feat(types): preserve dropped fields on AssistantMessage and ResultMessage

[qing-ant]

Summary

The message parser was discarding several fields that the CLI emits in its JSON output. SDK consumers had no way to access important data like per-model usage breakdown (modelUsage), message IDs, UUIDs, and permission denials.

Changes

AssistantMessage — 4 fields added

CLI field	SDK field	Source
message.id	message_id	Anthropic API message ID
message.stop_reason	stop_reason	Per-turn stop reason (end_turn, tool_use, etc.)
session_id	session_id	Session identifier
uuid	uuid	SDK message UUID for transcript correlation

ResultMessage — 3 fields added

CLI field	SDK field	Source
modelUsage	model_usage	Per-model usage breakdown (tokens, cost, context window)
permission_denials	permission_denials	List of denied permission requests
uuid	uuid	SDK message UUID for transcript correlation

All new fields are optional with None defaults for backward compatibility.

E2E Verification

Verified with a live SDK instance (v0.1.50, CLI 2.1.83):

=== AssistantMessage ===
  message_id: msg_013yKzePkD82yiUKxLjeMisU
  session_id: d815620f-c625-4179-9fd0-8fe26ef523d3
  uuid: 2acfe9fc-bab6-4829-9ac4-7894f36205e1

=== ResultMessage ===
  uuid: 582ae113-f2e5-43c2-90dc-9fcb0ec760da
  model_usage: True (per-model breakdown with costUSD)
  permission_denials: []

Tests

4 new test cases:

• test_parse_assistant_message_with_all_fields — verifies all new fields are parsed
• test_parse_assistant_message_optional_fields_absent — verifies defaults to None
• test_parse_result_message_with_model_usage — verifies modelUsage, permission_denials, uuid
• test_parse_result_message_optional_fields_absent — verifies defaults to None

All 292 tests pass. Lint clean (ruff), type check clean (mypy).

Fixes #562

[claude]

LGTM — straightforward additive change preserving optional CLI fields with full test coverage.

Extended reasoning...

Overview

This PR adds 7 optional fields across two dataclasses (AssistantMessage and ResultMessage) in types.py, extracts them via .get() in message_parser.py, and adds 4 test cases in test_message_parser.py. All new fields default to None for backward compatibility.

Security risks

None. The changes only preserve data that the CLI already emits — no new inputs are processed, no auth/permissions logic is touched, and no external APIs are called.

Level of scrutiny

Low scrutiny is appropriate. This is a mechanical, additive change that follows the exact same pattern already used throughout the parser (e.g., usage=data["message"].get("usage") was already present). The dataclass fields are all optional with None defaults, so there is zero risk of breaking existing consumers.

Other factors

The 4 new tests comprehensively cover both the "all fields present" and "all fields absent" scenarios for both message types. The PR description includes E2E verification against a live CLI. The change is small (7 new lines in types, 7 in parser, ~90 in tests) and self-contained.

[qing-ant]

E2E Test Results for PR #718

Test: Live claude_agent_sdk.query() call verifying that previously-dropped fields are now preserved on AssistantMessage and ResultMessage.

AssistantMessage Fields (Previously Dropped)

Field	Status	Sample Value
message_id	POPULATED	msg_01BqLK6eNJg8jqwqUDG5Lk7U
stop_reason	None	Expected for intermediate assistant chunks; final stop_reason appears on ResultMessage
session_id	POPULATED	9b56dedc-4e43-4bab-8812-e475dcab0c0c
uuid	POPULATED	3cbe1b7a-306f-4bf6-b7e1-bb81779e5097

ResultMessage Fields (Previously Dropped)

Field	Status	Sample Value
model_usage	POPULATED	Per-model breakdown with inputTokens, outputTokens, cacheReadInputTokens, costUSD, etc.
permission_denials	POPULATED	[] (empty list, no denials in this run -- correctly not None)
uuid	POPULATED	32683150-a744-440c-87bc-75ba623d0eef

Test Summary

Unit tests:           43/43 passed (test_message_parser.py)
AssistantMessage fields accessible: PASS
ResultMessage fields accessible:    PASS
AssistantMessage dataclass has all new fields: PASS
ResultMessage dataclass has all new fields: PASS
OVERALL RESULT: PASS

All previously-dropped fields are now correctly preserved and populated by the CLI. The stop_reason on AssistantMessage being None is expected behavior -- the CLI only sets it on intermediate messages when the stop reason is available at that point; the authoritative stop_reason is on ResultMessage (where it correctly shows end_turn).