Dapr Agents: document response_metadata on after_llm_call hooks

[CasperGN]

Description

Documents the response_metadata key that dapr/dapr-agents#752 added to the after_llm_call hook payload. With it, a hook can read a call's token usage without wrapping the LLM client.

Changes to developing-ai/dapr-agents/dapr-agents-hooks.md:

• The HookContext.payload row now mentions response_metadata for after_llm_call.
• The LLM hooks section now gives the after_llm_call signature (context plus a copy of the assistant message).
• New "Reading response metadata" section:
  • The common keys (provider, id, model, usage) and the fields some providers add on top.
  • How usage field names differ between OpenAI-style clients and AnthropicChatClient.
  • When the key is absent: structured output, streamed calls, a before_llm_call Skip/Deny, or empty metadata.
  • An example hook that logs token usage.
  • A note on hook order: on after_llm_call only Mutate stops later hooks, so an accounting hook placed after a Mutate hook is skipped whenever that hook returns Mutate. The page's general "first non-Proceed wins" rule now says so too.
• New row in "When to use which slot": per-call token usage.

The behaviour was checked against dapr_agents/agents/durable.py (call_llm) and dapr_agents/hooks.py on dapr-agents main.

Issue reference

Documents dapr/dapr-agents#752.

Checklist

• Doc conventions followed (matches the page's existing tables and code style)
• In-page anchors resolve
• Built locally with Hugo

🤖 Generated with Claude Code