Fix Codex named tool_choice serialization in reflect

[Sapientropic]

Summary

This fixes a Codex provider compatibility issue in reflect/tool-calling paths.

Reflect can force a specific tool on early iterations using the legacy named tool choice shape:

{"type":"function","function":{"name":"recall"}}

The current Codex Responses API expects the function name at the top level instead:

{"type":"function","name":"recall"}

Without that normalization, Codex rejects the request and reflect can end up with empty based_on.memories even when recallable evidence exists.

What changed

• normalize legacy named tool_choice dicts in CodexLLM.call_with_tools()
• add a regression test that verifies the serialized payload no longer contains tool_choice.function
• add a regression test that verifies forced tool choice still yields tool calls

Why this belongs in the provider

This follows the same design direction as the existing provider-specific tool-choice compatibility handling for LM Studio / OpenAI-compatible providers:

• the reflect agent keeps its provider-agnostic intent
• each provider adapts the request shape it needs

Validation

Targeted regression:

python -m pytest -o addopts="" --noconftest hindsight-api-slim/tests/test_codex_tool_choice.py -q

Result:

2 passed

I also verified the original failure mode before the patch via a live Codex-backed daemon, where reflect requests were rejected with:

400 Unknown parameter: 'tool_choice.function'

After this patch, the same integration recovered real reflect evidence instead of failing in the forced-tool iterations.

Non-goals

This PR does not claim to fix every possible reflect empty-evidence case.
It specifically fixes the Codex provider request-shape incompatibility for forced named tool choice.

[nicoloboschi]

Thansk, can you use pytest for the test? Just follow the codebase standard

[chriscoey]

Can confirm this bug. I'm using openai-codex provider with gpt-5.4 and reflect was returning "I don't have information" despite recall finding 12 results for the same query. The Responses API expects {"type": "function", "name": "recall"} but the agent sends {"type": "function", "function": {"name": "recall"}}. Applied the same fix locally and reflect works correctly now.