCard(items[0]!); |
| 276 | const name = document.querySelector(".tool__name")?.textContent ?? ""; |
| 277 | ok(name === "bash" || name.includes("bash"), `history path: shell name bash (got ${JSON.stringify(name)})`); |
| 278 | const body = document.body.textContent ?? ""; |
| 279 | ok( |
| 280 | body.includes("did not run") || body.includes("not run") || body.includes("未执行") || body.includes("命令未执行"), |
| 281 | "history path: preflight shows command-not-run (not partial)", |
| 282 | ); |
| 283 | ok( |
| 284 | !(body.includes("partially modified") || body.includes("部分修改文件")), |
| 285 | "history path: preflight must not claim partial file modification", |
| 286 | ); |
| 287 | await ui.cleanup(); |
| 288 | } |
| 289 | |
| 290 | // ── Path 3a: archive compaction keeps execution from the live result ── |
| 291 | { |
| 292 | let s = reducer(initialState, { type: "event", e: { kind: "turn_started" } }); |
| 293 | s = reducer(s, { |
| 294 | type: "event", |
| 295 | e: { |
| 296 | kind: "tool_dispatch", |
| 297 | tool: { id: "arch-live", name: "bash", args: `{"command":"exit 1"}`, readOnly: false }, |
| 298 | }, |
| 299 | }); |
| 300 | s = reducer(s, { |
| 301 | type: "event", |
| 302 | e: { |
| 303 | kind: "tool_result", |
| 304 | tool: { |
| 305 | id: "arch-live", |
| 306 | name: "bash", |
| 307 | readOnly: false, |
| 308 | output: "error: exit 1\n" + "y".repeat(5000), |
| 309 | err: "command exited: exit status 1", |
| 310 | durationMs: 42, |
| 311 | execution: failedPSExecution, |
| 312 | }, |
| 313 | }, |
| 314 | }); |
| 315 | const item = s.items.find((it): it is ToolItem => it.kind === "tool" && it.id === "arch-live"); |
| 316 | ok(!!item?.dataArchived, "archive path: tool_result archives large output"); |
| 317 | eq(item?.output, undefined, "archive path: output dropped after archive"); |
| 318 | eq(item?.execution?.shell, "powershell", "archive path: execution survives compactArchivedToolItems"); |
| 319 | eq(item?.execution?.exitCode, 1, "archive path: exitCode survives archive"); |
| 320 | |
| 321 | const ui = await renderCard(item!); |
| 322 | ok((document.querySelector(".tool__name")?.textContent ?? "").includes("Windows PowerShell"), |
| 323 | "archive path: card still shows Windows PowerShell without re-fetch"); |
| 324 | ok((document.body.textContent ?? "").includes("partially modified") || (document.body.textContent ?? "").includes("部分"), |
| 325 | "archive path: partial risk still visible when execution retained"); |
| 326 | await ui.cleanup(); |
| 327 | } |
| 328 | |
| 329 | // ── Path 3b: ToolResultForTab rehydrates execution when only fullData has it ── |
| 330 | { |
| 331 | const archived: ToolItem = { |
| 332 | kind: "tool", |
| 333 | id: "arch-ps", |
| 334 | name: "bash", |
| 335 | args: "", |
| 336 | readOnly: false, |
| 337 | status: "error", |
| 338 | error: "command exited: exit status 1", |
| 339 | dataArchived: true, |
| 340 | isShell: true, |
| 341 | execution: undefined, |
| 342 | durationMs: 42, |
| 343 | }; |
| 344 | |
| 345 | const dom = installDom(); |
| 346 | // Inject a desktop host stub so the bridge proxy picks up our methods |
| 347 | // instead of the browser mock (which always returns null for ToolResultForTab). |
| 348 | installDesktopHostStub(({ |
| 349 | main: { |
| 350 | App: { |
| 351 | ToolResultForTab: async () => ({ |
| 352 | args: `{"command":"Get-Content .\\\\中文\\\\app.ps1"}`, |
| 353 | output: "error: command exited: exit status 1\nSelect-String failed", |
| 354 | execution: failedPSExecution, |
| 355 | }), |
| 356 | }, |
| 357 | }, |
| 358 | }).main.App); |
| 359 | const rootEl = document.getElementById("root"); |
| 360 | if (!rootEl) throw new Error("missing root"); |
| 361 | const root = createRoot(rootEl); |
| 362 | |
| 363 | try { |
| 364 | await act(async () => { |
| 365 | root.render( |
| 366 | React.createElement(LocaleProvider, null, |
| 367 | React.createElement(ToolCard, { item: archived, tabId: "tab-1" }), |
| 368 | ), |
| 369 | ); |
| 370 | await flushTimers(); |
| 371 | }); |
| 372 | const head = document.querySelector(".tool__head") as HTMLButtonElement | null; |
| 373 | if (!head) throw new Error("tool head missing"); |
| 374 | await act(async () => { |
| 375 | head.click(); |
| 376 | await flushTimers(); |
| 377 | await flushTimers(); |
| 378 | await flushTimers(); |
| 379 | }); |
| 380 | |
| 381 | const name = document.querySelector(".tool__name")?.textContent ?? ""; |
| 382 | ok(name.includes("Windows PowerShell"), `archive rehydrate: shell name after expand (got ${JSON.stringify(name)})`); |
| 383 | const duration = document.querySelector(".tool__duration")?.textContent ?? ""; |
| 384 | ok( |
| 385 | duration.includes("exit 1") || duration.includes("execution"), |
| 386 | `archive rehydrate: exit/phase after expand (got ${JSON.stringify(duration)})`, |
| 387 | ); |
| 388 | const body = document.body.textContent ?? ""; |
| 389 | ok(body.includes("partially modified") || body.includes("部分"), "archive rehydrate: partial risk after expand"); |
| 390 | } finally { |
| 391 | await act(async () => { |
| 392 | root.unmount(); |
| 393 | }); |
| 394 | delete (window as unknown as { go?: unknown }).go; |
| 395 | dom.window.close(); |
| 396 | } |
| 397 | } |
| 398 | |
| 399 | // ── Nil execution safety ── |
| 400 | { |
| 401 | const plain: ToolItem = { |
| 402 | kind: "tool", |
| 403 | id: "plain", |
| 404 | name: "bash", |
| 405 | args: `{"command":"echo hi"}`, |
| 406 | readOnly: false, |
| 407 | status: "done", |
| 408 | output: "hi\n", |
| 409 | isShell: true, |
| 410 | }; |
| 411 | const ui = await renderCard(plain); |
| 412 | ok(document.querySelector(".tool__name")?.textContent === "bash", "nil execution falls back to bash label"); |
| 413 | ok(!document.querySelector("[data-shell]") || document.querySelector("[data-shell]")?.getAttribute("data-shell") === "bash", |
| 414 | "nil execution still renders shell card without throwing"); |
| 415 | await ui.cleanup(); |
| 416 | } |
| 417 | |
| 418 | console.log(`\n${passed} passed, ${failed} failed, ${passed + failed} total`); |
| 419 | if (failed > 0) process.exit(1); |
| 420 |