fix: allow Responses API auto-upgrade for custom providers serving GPT-5.x

[moxi-2534]

Summary

Fix the auto-upgrade guard in AIAgent.__init__ that prevents custom providers from switching to codex_responses mode for GPT-5.x models.

Problem

The condition api_mode is None at line ~1268 of run_agent.py is always False for custom providers, because runtime_provider.py always passes api_mode="chat_completions" explicitly. This means _provider_model_requires_responses_api() (which correctly detects GPT-5.x) never gets a chance to trigger the upgrade.

Users with custom OpenAI-compatible proxies (Codex Manager, one-api, etc.) serving GPT-5.x models get empty responses on every tool call.

Fix

-            api_mode is None
+            (api_mode is None or api_mode == "chat_completions")

Allows the default chat_completions to be auto-upgraded for GPT-5.x, while still respecting explicitly chosen non-default modes.

Fixes #23893

[teknium1]

Thanks for isolating the original api_mode handoff issue. Current main has since taken an intentionally different safety direction.

Problems

• run_agent.py:1397-1401 now explicitly keeps generic provider="custom" GPT-5 relays on Chat Completions because they may not implement full Responses semantics. Commit 0e4c879 added this policy and regression coverage for https://relay.example.com/v1.
• The proposed api_mode == "chat_completions" branch would also override an explicitly supplied Chat Completions mode, contrary to the contract documented in agent/agent_init.py:481-482.
• The target initialization code moved to agent/agent_init.py:486-505; this one-line patch no longer applies to current main.

Suggested changes

• Any follow-up should use a verified endpoint capability signal or a named-provider configuration path, while retaining the plain-custom-relay safeguard.
• Add coverage for both a Responses-capable proxy and a GPT-5 relay that only supports Chat Completions.

Automated hermes-sweeper review.

[teknium1]

Current main intentionally keeps generic provider="custom" GPT-5 relays on Chat Completions (run_agent.py:1397-1401, commit 0e4c879) because many OpenAI-compatible relays do not implement full Responses semantics. This broad condition would undo that regression fix and also override an explicit api_mode="chat_completions" choice.