parseSseChunk does not unwrap the SSE envelope — tool_result AND done diverge from Python (TS↔Python drift)

[AndreLFSMartins]

Summary

parseSseChunk in packages/core/src/session.ts assigns the entire SSE data object directly for envelope-style events instead of unwrapping the inner payload the way the Python SDK does. This affects two event types (same root cause):

Event	TS (parseSseChunk)	Python (_parse_stream_event)
tool_result	toolCall: data (whole object)	raw.get("toolCall") → inner ToolCallRecord
done	result: data as SendResult (whole object)	raw.get("result") → inner SendResult

Python (packages/python/src/codespar/_async_session.py) and its tests (packages/python/tests/test_streaming.py) treat the canonical wire payload as { "type": "<evt>", "toolCall"|"result": { … } }. TS does not unwrap, so against that canonical payload:

• tool_result → event.toolCall.tool_name is undefined (record is one level deeper).
• done → event.result.message is undefined (SendResult is one level deeper).

A single backend payload cannot satisfy both SDKs for either event. Sibling of #51 (malformed SSE); same family of parseSseChunk wire drift.

Consequence

Any TS sendStream consumer doing event.toolCall.tool_name or event.result.message (the documented StreamEvent shape in @codespar/types) gets undefined on canonical backend payloads — a silent logic failure, not a network error. Forbidden by CLAUDE.md ("NEVER let TS + Python drift").

Decision needed

Confirm the backend's real tool_result and done data shapes (codespar-enterprise Fastify routes), then align parseSseChunk to unwrap data.toolCall / data.result (mirroring Python) — or, if the backend truly sends the bare objects, stop unwrapping in Python. Parser/contract change → escalated for a maintainer decision, not changed unilaterally.

Acceptance criteria

• TS and Python yield an equivalent ToolCallRecord (tool_result) and SendResult (done) from the same canonical payloads, with regression tests on both sides using identical fixtures.

Test coverage status

PR #52 pins both as narrow drift tripwires (normal passing tests asserting the current TS-wrapped shape; they fail loudly when the parser is aligned, forcing this issue to be closed deliberately). No parser change in #52 (test-only).