Fixes #1139.
Fixes #1140.
Fixes #1141.

Why

from_dict(to_dict(x)) was not the identity for three generated Python dataclasses whenever the affected optional field was None. The dataclass declares the field as T | None = None and to_dict() omits it when None, but from_dict() was substituting the JSON-Schema default value when the key was missing, silently mutating unset fields:

• SessionTaskCompleteData.summary: None → ""
• PermissionPromptRequest.action: None → PermissionPromptRequestMemoryAction.STORE
• PermissionRequest.action: None → PermissionRequestMemoryAction.STORE

Any code that round-trips events (caching, replay, audit logs, dedup) or pattern-matches on None to mean "absent" was getting the wrong answer end-to-end. The bug was reproducible on main as of today.

What changed

Root cause is in scripts/codegen/python.ts: for optional fields with a schema default, the codegen was emitting obj.get("key", "<default>"), then passing that into from_union([from_none, parse_X], ...), which happily parsed the default and returned it. The dataclass default and the wire default were inconsistent, so the round-trip broke.

• Dropped the defaultLiteral mechanism from both emitPyClass and emitPyFlatDiscriminatedUnion. from_dict() now uniformly emits obj.get(key), so missing/null keys flow through from_none and land at the dataclass-declared None. Removed the now-unused toPythonLiteral helper.
• Regenerated python/copilot/generated/session_events.py via npm run generate:python. Exactly three call sites changed, all matching the bug reports.
• Flipped the two codegen-level assertions in nodejs/test/python-codegen.test.ts that previously locked in the buggy output, and added negative assertions to prevent regression at the generator level.
• Replaced test_schema_defaults_are_applied_for_missing_optional_fields in python/test_event_forward_compatibility.py (which asserted the bug as expected behavior) with two regression tests covering missing-key parsing, explicit-null parsing, and the full from_dict(to_dict(x)) round-trip for all three affected classes.

Surface area / non-obvious notes

• Other languages (Go, .NET, Rust, TypeScript) are not affected. Their generators never read propSchema.default to construct deserialization fallbacks; they use real nullable types end-to-end.
• Grep on the post-fix generated session_events.py and rpc.py confirms zero remaining obj.get("...", "...") patterns, so no other class was silently relying on baked-in schema defaults.
• The test_schema_defaults_are_applied_for_missing_optional_fields test was the only consumer in the repo of the old behavior. Nothing in runtime code depended on request.action == STORE or task_complete.summary == "" for unset fields.
• JSON Schema default is an annotation, not validation behavior, so applying it only during deserialization while ignoring it in the dataclass field and in to_dict() was internally inconsistent. This change makes Python deserialization symmetric with serialization and with the in-memory default.

Validation

• python -m pytest test_event_forward_compatibility.py -> 10/10 pass (including new regression tests).
• Broader Python suite: 136/136 non-E2E tests pass.
• npx vitest run on python-codegen.test.ts and other non-CLI test files: pass.
• ruff check, ruff format --check, ty check copilot, tsc --noEmit, eslint: clean.
• Manual reproduction of each issue's repro script on unmodified main (changes stashed) confirms the bug; with the fix applied, all three round-trip with equal? True.

Independent review by the code-review agent on claude-opus-4.7-xhigh found no significant issues and confirmed both codegen paths and the test coverage.