feat: add context management configuration (caching and compaction)

[dobesv]

Summary

Add context management configuration (event compaction and context caching) for kagent agents, wiring the feature through the full pipeline: CRD → Go controller → config.json → Python runtime → Google ADK App().

This completes the work started in #1198 (stalled since January) with a clean implementation from scratch, addressing all identified wiring gaps.

Changes

Go CRD Types (go/api/v1alpha2/agent_types.go)

• Add Context *ContextConfig field to DeclarativeAgentSpec
• New types: ContextConfig, ContextCompressionConfig, ContextCacheConfig, ContextSummarizerConfig
• CEL validation for required compaction fields
• Kubebuilder default values matching ADK defaults

Go ADK Config Types (go/pkg/adk/types.go)

• Add AgentContextConfig, AgentCompressionConfig, AgentCacheConfig types for config.json serialization
• Critical: Update UnmarshalJSON to include new ContextConfig field (prevents silent data loss)

Go Translator (go/internal/controller/translator/agent/adk_api_translator.go)

• Extract context config from Agent CRD in translateInlineAgent()
• Resolve summarizer model name from ModelConfig CRD (falls back to agent's own model)
• 6 table-driven test cases covering all combinations

Python Types (python/packages/kagent-adk/src/kagent/adk/types.py)

• Add ContextCompressionSettings, ContextCacheSettings, ContextConfig Pydantic models
• Add build_adk_context_configs() builder function that converts kagent types to Google ADK EventsCompactionConfig and ContextCacheConfig
• Uses LiteLlm for universal summarizer model support

Python Runtime Wiring

• KAgentApp accepts and passes events_compaction_config and context_cache_config to ADK App()
• All 3 CLI commands (static, run, test) wire context configs from AgentConfig

Tests

• Go: 6 translator test cases (no config, compaction only, cache only, both, summarizer with agent model, summarizer with separate model)
• Python: 17 test cases covering type parsing, builder function, JSON round-trip with Go output format, edge cases

Example Usage

apiVersion: kagent.dev/v1alpha2
kind: Agent
metadata:
  name: my-agent
spec:
  type: Declarative
  declarative:
    modelConfig: default-model-config
    systemMessage: "You are a helpful assistant"
    context:
      compaction:
        compactionInterval: 10
        overlapSize: 3
        summarizer:
          promptTemplate: "Summarize the conversation so far"
      cache:
        cacheIntervals: 10
        ttlSeconds: 1800

Related

• Completes work from add feature: config summarisation, context compression, caching #1198
• Uses Google ADK experimental EventsCompactionConfig and ContextCacheConfig APIs

[Copilot]

Pull request overview

This PR adds context management configuration (event compaction and context caching) for kagent agents, completing work from stalled PR #1198. The feature is wired through the complete pipeline: CRD definitions → Go controller → config.json → Python runtime → Google ADK App().

Changes:

• Added CRD types for context management (compaction and caching) with CEL validation and kubebuilder defaults
• Extended Go ADK config types and translator to handle context configuration, including model resolution for summarizers
• Implemented Python types and builder function to convert kagent configs to Google ADK objects (EventsCompactionConfig and ContextCacheConfig)
• Wired context configs through all Python CLI commands (static, run, test) to KAgentApp and ADK App()

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
go/api/v1alpha2/agent_types.go	Defines CRD types for context management: ContextConfig, ContextCompressionConfig, ContextCacheConfig, ContextSummarizerConfig with validation
go/api/v1alpha2/zz_generated.deepcopy.go	Auto-generated DeepCopy methods for new context config types
go/config/crd/bases/kagent.dev_agents.yaml	Generated CRD YAML schema with validation rules and default values
go/pkg/adk/types.go	ADK config types for JSON serialization and critical UnmarshalJSON update to include ContextConfig
go/internal/controller/translator/agent/adk_api_translator.go	Translates CRD context config to ADK config, resolves summarizer models with fallback to agent model
go/internal/controller/translator/agent/adk_api_translator_test.go	6 comprehensive test cases covering all context config scenarios
python/packages/kagent-adk/src/kagent/adk/types.py	Pydantic models and builder function to convert kagent types to Google ADK configs
python/packages/kagent-adk/src/kagent/adk/_a2a.py	KAgentApp accepts and passes context configs to ADK App()
python/packages/kagent-adk/src/kagent/adk/cli.py	All CLI commands wire context configs via helper function
python/packages/kagent-adk/tests/unittests/test_context_config.py	17 test cases covering parsing, builder function, and JSON round-trip validation

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The ContextSummarizerSettings class is defined but never used anywhere in the codebase. The summarizer_model_name and prompt_template fields are duplicated in ContextCompressionSettings instead. This class should either be removed as dead code, or ContextCompressionSettings should be refactored to use it as a nested field (e.g., summarizer: ContextSummarizerSettings | None = None).

Suggested change

	class ContextSummarizerSettings(BaseModel):
	"""Settings for the LLM-based event summarizer."""
	summarizer_model_name: str | None = None
	prompt_template: str | None = None

[Copilot]

Unused imports: MagicMock and patch are imported from unittest.mock but never used in the test file. These should be removed to keep the code clean.

Suggested change

from unittest.mock import MagicMock, patch

[Copilot]

The EventsCompactionConfig is created but token_threshold and event_retention_size fields from comp are not passed to it. These fields are defined in the CRD (agent_types.go), collected by the Go translator (adk_api_translator.go), and present in ContextCompressionSettings, but are silently dropped here.

If EventsCompactionConfig supports these parameters, they should be passed through. If not, either remove them from the earlier stages or document why they're being ignored.

Suggested change

	events_compaction_config = EventsCompactionConfig(
	compaction_interval=comp.compaction_interval,
	overlap_size=comp.overlap_size,
	summarizer=summarizer,
	)
	compaction_kwargs = {
	"compaction_interval": comp.compaction_interval,
	"overlap_size": comp.overlap_size,
	"summarizer": summarizer,
	}
	# Optionally pass through additional compaction settings if available.
	token_threshold = getattr(comp, "token_threshold", None)
	if token_threshold is not None:
	compaction_kwargs["token_threshold"] = token_threshold
	event_retention_size = getattr(comp, "event_retention_size", None)
	if event_retention_size is not None:
	compaction_kwargs["event_retention_size"] = event_retention_size
	events_compaction_config = EventsCompactionConfig(**compaction_kwargs)