feat(vertex_ai): support GenerateContentConfig labels for Dify app_id propagation

[mas-sakai]

Note

This is a Draft PR. It depends on
langgenius/dify-plugin-sdks#313
being merged and a new SDK released. pyproject.toml / uv.lock currently
pin dify_plugin to the #313 branch so the Gemini route can import
get_current_session(). The pin will be reverted to a released version
constraint (dify_plugin>=0.3.0,<0.6.0) before this PR is marked Ready
for review. The CI "Validate Dependencies" / "Check Plugin Install"
steps may fail until then.

Summary

Enable Cloud Billing visibility per Dify app for the Vertex AI plugin (Gemini
route). When the new enable_request_metadata credential is set to enabled,
the plugin attaches dify_app_id and dify_source to
GenerateContentConfig.labels, so operators can break Vertex AI cost down by
labels.dify_app_id in GCP Cloud Billing without any ETL.

Related issues / PRs:

• langgenius/dify#35772 — host issue (app_id propagation)
• langgenius/dify-plugin-sdks#311 — SDK issue
• langgenius/dify#35859 — host PR (contextvars-based app_id propagation)
• langgenius/dify-plugin-sdks#313 — SDK PR (get_current_session() public API) — this PR depends on it

Scope (intentionally narrow)

• Gemini route only (models/llm/llm.py::_generate). The Anthropic-on-Vertex
  route (_generate_anthropic) and other providers are out of scope.
• Label keys: dify_app_id and dify_source only. dify_tenant_id /
  dify_user_id are deferred to a follow-up to keep the surface minimal.
• Default-off: enable_request_metadata defaults to disabled, so existing
  users who do not touch their credentials are completely unaffected. With
  disabled, the control flow is identical to the previous version.

Implementation notes

• New helper models/llm/_labels.py with normalize_label_value and
  build_dify_labels. Output conforms to the GCP labels spec (lowercase,
  [a-z0-9_-] only, 63-char cap). UUID app_ids pass through unchanged.
• get_current_session() is imported locally inside _generate and wrapped in
  try/except ImportError, so the plugin keeps working against older
  dify_plugin releases (no hard runtime dependency on fix getnumtoken retrun list[int] #313 once it ships).
• 9 unit tests added in tests/test_labels.py (passthrough, character
  replacement, truncation, casing, empty input, None / empty short-circuits
  in build_dify_labels).

Change Type

• Documentation / non-plugin change
• Non-LLM plugin (tools, extensions, datasource, etc.)
• LLM plugin

Screenshots / Videos

🚧 Cloud Billing breakdown screenshots (by labels.dify_app_id) will be
added once E2E verification completes (currently testing on GCP project
dify-vertex-labels-poc). All 9 unit tests pass (uv run pytest tests/).

Before	After

LLM Plugin Checklist

Areas affected by this change (check all that apply)

• Message flow (system messages, user ↔ assistant turn-taking)
• Tool interaction flow (multi-round usage, Agent App and Agent Node)
• Multimodal input (images, PDFs, audio, video, etc.)
• Multimodal output (images, audio, video, etc.)
• Structured output (JSON, XML, etc.)
• Token consumption metrics
• Other LLM functionality (reasoning, grounding, prompt caching, etc.) — request metadata / Cloud Billing labels
• New models / model parameter fixes

Version

• Bumped top-level version in manifest.yaml (not the one under meta) — 0.0.53 → 0.0.54
• dify_plugin>=0.3.0,<0.6.0 is declared in pyproject.toml and locked in uv.lock (or kept in requirements.txt for legacy plugins without uv.lock) — SDK docs

⚠️ The dify_plugin constraint is currently a temporary git-branch pin
targeting #313.
It will be reverted to dify_plugin>=0.3.0,<0.6.0 before this PR leaves
Draft.

Testing

• Local deployment — Dify version: 1.14.x (E2E Cloud Billing breakdown verification in progress on GCP project dify-vertex-labels-poc)
• SaaS (cloud.dify.ai)

Unit tests: uv run pytest tests/ — 9 passed (label normalization helper).

[gemini-code-assist]

Code Review

This pull request introduces the capability to track Cloud Billing costs on a per-application basis by attaching Dify metadata as labels to Vertex AI requests. The implementation includes a utility for normalizing label values according to GCP constraints, logic in the Gemini model to extract session metadata, and a new configuration setting to enable this feature. Feedback suggested optimizing the label normalization helper by truncating long input strings prior to processing to enhance performance and reduce memory usage.