Ridiculous speed.
Readable code.

1import { agent } from "@b4run/sdk"2 3export default agent({4  model: process.env.B4_CODE_FIXER_MODEL ?? "gpt-5-mini",5  recursionLimit: 60,6  description: "Repairs a failing test and verifies the change.",7  tools: { approve: ["exportForReview"] }, systemPrompt · instructions folded22})

14export function projectWorkspace(id: string): WorkspaceDefinition {15  const manifest = projectManifest(id)16  return {17    source: {18      directory: "sample/project",19      include: [...manifest.allowedSourcePaths, ...manifest.immutablePaths],20      files: [21        { path: "TASK.md", file: "sample/task.md" },22        { path: ".gitignore", text: "node_modules/\n" },23        { path: "project.json", text: JSON.stringify({ id: manifest.id }) },24      ],25    },26    environmentLinks: [27      { path: "node_modules", target: `/opt/fixtures/${manifest.id}/node_modules` },28    ],29    baseline: "git",30  }31}

11  sandbox: {12    ...sandboxPolicy,13    provider: dockerSandbox({ scope: "code-fixer-local", image: sandboxImage }),14    workspace: projectWorkspace(task),15  },

1import type { B4ToolContext } from "@b4run/sdk"2import { inspectCandidate } from "../../../review/inspect.js"3import { renderReviewDiff } from "../../../review/patch.js"4import { verifyChanges } from "../../../review/verifier.js"5 6/** Verify the source repair independently and return the exact candidate for review. */7export default async function prepareReview(_input: Record<string, never>, ctx: B4ToolContext) {8  // Compare the workspace with its captured source and reject edits outside the allowed files.9  const { manifest, candidate, baseline, initial } = await inspectCandidate(ctx)10 11  // Apply these exact changes in a fresh workspace and run both test suites.12  const verification = await verifyChanges(manifest.id, candidate.changes, ctx.signal, initial)13  if (!verification.passed) throw new Error("Independent verification failed")14 15  // Give the agent a readable diff and the exact candidate it must submit for approval.16  return {17    task: manifest.id,18    candidate,19    diff: renderReviewDiff(baseline, candidate.changes),20    verification,21  }22}

1- [ ] Read the task and reproduce the failing test2- [ ] Inspect the relevant source and identify the cause3- [ ] Make a focused source repair4- [ ] Run the test, check preservation requirements, and explain the verified result5- [ ] Call prepareReview and inspect the independently verified candidate6- [ ] Pass that exact candidate to exportForReview for runtime approval

1{2  "visible": {3    "file": "test/cli.test.ts",4    "assertions": [5      "documented dry-run flag reaches the handler"6    ]7  },8  "independent": {9    "file": "checks/independent.test.ts",10    "assertions": [11      "forwards cap and memory-level cwd",12      "rejects unknown and incomplete arguments",13      "dry-run preserves memory state and creates no files"14    ]15  }16}

7  tools: { approve: ["exportForReview"] },

15  const candidate = validateCandidate(input.candidate)16  const inspected = await inspectCandidate(ctx)17  if (candidate.receiptDigest !== inspected.candidate.receiptDigest)18    throw new Error("Workspace changed since review; prepare and approve a new candidate")19  const verification = await verifyChanges(20    inspected.manifest.id,21    candidate.changes,22    ctx.signal,23    inspected.initial,24  )25  if (!verification.passed) throw new Error("Independent verification failed")