[Zheng-Lu]

Related to #335 (split from #350)

Summary

OpenAI marks assistant messages with phase (commentary or final_answer). The gateway's message types had no such field, so the value was dropped from the public response, from stored history, and from the input replayed on the next turn.

• Add an optional, closed MessagePhase to input and output messages. It passes through completion, output-to-input conversion, storage and rehydration unchanged. The gateway never infers a phase: absent or null stays absent, and vLLM's "phase": null output is unchanged.
• Reject phase on non-assistant input messages with a 400 before any inference.
• Move InputMessage and OutputMessage into types/io/message.rs to stay within the file-size policy. Existing import paths still work. Rust struct literals must now set phase; the constructors default it to None.

Test Plan

• cargo test -p agentic-server-core -p agentic-server: 1,728 passed, 0 failed, 17 ignored. cargo clippy --workspace --all-targets -- -D warnings, cargo fmt --check and pre-commit (including Rust file sizes) pass.
• message_phase_test.rs replays the recorder-generated custom-tool-openai-reference-gpt-5.6 cassettes (JSON and SSE):
  • phase: final_answer reaches the public output.
  • A previous_response_id continuation replays it upstream.
  • Negative control: on main both values are missing, in both modes.
• The same file covers the closed enum, absent and null legacy values, storage round trips, the upstream wire projection, and rejection for user, system and developer. openapi.rs checks the schema accepts absent, null and both values, and rejects unknown ones.
• No cassette was added or edited.

[maralbahari]

@Zheng-Lu thanks for the split. there seems to be overlap for the changes with #372 it adds support for phase and final_answer

[Zheng-Lu]

@Zheng-Lu thanks for the split. there seems to be overlap for the changes with #372 it adds support for phase and final_answer

Thanks, you're right. I'll close #383.

The only behavior #383 adds is rejecting phase on non-assistant input messages with a 400; main currently forwards it upstream. If you'd like that check, I can open a small follow-up with the recorded-cassette regression test; otherwise nothing is lost.