Validate runner workspace patch base identity

Author: chubes4

.github/scripts/run-agent-task/execute-native-agent-task.mjs

@@ -590,9 +590,9 @@ if (execution.code === 0 && runtimeResult.success === true && request.runner_wor
     const publicCore = await import(pathToFileURL(join(codeboxRoot, "packages/runtime-core/dist/public.js")).href)
     const refs = publicCore.normalizePublicArtifactRefDTOs(runtimeResult)
       .filter((ref) => ref.kind === "codebox-patch" || ref.kind === "codebox-changed-files")
-     workspaceApply = await runnerWorkspaceCore.applyRunnerWorkspacePatch({ artifactRoot: artifactsPath, artifactRefs: refs, workspaceRoot: workspace, writablePaths })
+      workspaceApply = await runnerWorkspaceCore.applyRunnerWorkspacePatch({ artifactRoot: artifactsPath, artifactRefs: refs, workspaceRoot: workspace, writablePaths, seedIdentity: runnerWorkspaceSeedSnapshot?.provenance.identity })
   } catch (error) {
-    downstreamFailure = { stage: "apply", message: bounded(error instanceof Error ? error.message : String(error), MAX_OUTPUT_CHARS) }
+    downstreamFailure = { stage: "apply", message: bounded(error instanceof Error ? error.message : String(error), MAX_OUTPUT_CHARS), ...(error?.evidence ? { evidence: error.evidence } : {}) }
   }
 }
 runtimeResult = sanitizeRuntimeSourceValue(runtimeResult, runtimeSourceOutputRoots)

.github/scripts/run-agent-task/prepare-agent-task-upload.mjs

@@ -223,6 +223,22 @@ function projectWorkflowResult(value) {
   }).filter(([, entry]) => entry !== undefined)))
 }
 
+async function stageApplyFailureEvidence(result) {
+  const evidence = record(record(result.failure).evidence)
+  const files = [
+    ["patch", record(evidence.patch)],
+    ["changed-files", record(evidence.changed_files)],
+  ]
+  for (const [name, descriptor] of files) {
+    const artifactPath = safeRelativeArtifactPath(descriptor.artifact_path)
+    if (!artifactPath) continue
+    const source = resolve(artifactsPath, artifactPath)
+    if (relative(artifactsPath, source).startsWith("..")) continue
+    const destination = join(uploadPath, ".codebox", "agent-task-artifacts", "apply-failure", name === "patch" ? "rejected.patch" : "changed-files.json")
+    await stageTextFile(source, destination, { allowTargetCode: true })
+  }
+}
+
 function safeTargetPath(value) {
   const path = safeRelativeArtifactPath(value)
   return path ? `workspace/${path}` : undefined
@@ -469,6 +485,7 @@ await mkdir(uploadPath, { recursive: true })
 await stageTextFile(requestPath, join(uploadPath, ".codebox", "agent-task-request.json"))
 await stageTextFile(resultSource, join(uploadPath, ".codebox", "agent-task-workflow-result.json"), { projectWorkflowResult: true })
 await stageTextFile(join(workspace, ".codebox", "native-agent-task-input.json"), join(uploadPath, ".codebox", "native-agent-task-input.json"), { compactNativeInput: true })
+await stageApplyFailureEvidence(result)
 for (const path of declaredPaths) {
   const source = resolve(artifactsPath, path)
   if (relative(artifactsPath, source).startsWith("..") || sourceCategory(path, source)) {

.github/scripts/run-agent-task/runner-workspace-seed-snapshot.mjs

@@ -2,6 +2,8 @@ import { createHash } from "node:crypto"
 import { chmod, copyFile, lstat, mkdir, mkdtemp, readdir, readFile, rm } from "node:fs/promises"
 import { tmpdir } from "node:os"
 import { join, relative, resolve } from "node:path"
+import { execFile } from "node:child_process"
+import { promisify } from "node:util"
 
 // Snapshot policy is intentionally allow-by-exception: disposable build trees and
 // credential-bearing files never enter the agent-readable workspace. `.env.example`
@@ -10,6 +12,7 @@ export const RUNNER_WORKSPACE_SEED_EXCLUDES = [".git/**", ".codebox/**", "node_m
 const EXCLUDED_ROOT_NAMES = new Set(RUNNER_WORKSPACE_SEED_EXCLUDES.filter((pattern) => pattern.endsWith("/**")).map((pattern) => pattern.slice(0, -3)))
 const MAX_FILES = 10_000
 const MAX_BYTES = 256 * 1024 * 1024
+const execFileAsync = promisify(execFile)
 
 function snapshotError(message) {
   const error = new Error(message)
@@ -80,11 +83,17 @@ export async function createRunnerWorkspaceSeedSnapshot(source) {
     const sourceStat = await lstat(sourceRoot)
     if (!sourceStat.isDirectory() || sourceStat.isSymbolicLink()) throw snapshotError("Runner workspace seed source must be a real directory.")
     await copyTree(sourceRoot, root)
+    const contentDigest = digest.digest("hex")
+    const head = await gitHead(sourceRoot)
     return {
       source: root,
       provenance: {
         schema: "wp-codebox/runner-workspace-seed-snapshot/v1",
-        digest: { sha256: digest.digest("hex") },
+        digest: { sha256: contentDigest },
+        identity: {
+          content_digest: { algorithm: "sha256", value: contentDigest },
+          ...(head ? { git: { head } } : {}),
+        },
         files: fileCount,
         bytes: byteCount,
         excludes: RUNNER_WORKSPACE_SEED_EXCLUDES,
@@ -99,3 +108,13 @@ export async function createRunnerWorkspaceSeedSnapshot(source) {
     throw error
   }
 }
+
+async function gitHead(source) {
+  try {
+    const { stdout } = await execFileAsync("git", ["rev-parse", "HEAD"], { cwd: source })
+    const head = stdout.trim()
+    return /^[a-f0-9]{40}$/i.test(head) ? head : ""
+  } catch {
+    return ""
+  }
+}

packages/runtime-core/src/runner-workspace-apply.ts

@@ -28,9 +28,22 @@ export interface RunnerWorkspaceApplyRequest {
   artifactRefs: RunnerWorkspaceArtifactRef[]
   workspaceRoot: string
   writablePaths: string[]
+  seedIdentity?: RunnerWorkspaceSeedIdentity
   verify?: () => Promise<void>
 }
 
+export interface RunnerWorkspaceSeedIdentity {
+  content_digest: { algorithm: "sha256"; value: string }
+  git?: { head: string }
+}
+
+export interface RunnerWorkspaceApplyFailureEvidence {
+  expected_identity?: RunnerWorkspaceSeedIdentity
+  actual_identity: RunnerWorkspaceSeedIdentity
+  patch: { artifact_path: string; sha256: string }
+  changed_files: { artifact_path: string; sha256: string }
+}
+
 export interface RunnerWorkspaceApplyResult {
   schema: "wp-codebox/runner-workspace-apply-result/v1"
   status: "applied" | "no-op"
@@ -72,15 +85,27 @@ export async function applyRunnerWorkspacePatch(request: RunnerWorkspaceApplyReq
 
   const changed = parseChangedFiles(changedRaw)
   validateChangedFiles(changed, request.writablePaths)
+  const failureEvidence = runnerWorkspaceFailureEvidence(request.seedIdentity, await runnerWorkspaceIdentity(workspaceRoot), artifactRoot, patchPath, patch, changedPath, changedRaw)
+  if (request.seedIdentity) {
+    if (!sameIdentity(request.seedIdentity, failureEvidence.actual_identity)) {
+      throw applyFailure("Runner workspace seed identity does not match the host workspace; refusing to apply patch.", {
+        ...failureEvidence,
+      })
+    }
+  }
   if (changed.length === 0) {
     if (patch.trim()) throw new Error("Canonical patch is non-empty but changed-files declares no changes.")
     return { schema: "wp-codebox/runner-workspace-apply-result/v1", status: "no-op", changedFiles: [] }
   }
   if (!patch.trim()) throw new Error("Canonical changed-files declares changes but patch is empty.")
   validatePatchPaths(patch, changed)
 
-  await execGit(workspaceRoot, ["apply", "--check", "--whitespace=error", "--", patchPath])
-  await execGit(workspaceRoot, ["apply", "--whitespace=error", "--", patchPath])
+  try {
+    await execGit(workspaceRoot, ["apply", "--check", "--whitespace=error", "--", patchPath])
+    await execGit(workspaceRoot, ["apply", "--whitespace=error", "--", patchPath])
+  } catch (error) {
+    throw applyFailure(error instanceof Error ? error.message : String(error), failureEvidence)
+  }
   const files = await snapshotWorkspace(workspaceRoot)
   validateAppliedWorkspace(baseline, files, changed)
   if (request.verify) await request.verify()
@@ -98,6 +123,36 @@ export async function applyRunnerWorkspacePatch(request: RunnerWorkspaceApplyReq
   }
 }
 
+export async function runnerWorkspaceIdentity(root: string): Promise<RunnerWorkspaceSeedIdentity> {
+  const digest = createHash("sha256")
+  async function visit(directory: string): Promise<void> {
+    for (const entry of (await readdir(directory, { withFileTypes: true })).sort((left, right) => left.name.localeCompare(right.name))) {
+      if ([".git", ".codebox", "node_modules", "vendor", "dist", "build", "coverage", ".cache"].includes(entry.name)) continue
+      const absolute = resolve(directory, entry.name)
+      const path = relative(root, absolute).replaceAll("\\", "/")
+      const stat = await lstat(absolute)
+      if (excludedSeedFile(path)) continue
+      if (stat.isSymbolicLink() || (!stat.isDirectory() && !stat.isFile())) throw new Error(`Runner workspace contains an unsupported path type: ${path}`)
+      if (stat.isDirectory()) {
+        digest.update(`directory\0${path}\n`)
+        await visit(absolute)
+        continue
+      }
+      const bytes = await readFile(absolute)
+      digest.update(`file\0${path}\0${(stat.mode & 0o111 ? 0o755 : 0o644).toString(8)}\0${bytes.length}\n`)
+      digest.update(bytes)
+    }
+  }
+  await visit(root)
+  const identity: RunnerWorkspaceSeedIdentity = { content_digest: { algorithm: "sha256", value: digest.digest("hex") } }
+  try {
+    const { stdout } = await execFileAsync("git", ["rev-parse", "HEAD"], { cwd: root, maxBuffer: 1024 })
+    const head = stdout.trim()
+    if (/^[a-f0-9]{40}$/i.test(head)) identity.git = { head }
+  } catch { /* A non-git workspace still has a content identity. */ }
+  return identity
+}
+
 /** Ensures checks did not mutate approved output or introduce unrelated files. */
 export async function verifyRunnerWorkspaceIntegrity(snapshot: RunnerWorkspaceIntegritySnapshot): Promise<void> {
   const current = await snapshotWorkspace(snapshot.workspaceRoot)
@@ -217,6 +272,34 @@ function validateAppliedWorkspace(baseline: RunnerWorkspacePublicationFile[], cu
   }
 }
 
+function sameIdentity(expected: RunnerWorkspaceSeedIdentity, actual: RunnerWorkspaceSeedIdentity): boolean {
+  return expected.content_digest.algorithm === "sha256"
+    && expected.content_digest.value === actual.content_digest.value
+    && (!expected.git?.head || expected.git.head === actual.git?.head)
+}
+
+function runnerWorkspaceFailureEvidence(expected: RunnerWorkspaceSeedIdentity | undefined, actual: RunnerWorkspaceSeedIdentity, artifactRoot: string, patchPath: string, patch: string, changedPath: string, changedRaw: string): RunnerWorkspaceApplyFailureEvidence {
+  return {
+    ...(expected ? { expected_identity: expected } : {}),
+    actual_identity: actual,
+    patch: { artifact_path: relative(artifactRoot, patchPath).replaceAll("\\", "/"), sha256: createHash("sha256").update(patch).digest("hex") },
+    changed_files: { artifact_path: relative(artifactRoot, changedPath).replaceAll("\\", "/"), sha256: createHash("sha256").update(changedRaw).digest("hex") },
+  }
+}
+
+function excludedSeedFile(path: string): boolean {
+  const name = path.split("/").at(-1)?.toLowerCase() ?? ""
+  return (name === ".env" || (name.startsWith(".env.") && name !== ".env.example"))
+    || [".npmrc", ".yarnrc.yml", ".pypirc", ".netrc", "auth.json", "credentials", "credentials.json", "secrets.json", "token.json", "id_rsa", "id_ed25519", "id_ecdsa", "id_dsa"].includes(name)
+    || /\.(?:pem|key|p12|pfx)$/i.test(name)
+}
+
+function applyFailure(message: string, evidence: RunnerWorkspaceApplyFailureEvidence): Error {
+  const error = new Error(message) as Error & { evidence?: RunnerWorkspaceApplyFailureEvidence }
+  error.evidence = evidence
+  return error
+}
+
 async function execGit(cwd: string, args: string[]): Promise<void> {
   try {
     await execFileAsync("git", args, { cwd, maxBuffer: MAX_PATCH_BYTES })

tests/execute-native-agent-task-lifecycle.test.mjs

@@ -3,9 +3,13 @@ import { chmod, mkdtemp, mkdir, readFile, writeFile } from "node:fs/promises"
 import { tmpdir } from "node:os"
 import { join, resolve } from "node:path"
 import { spawn } from "node:child_process"
+import { execFile } from "node:child_process"
+import { promisify } from "node:util"
 
 const root = resolve(".")
 const executor = join(root, ".github/scripts/run-agent-task/execute-native-agent-task.mjs")
+const uploader = join(root, ".github/scripts/run-agent-task/prepare-agent-task-upload.mjs")
+const execFileAsync = promisify(execFile)
 
 async function run(mode = "success") {
   const temp = await mkdtemp(join(tmpdir(), "wp-codebox-native-lifecycle-"))
@@ -61,19 +65,22 @@ async function run(mode = "success") {
 
   const fakeCli = join(temp, "fake-cli.mjs")
   const noOp = mode.startsWith("no-op")
+  const mismatch = mode === "mismatch"
+  const applyReject = mode === "apply-reject"
   await writeFile(fakeCli, `
 import { mkdir, readFile, writeFile } from "node:fs/promises"
 import { join } from "node:path"
 const output = process.argv[process.argv.indexOf("--result-file") + 1]
 const input = JSON.parse(await readFile(process.argv[process.argv.indexOf("--input-file") + 1], "utf8"))
 const root = join(process.cwd(), ".codebox", "agent-task-artifacts", "files")
 await mkdir(root, { recursive: true })
-const patch = ${noOp} ? "" : "--- a/README.md\\n+++ b/README.md\\n@@ -1 +1 @@\\n-before\\n+after\\n"
+const patch = ${JSON.stringify(noOp ? "" : applyReject ? "--- a/README.md\n+++ b/README.md\n@@ -1 +1 @@\n-not-before\n+after\n" : "--- a/README.md\n+++ b/README.md\n@@ -1 +1 @@\n-before\n+after\n")}
 await writeFile(join(root, "patch.diff"), patch)
 await writeFile(join(root, "changed-files.json"), JSON.stringify({
   schema: "wp-codebox/changed-files/v1",
   files: ${noOp} ? [] : [{ path: "/workspace/README.md", relativePath: "README.md", status: "modified", beforeMode: "100644", afterMode: "100644" }],
 }))
+${mismatch ? "await writeFile(join(process.cwd(), \"README.md\"), \"diverged\\n\")" : ""}
 await writeFile(join(process.cwd(), ".codebox", "order"), "runtime\\n")
 const summary = {
   schema: "wp-codebox/agent-task-run-result/v1",
@@ -190,4 +197,44 @@ const noOpMaintenance = await run("no-op-maintenance")
 assert.equal(noOpMaintenance.code, 0)
 assert(!noOpMaintenance.order.includes("publish"))
 
+const mismatch = await run("mismatch")
+assert.equal(mismatch.code, 1)
+assert.equal(await readFile(join(mismatch.workspace, "README.md"), "utf8"), "diverged\n", "base mismatch fails before the rejected patch can mutate the host workspace")
+assert.equal(mismatch.result.failure.stage, "apply")
+assert.match(mismatch.result.failure.message, /seed identity does not match/)
+assert.notEqual(mismatch.result.failure.evidence.expected_identity.content_digest.value, mismatch.result.failure.evidence.actual_identity.content_digest.value)
+assert.equal(mismatch.result.failure.evidence.patch.artifact_path, "files/patch.diff")
+assert.equal(mismatch.result.failure.evidence.changed_files.artifact_path, "files/changed-files.json")
+await execFileAsync(process.execPath, [uploader], {
+  cwd: mismatch.workspace,
+  env: {
+    ...process.env,
+    AGENT_TASK_WORKSPACE: mismatch.workspace,
+    AGENT_TASK_REQUEST_PATH: join(mismatch.workspace, ".codebox", "agent-task-request.json"),
+    AGENT_TASK_UPLOAD_PATH: join(mismatch.workspace, ".codebox", "agent-task-upload"),
+  },
+})
+assert.equal(await readFile(join(mismatch.workspace, ".codebox", "agent-task-upload", ".codebox", "agent-task-artifacts", "apply-failure", "rejected.patch"), "utf8"), "--- a/README.md\n+++ b/README.md\n@@ -1 +1 @@\n-before\n+after\n")
+assert.match(await readFile(join(mismatch.workspace, ".codebox", "agent-task-upload", ".codebox", "agent-task-artifacts", "apply-failure", "changed-files.json"), "utf8"), /README\.md/)
+
+const applyReject = await run("apply-reject")
+assert.equal(applyReject.code, 1)
+assert.equal(await readFile(join(applyReject.workspace, "README.md"), "utf8"), "before\n", "a rejected patch does not partially mutate the matching host workspace")
+assert.equal(applyReject.result.failure.stage, "apply")
+assert.match(applyReject.result.failure.message, /Host git apply failed/)
+assert.deepEqual(applyReject.result.failure.evidence.expected_identity, applyReject.result.failure.evidence.actual_identity, "the failure records the matching seed and host identities")
+assert.equal(applyReject.result.failure.evidence.patch.artifact_path, "files/patch.diff")
+assert.equal(applyReject.result.failure.evidence.changed_files.artifact_path, "files/changed-files.json")
+await execFileAsync(process.execPath, [uploader], {
+  cwd: applyReject.workspace,
+  env: {
+    ...process.env,
+    AGENT_TASK_WORKSPACE: applyReject.workspace,
+    AGENT_TASK_REQUEST_PATH: join(applyReject.workspace, ".codebox", "agent-task-request.json"),
+    AGENT_TASK_UPLOAD_PATH: join(applyReject.workspace, ".codebox", "agent-task-upload"),
+  },
+})
+assert.match(await readFile(join(applyReject.workspace, ".codebox", "agent-task-upload", ".codebox", "agent-task-artifacts", "apply-failure", "rejected.patch"), "utf8"), /-not-before/)
+assert.match(await readFile(join(applyReject.workspace, ".codebox", "agent-task-upload", ".codebox", "agent-task-artifacts", "apply-failure", "changed-files.json"), "utf8"), /README\.md/)
+
 console.log("native agent task lifecycle ok")

tests/execute-native-agent-task-playground-e2e.test.ts

@@ -31,6 +31,11 @@ try {
   await writeFile(join(workspace, "id_ed25519"), "PRIVATE_KEY_SENTINEL\n")
   await writeFile(join(workspace, ".codebox", "private.txt"), "PRIVATE_CODEBOX_SENTINEL\n")
   await writeFile(join(workspace, "node_modules", "private.txt"), "PRIVATE_NODE_MODULES_SENTINEL\n")
+  await execFileAsync("git", ["init", "--quiet"], { cwd: workspace })
+  await execFileAsync("git", ["config", "user.email", "tests@example.invalid"], { cwd: workspace })
+  await execFileAsync("git", ["config", "user.name", "Tests"], { cwd: workspace })
+  await execFileAsync("git", ["add", "."], { cwd: workspace })
+  await execFileAsync("git", ["commit", "--quiet", "-m", "seed"], { cwd: workspace })
   const packageBytes = Buffer.from(`${JSON.stringify({ schema_version: 1, bundle_slug: "fixture-agent", agent: { agent_slug: "fixture-agent", agent_name: "Fixture Agent", description: "Playground imported-agent fixture.", agent_config: { instructions: "Read and update README.", enabled_tools: ["workspace_read", "workspace_edit"], modes: ["chat"] } } })}\n`)
   await writeFile(packagePath, packageBytes)
   await mkdir(interceptor, { recursive: true })
@@ -68,10 +73,12 @@ add_filter( 'pre_http_request', static function( $preempt, $args, $url ) {
   assert.deepEqual(input.task_input.workspaces.map((entry: { target: string, mode: string }) => ({ target: entry.target, mode: entry.mode })), [{ target: "/workspace", mode: "readwrite" }])
   assert.notEqual(seededWorkspace.seed.source, workspace)
   const seedProvenance = input.task_input.runtime_task.input.metadata.runner_workspace_seed
-  assert.match(seedProvenance.digest.sha256, /^[a-f0-9]{64}$/)
+   assert.match(seedProvenance.digest.sha256, /^[a-f0-9]{64}$/)
+   assert.match(seedProvenance.identity.content_digest.value, /^[a-f0-9]{64}$/)
+   assert.match(seedProvenance.identity.git.head, /^[a-f0-9]{40}$/)
   assert.equal(seedProvenance.files, 2, "only README and the explicit .env.example template are copied")
-  assert.equal(seedProvenance.excluded.files, 6)
-  assert.deepEqual(seedProvenance.excluded.categories, [{ category: "credentials", count: 2 }, { category: "environment", count: 1 }, { category: "generated-tree", count: 2 }, { category: "private-key", count: 1 }])
+   assert.equal(seedProvenance.excluded.files, 7)
+   assert.deepEqual(seedProvenance.excluded.categories, [{ category: "credentials", count: 2 }, { category: "environment", count: 1 }, { category: "generated-tree", count: 3 }, { category: "private-key", count: 1 }])
   assert.equal(execution.code, 0, `${execution.stderr ?? ""}\n${JSON.stringify(result)}`)
   assert.equal(result.success, true)
   assert.equal(await readFile(join(workspace, "README.md"), "utf8"), "after\n")