Skip to content

feat: Add TealTiger governance integration for deterministic policy enforcement #6418

Description

@nagasatish007

Issue Body

Summary

We've published tealtiger-adk — a standalone governance callback package that integrates with Google ADK's before_tool_callback / after_tool_callback hooks to provide deterministic policy enforcement, PII detection, cost tracking, and audit evidence for ADK agents.

What it does

tealtiger-adk provides a TealTigerCallback class that plugs into ADK's existing callback system:

from google.adk import Agent
from tealtiger_adk import TealTigerCallback

governance = TealTigerCallback(
    policies=[
        {"type": "pii_block", "categories": ["ssn", "credit_card"]},
        {"type": "cost_limit", "max_per_session": 5.00},
        {"type": "tool_allowlist", "allowed": ["google_search", "code_*"]},
        {"type": "secret_detection"},
    ],
    mode="ENFORCE",  # OBSERVE | MONITOR | ENFORCE
)

agent = Agent(
    model="gemini-2.0-flash",
    tools=[google_search, code_execution],
    before_tool_callback=governance.before_tool,
    after_tool_callback=governance.after_tool,
)

Key Features

Feature Description
PII Detection Block SSN, credit cards, emails, phone numbers, IP addresses
Secret Detection Prevent API key leaks (OpenAI, GitHub, AWS, Google, Groq)
Cost Tracking Per-session budget limits with automatic enforcement
Tool Allowlisting Restrict tools with wildcard support
Kill Switch Instant freeze/unfreeze of agent tool access
Audit Trail Structured decision log with correlation IDs and risk scores

Architecture

  • Zero LLM in governance path — all evaluation is deterministic regex/rule-based
  • Sub-millisecond overhead — <0.1ms per evaluation (30 tests pass in <1s total)
  • Three modes: OBSERVE (log only), MONITOR (log + warn), ENFORCE (log + block)
  • Native ADK integration — uses before_tool_callback/after_tool_callback exactly as designed

Why this matters for ADK users

  1. Compliance — Prove governance was applied (audit trail with correlation IDs)
  2. Safety — Prevent PII/secrets from leaking through tool calls
  3. Cost control — Budget limits prevent runaway agent costs
  4. Kill switch — Emergency stop without redeploying

Installation

pip install tealtiger-adk

Links

Request

Would the team be open to:

  1. Listing tealtiger-adk in the ADK ecosystem/integrations documentation?
  2. Any feedback on how we're using the callback API — happy to adjust to better fit ADK patterns.

We're part of NVIDIA Inception and have existing integrations with LangChain, Haystack, AG2, CrewAI, Strands, and PydanticAI.

Labels

enhancement, integration

Metadata

Metadata

Assignees

No one assigned

    Labels

    spam[Status] Issues suspected of having comments which are spam

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions