Fix agent-openai-agents-sdk multi-turn failure on 2nd+ prompt (id-less assistant history item)

[maskebail]

Summary

Apps built from the agent-openai-agents-sdk template fail on the second and later prompts with:

agents.exceptions.UserError: Unhandled item type or structure:
  {'status': None, 'content': [{'text': ..., 'type': 'output_text'}],
   'role': 'assistant', 'type': 'message'}

The first prompt always succeeds; every follow-up fails. Reproduces on fresh installs.

Root cause

openai-agents >= 0.19 tightened Converter.maybe_response_output_message in chatcmpl_converter.py to require {"id", "content"} on assistant history items:

if (isinstance(item, dict)
    and item.get("type") == "message"
    and item.get("role") == "assistant"
    and {"id", "content"} <= set(item)):   # now requires an id
    return cast(ResponseOutputMessageParam, item)
return None

Failure chain on the 2nd+ prompt:

1. The built-in chat UI replays the prior assistant turn as an id-less {"type":"message","role":"assistant","content":[output_text]} item.
2. MLflow's ResponsesAgentRequest strips any client-supplied id during normalization, so the item reaching the converter is guaranteed id-less.
3. The tightened guard no longer recognizes it; it matches no other branch and hits the catch-all raise UserError("Unhandled item type or structure").

The template pins openai-agents>=0.4.1 with no upper bound and no lockfile, so fresh installs resolve onto a build containing the change.

Fix

Collapse replayed id-less assistant output_text items to the easy-input {"role","content"} form before Runner.run, in both invoke_handler and stream_handler. That form is recognized by maybe_easy_input_message (no id required), so the fix is independent of the installed openai-agents version. Multiple output_text segments are joined with a newline to match the SDK converter byte-for-byte.

The helper lives in a new dependency-free agent_server/history.py so it is unit-testable in isolation.

Why not just pin openai-agents?

A version pin was considered and rejected — pinning <0.18 resolves to 0.17.8, whose older model layer is incompatible with the current openai token-usage schema and fails every prompt (including the 1st) with InputTokensDetails: cache_write_tokens Field required. The code fix above is the correct, version-independent solution.

Maintainers may also want to commit a uv.lock for reproducible installs (kept out of this PR so it's generated in your own resolver environment).

Validation

Verified on a live Databricks Apps deployment (openai-agents 0.19.1, guard active):

Scenario	Result
1st prompt	✅
2nd prompt, non-streaming	✅ agent correctly recalls prior turn
2nd prompt, streaming	✅ clean response.completed + [DONE]
2nd prompt, multi-segment assistant history	✅ newline separator preserved

New offline unit tests (tests/test_history_normalization.py, 4 cases, no network) lock in the transformation so the regression can't return silently.

Changes

• agent_server/history.py — new, dependency-free normalize_history_items()
• agent_server/agent.py — import + wire both handlers (+3/−2)
• tests/test_history_normalization.py — regression tests