feat: validate OneComputer sandbox artifacts

Author: Gini

docs/MANUS-PARITY.md

@@ -116,7 +116,7 @@ This is the implementation gate, not a marketing checklist. **I** means behavior
 
 ## Validation, delivery, and reuse
 
-91. **P** Production-build validation and portability — ONEVibe ZIPs source and writes a versioned per-task static contract report. Generated React/Vite projects now validate their portable dependency/scripts contract and output semantics, but are not yet dependency-installed or built in an attested sandbox.
+91. **P** Production-build validation and portability — ONEVibe ZIPs source and writes a versioned per-task static contract report. Native Claude SDK and ONEComputer sandbox runs now both emit that bounded report after workspace delivery; generated React/Vite projects validate their portable dependency/scripts contract and output semantics, but are not yet dependency-installed or built in an attested sandbox.
 92. **P** Accessibility-validation step — static preview semantics plus generated Website landmarks, native FAQ disclosure, compact layout, reduced-motion, and keyboard-focus affordances are checked. Automated accessibility scans remain pending.
 93. **P** Live-preview delivery and agent browser validation — isolated local preview works, and an attested sandbox may surface server-proxied X11 PNG frames. For Website, App, and Game outputs, a gateway-enforced browser runtime now additionally renders `file://` `index.html` inside the sandbox into a preserved artifact-rail screenshot with hostname resolution blocked. Allowlisted agent browser activity remains separately recorded; managed HTTPS and a deployed microVM proof remain pending.
 94. **P** Publish control — external approval is required and publication is withheld; approved deployment is pending.

server/onecomputer-sandbox-runner.test.ts

@@ -108,6 +108,8 @@ describe('OneComputerSandboxRuntimeAdapter', () => {
     expect(commands.some((command) => command.includes('onevibe-browser-review.png'))).toBe(true)
     expect(store.listEvents(task.id).some((event) => event.label === 'Sandbox browser review observed' && event.payload.generatedArtifactPreview === true)).toBe(true)
     expect(frames.at(-1)?.payload.uri).toContain(`/api/tasks/${task.id}/file?path=evidence%2Fvisual%2Fbrowser-review-`)
+    expect(store.listEvents(task.id).some((event) => event.label === 'Static artifact contract needs review' && event.content === 'validation-report.json')).toBe(true)
+    expect(await store.readWorkspaceFile(task.id, 'validation-report.json')).toContain('Static contract validation only')
     expect(store.listEvents(task.id).at(-1)?.type).toBe('run_completed')
     expect(store.verifyChain(task.id)).toBe(true)
   })

server/onecomputer-sandbox-runner.ts

@@ -2,6 +2,7 @@ import { createHash } from 'node:crypto'
 import path from 'node:path'
 import type { OneComputerClient } from './onecomputer-client.js'
 import type { RuntimeAdapter, RuntimeContext } from './runtime-adapter.js'
+import { validateModeArtifacts } from './artifact-validation.js'
 
 const PNG_SIGNATURE = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])
 const isPng = (bytes: Buffer) => bytes.byteLength >= PNG_SIGNATURE.byteLength && bytes.subarray(0, PNG_SIGNATURE.byteLength).equals(PNG_SIGNATURE)
@@ -434,6 +435,13 @@ export class OneComputerSandboxRuntimeAdapter implements RuntimeAdapter {
         })
       }
       await store.setPlanStep(task.id, 'build', 'completed')
+      await store.setPlanStep(task.id, 'verify', 'running')
+      const validation = await validateModeArtifacts(store.getTask(task.id), store)
+      await store.appendEvent(task.id, {
+        type: 'artifact_created', lane: 'artifact', label: validation.passed ? 'Static artifact contract passed' : 'Static artifact contract needs review',
+        content: 'validation-report.json',
+        payload: { executionRoute: 'onecomputer_sandbox', kind: 'validation_report', passed: validation.passed, checkCount: validation.checks.length, limitation: validation.limitation },
+      })
       await store.setPlanStep(task.id, 'verify', 'completed')
       await store.setPlanStep(task.id, 'deliver', 'running')
       await destroy()