Add agent_id/agent_type to tool-lifecycle hook inputs

[james-m-thomas]

What

Surfaces the agent_id / agent_type fields the CLI started sending in v2.1.64 (anthropics/claude-cli-internal#20180) on the four tool-lifecycle hook input types: PreToolUseHookInput, PostToolUseHookInput, PostToolUseFailureHookInput, PermissionRequestHookInput.

Why

When the parent spawns multiple Task sub-agents in parallel, their PreToolUse/PostToolUse hooks interleave over the same control channel with nothing tying each callback to the sub-agent that fired it:

SubagentStart(agent_id=A)
PreToolUse(tool=databricks, tuid=t1)        ← only A is alive, so A ✓
SubagentStart(agent_id=B)
PostToolUse(tuid=t1)
PreToolUse(tool=databricks, tuid=t2)        ← A's follow-up? B's first? ambiguous
SubagentStart(agent_id=C)
PreToolUse(tool=databricks, tuid=t3)        ← A, B, or C? ambiguous

session_id and transcript_path are both parent-level and don't discriminate between sub-agents. Once more than one sub-agent is unblocked, there's no heuristic that can disambiguate — an older sub-agent might be writing its final response or about to fire a follow-up query; both states look identical from the hook stream.

agent_id is the correlation key: it's the same value the CLI emits in that sub-agent's SubagentStart/SubagentStop. For main-thread tool calls it's absent (check .get("agent_id") is None).

Implementation note

The CLI puts these on BaseHookInput, but doing that here would conflict with SubagentStartHookInput/SubagentStopHookInput, which already declare agent_id: str as required — PEP 655 forbids narrowing NotRequired → Required in a TypedDict subclass. So this uses a total=False mixin applied to exactly the four hook types the CLI populates. Those are also the only callsites where the CLI has a toolUseContext in scope, so it's a wash semantically.

Scope

Types only. CHANGELOG.md and _cli_version.py are intentionally untouched: the auto-release workflow generates changelog entries and keys on the exact chore: bump bundled CLI version to X.Y.Z commit message to trigger a release. Bundling the bump here would silently skip the 2.1.64 auto-release.

The NotRequired fields are inert until the bundled CLI reaches 2.1.64 via the normal bump-commit flow, at which point both sides line up.

Test plan

• ruff check / ruff format — clean
• mypy src/claude_agent_sdk/types.py — clean
• pytest tests/ — 246/246 pass (2 new)
  • PreToolUseHookInput with agent_id present (sub-agent) and absent (main thread) — both type-valid
  • PostToolUseHookInput with agent_id present