Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

• No existing issue; see the description below.

2. Or, if no issue exists, describe the change:

Problem:
adk-python ships a VertexAiCodeExecutor that runs model-generated code on the
managed Vertex AI Code Interpreter Extension, but adk-js has no equivalent.
Users who want a managed, stateful, sandboxed Python execution backend
(returning stdout plus generated image/data files) have no option in the
TypeScript SDK. This is the last remaining code-executor parity gap (the
Container and GKE executors are tracked separately and are out of scope).

Solution:
Add VertexAiCodeExecutor
(core/src/code_executors/vertex_ai_code_executor.ts), a BaseCodeExecutor
that:

• Resolves a Code Interpreter extension in the reference order — explicit
  resourceName -> CODE_INTERPRETER_EXTENSION_NAME env var -> create a new one
  from the public hub manifest
  (gs://vertex-extension-public/code_interpreter.yaml) and write the created
  resource name back to the env var for reuse.
• Prepends the ported Python imported-library preamble to every code block, runs
  the extension's execute operation (operationId: 'execute', snake_case
  operationParams), and reuses the session_id across calls so state persists.
• Maps the response to stdout/stderr and File[], applying the reference
  MIME rules: png/jpg/jpeg -> image/<ext> (faithful, non-normalized so
  .jpg -> image/jpg), csv -> text/csv, everything else -> a best-effort
  guessMimeType fallback.

Because @google-cloud/vertexai exposes no Extensions API surface, the default
client talks to the extension over its public REST endpoints
(<location>-aiplatform.googleapis.com/v1beta1) using GoogleAuth (ADC,
cloud-platform scope) + fetch, mirroring the approach already used by other
Vertex-backed adk-js services. No new package dependencies are added. The
extension client is injectable for hermetic testing, the class sets
stateful = true / optimizeDataFile = true (so the request processor
populates executionId, enabling session reuse), and it is marked
@experimental and exported from the core barrel.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

core/test/code_executors/vertex_ai_code_executor_test.ts covers the executor
with an injected mock client, and the default REST client with mocked fetch +
GoogleAuth, reaching 100% line and branch coverage of the new file: extension
resolution order (explicit / env var / auto-create + env write-back / concurrent
single import), stateful / optimizeDataFile parity, the
Project ID is required. guard, session-id forwarding and reuse, input-file
forwarding, the imported-library preamble, all MIME mapping rules, empty/missing
output and stdout/stderr defaults, the correct :execute URL and content
envelope parsing, LRO import polling, and error paths (non-2xx, invalid resource
name, import timeout).

Run locally:

npm run build
npx vitest run --project unit:core core/test/code_executors/vertex_ai_code_executor_test.ts

Result:

 ✓ |unit:core| core/test/code_executors/vertex_ai_code_executor_test.ts (33 tests)

 Test Files  1 passed (1)
      Tests  33 passed (33)

npm run lint, npm run format:check, and npm run docs:check are also clean
for the touched files.

Manual End-to-End (E2E) Tests:

An env-gated live test lives at
tests/e2e/code_executors/vertex_ai_code_executor_e2e_test.ts. It is skipped by
default so CI stays hermetic, and only runs when explicitly opted in. It runs
Python twice in one session and asserts stdout, a generated .png (with
image/png MIME type), and persisted session state (a variable defined in the
first call is used in the second).

gcloud auth application-default login
export GOOGLE_CLOUD_PROJECT=<your-project>
# optional: export CODE_INTERPRETER_EXTENSION_NAME=<existing-extension>
RUN_VERTEX_AI_CODE_EXECUTOR_E2E=true npx vitest run --project e2e tests/e2e/code_executors

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

The live E2E test above is provided but left unchecked: it requires a real
Google Cloud project with the Vertex AI Code Interpreter Extension available, so
it is opt-in rather than part of the default suite. No dependent changes are
required — this PR is additive and introduces no new package dependencies.