Skip to content

feat(source/cloud-storage): add Cloud Storage source with list_objects and read_object tools#3081

Open
huangjiahua wants to merge 13 commits intogoogleapis:mainfrom
huangjiahua:feat/cloud-storage-source
Open

feat(source/cloud-storage): add Cloud Storage source with list_objects and read_object tools#3081
huangjiahua wants to merge 13 commits intogoogleapis:mainfrom
huangjiahua:feat/cloud-storage-source

Conversation

@huangjiahua
Copy link
Copy Markdown

@huangjiahua huangjiahua commented Apr 16, 2026

Description

Adds Google Cloud Storage as a first-class source in MCP Toolbox, enabling LLM agents to work with objects across buckets in a GCP project. The source is project-scoped and authenticates via Application Default Credentials, mirroring Firestore/Bigtable.

This first PR ships the source plus two read-only tools from the approved design (14 total):

  • cloud-storage-list-objects — prefix filter, delimiter-based grouping (returns prefixes), and pagination via max_results / page_token. Passes through whatever metadata the GCS client returns (*storage.ObjectAttrs) so we don't have to plumb new fields later.
  • cloud-storage-read-object — reads an object's bytes, textual data only, with optional HTTP-style byte ranges (bytes=0-999, bytes=-500, bytes=500-).

GCS-aware error categorization (per DEVELOPER.md) is implemented in a new cloudstoragecommon helper that maps GCS sentinels and *googleapi.Error codes to Agent errors (missing bucket/object, bad request, unsatisfiable range) vs. Server errors (auth, IAM denial, quota, 5xx, context cancellation). This replaces the coarse util.ProcessGcpError for the two new tools.

Remaining 12 tools from the design doc (list_buckets, create_bucket, copy/move/delete_object, etc.) will land in follow-up PRs.

CI note: the cloud-storage shard in .ci/integration.cloudbuild.yaml expects CLOUD_STORAGE_PROJECT=$PROJECT_ID and requires the test service account to have a Cloud Storage admin role in the test project. Integration test self-manages its own UUID-suffixed bucket with defer-based cleanup.

PR Checklist

  • Make sure you reviewed CONTRIBUTING.md
  • Make sure to open an issue as a bug/issue before writing your code! That way we can discuss the change, evaluate designs, and agree on the general idea (communicated internally)
  • Ensure the tests and linter pass
  • Code coverage does not decrease (if any source code was changed)
  • Appropriate docs were updated (if necessary)
  • Make sure to add ! if this involve a breaking change

What's included

  • New source: internal/sources/cloudstorage/ (+ YAML-parse unit tests)
  • Two tools: internal/tools/cloudstorage/cloudstoragelistobjects/, .../cloudstoragereadobject/ (+ YAML-parse + range-parser unit tests)
  • New cloudstoragecommon error classifier (+ 17-case unit test covering sentinels, HTTP statuses, context.Canceled/DeadlineExceeded, and fallback)
  • Integration test: tests/cloudstorage/cloud_storage_integration_test.go — 12 sub-tests against a real bucket (self-created, self-cleaned)
  • Docs: docs/en/integrations/cloud-storage/ (source + both tool pages; passes .ci/lint-docs-{source,tool}-page.sh)
  • CI shard: cloud-storage in .ci/integration.cloudbuild.yaml
  • Dependency: cloud.google.com/go/storage v1.62.1

Opening as draft for initial review — happy to split the error-classifier refactor into a separate commit if reviewers prefer.

Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request adds Google Cloud Storage integration, introducing a new source and tools for listing and reading objects. The implementation includes configuration, error handling, and tests. Feedback recommends capping listing page sizes at 1000 for consistency, implementing memory safety limits when reading objects, and updating documentation titles to include the 'Tool' suffix.

Comment thread internal/sources/cloudstorage/cloudstorage.go
Comment thread internal/sources/cloudstorage/cloudstorage.go
@huangjiahua huangjiahua marked this pull request as ready for review April 16, 2026 23:28
@huangjiahua huangjiahua requested a review from a team as a code owner April 16, 2026 23:28
…s and read_object tools

Adds a new project-scoped `cloud-storage` source using ADC, plus two read-only
tools: `cloud-storage-list-objects` (with prefix/delimiter/pagination) and
`cloud-storage-read-object` (with HTTP-style byte range and base64 payload).

Introduces a GCS-aware error classifier in `cloudstoragecommon` that splits
failures into Agent errors (missing bucket/object, bad request, unsatisfiable
range) and Server errors (auth, IAM denial, quota, 5xx, cancellation) per
DEVELOPER.md, replacing the coarse-grained `util.ProcessGcpError`.

Ships YAML-parse unit tests, an error-classifier unit test, a range-parser unit
test, a live-GCS integration test (12 sub-tests, UUID-suffixed bucket with
self-cleanup), docs under `docs/en/integrations/cloud-storage/`, and a
`cloud-storage` CI shard.

The remaining 12 tools from the approved design doc land in follow-up PRs.
…dObject at 1 MiB

- ListObjects: pageSize() now clamps to the GCS API max of 1000 so callers that
  pass a larger max_results don't pre-allocate oversized buffers.
- ReadObject: reject objects/ranges over 1 MiB with the new sentinel
  cloudstoragecommon.ErrReadSizeLimitExceeded, which the classifier maps to an
  Agent error so the LLM can retry with a narrower 'range'.
- Docs + integration tests updated (two new sub-tests: oversize rejection and
  oversize-narrowed-by-range success).
… MiB

8 MiB gives agents more headroom for typical text/JSON/log payloads while
still guarding against OOM. Doc and the oversize integration seed updated to
match.
…ckage

DefaultMaxReadBytes doesn't belong in errors.go — the limit is a source-side
invariant, not an error-classification concern. The sentinel
ErrReadSizeLimitExceeded stays in cloudstoragecommon because the classifier
still needs to recognize it.
…geSize bounds

Cleanup loop in the integration test was treating any iterator error as
iterator.Done; now distinguishes the two and logs non-Done errors so
flaky teardowns are debuggable. Also adds an internal unit test for
pageSize covering 0, negative, in-range, and over-cap inputs.
MCP tool results only carry text today, so the previous base64-encoded
content was unusable by the LLM. Validate object bytes with utf8.Valid
and return plain-text content; non-UTF-8 objects surface as an
agent-fixable ErrBinaryContent error. TODO notes mark the spots to
revisit once MCP supports embedded resources.
@huangjiahua huangjiahua force-pushed the feat/cloud-storage-source branch from 91a222a to 4919821 Compare April 17, 2026 19:30
Copy link
Copy Markdown
Contributor

@Yuan325 Yuan325 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @huangjiahua Thank you for the contribution! Please let me know if you need any clarifications


type Source struct {
Config
Client *storage.Client
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
Client *storage.Client
client *storage.Client

Probably don't need to export this~

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we move this test function into cloudstorage_test.go instead?

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

// results at 1000 per page; we enforce the same cap here so callers don't
// pre-allocate larger buffers and so the contract matches the tool's
// 'max_results' documentation.
func pageSize(maxResults int) int {
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we trigger an AgentError in the tool during parameter extraction if the value exceeds 1,000? This makes the limit explicit to the agent/user, preventing confusion when the returned page count is lower than requested.

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, I did that in internal/tools/cloudstorage/cloudstoragelistobjects/cloudstoragelistobjects.go

func (s *Source) ListObjects(ctx context.Context, bucket, prefix, delimiter string, maxResults int, pageToken string) (map[string]any, error) {
it := s.Client.Bucket(bucket).Objects(ctx, &storage.Query{
Prefix: prefix,
Delimiter: delimiter,
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

just confirming, will be okay if these 2 values are ""?

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can confirm. I've also add integration test with these 2 values being "".

Prefix: prefix,
Delimiter: delimiter,
})
pager := iterator.NewPager(it, pageSize(maxResults), pageToken)
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

will this be okay if pageToken is ""

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes. Also added an integration test case.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Let's just move this test to cloudstoragereadobject_test.go

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

}

func initStorageClient(ctx context.Context) (*storage.Client, error) {
return storage.NewClient(ctx, option.WithUserAgent("genai-toolbox-integration-test"))
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
return storage.NewClient(ctx, option.WithUserAgent("genai-toolbox-integration-test"))
return storage.NewClient(ctx)

Can we just init without the user agent option for int test?

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

t.Fatalf("toolbox didn't start successfully: %s", err)
}

runCloudStorageToolGetTest(t)
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Probably wouldn't need this function. We can just utilize this existing function

func RunToolGetTestByName(t *testing.T, name string, want map[string]any) {

ref:

tests.RunToolGetTestByName(t, "get_models",
map[string]any{
"get_models": map[string]any{
"description": "Simple tool to test end to end functionality.",
"authRequired": []any{},
"parameters": []any{},
},
},
)

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do you mean we can remove this test?

}
}

func runCloudStorageListObjectsTest(t *testing.T, bucket string) {
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For this function and runCloudStorageReadObjectTest(), is it possible to use the golang's table-driven tests? There's probably alot of duplication here in each t.Run().

Reference:

func RunToolInvokeTest(t *testing.T, select1Want string, options ...InvokeTestOption) {

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

The storage.Client is an implementation detail; external callers that
need it use the StorageClient() accessor, so the field itself doesn't
need to be exported.
…e tests into single test file per package

Merge TestPageSize into cloudstorage_test.go and TestParseRange into
cloudstoragereadobject_test.go. Both test files now use the internal
package so they can exercise the unexported pageSize and parseRange
helpers directly, removing the need for separate *_internal_test.go
files.
…with AgentError

Previously, values above the GCS per-page cap of 1000 were silently
clamped by the pageSize helper, which could confuse agents when the
returned page was smaller than requested. Validate max_results during
Invoke and return an AgentError so the limit is explicit. Docs and the
parameter description are updated to match; the pageSize clamp remains
as defense in depth. A unit test covers the rejection path and an
integration test exercises it over HTTP.
…age_token inputs

Add two integration sub-tests confirming that empty-string inputs are
accepted by the GCS client as expected: ListObjects with empty prefix
and delimiter returns an unfiltered listing, and an empty page_token
returns the first page rather than erroring. These cases address
review questions about whether the values passed through to
storage.Query and iterator.NewPager are safe when unset.
… simplify storage client init in integration test

Drop the initStorageClient wrapper and the option.WithUserAgent call;
the integration test now uses storage.NewClient(ctx) directly, matching
the suggestion in review and removing a needless indirection.
… table-drive integration tests and reuse RunToolGetTestByName

Replace the bespoke runCloudStorageToolGetTest with two
tests.RunToolGetTestByName calls that assert the full manifest for
each tool. Convert the list_objects and read_object sub-tests to
table-driven form: each case declares a request body plus substring,
content, or contentType expectations, driven by a single assertion
loop. The inherently two-step pagination test stays as its own
t.Run. Behaviour is unchanged; the file is ~220 lines shorter in
boilerplate.
… drop tool-get manifest test

Remove runCloudStorageToolGetTest entirely. The manifest-shape check
it performed was redundant: unit tests already cover ParseFromYaml for
each config, and the invoke sub-tests exercise the tool handlers over
HTTP. Keeping a full-manifest deep-equal here just duplicates that
coverage and has to be updated whenever parameter docs change.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants