Callback Hooks for Dapr Agents

[bibryam]

Context
Dapr agents currently lack a built-in mechanism to intercept, observe, or modify execution flow at critical points, such as before/after LLM or tool calls. Introducing a standardized callback interface enhances extensibility, debugging, and control, enabling advanced use cases like auditing, response caching, guardrails, and dynamic modification of inputs/outputs.

Acceptance Criteria

• Support four callback types:
• before_llm_call(context): invoked before calling the LLM.
• after_llm_call(context, response): invoked after receiving LLM response.
• before_tool_call(context): invoked before invoking a tool.
• after_tool_call(context, result): invoked after tool execution.
• Each callback receives a context object with execution metadata (e.g., agent name, session ID, invocation details).

Callbacks can:

• Log, inspect, or modify request data.
• Return None to proceed or return a specific value to override execution.

Callback return behavior:

• Returning a result from a before_* callback skips execution and uses that value.
• Returning a result from an after_* callback replaces the original output.