feat(openai): support Responses text verbosity

[JiehoonKwak]

Summary

• add agent.text_verbosity config parsing for OpenAI Responses output verbosity
• automatically inject text.verbosity only for GPT-5 models on Codex OAuth or the exact api.openai.com host; xAI, GitHub Models, and arbitrary custom endpoints fail closed
• merge configured verbosity into request_overrides.text without dropping text.format or replacing an explicit verbosity override
• validate text in the Responses preflight, document the option, and invalidate cached gateway agents when it changes

Context

Closes #20203.

#20258 also proposes this feature. This PR keeps the implementation at the existing agent-to-Responses transport boundary and does not add a generic capability framework or custom-proxy opt-in.

Tests

• scripts/run_tests.sh tests/agent/transports/test_codex_transport.py tests/run_agent/test_run_agent.py tests/gateway/test_agent_cache.py (581 passed)
• ruff check on all changed Python files
• python -m compileall on all changed Python files
• git diff --check

[teknium1]

Thanks for the focused config-to-transport implementation; current main still rejects top-level text, so the feature request remains valid.

Problems

• agent/transports/codex.py:152 sets configured text.verbosity before request_overrides is merged. The existing kwargs.update(request_overrides) immediately afterward replaces the entire text object, dropping verbosity when an override supplies text.format or another text field. Merge the dictionaries after applying overrides instead.
• agent/transports/codex.py:147 excludes only GitHub and xAI. Current named custom-provider resolution can select codex_responses (hermes_cli/runtime_provider.py:695-697), so this does not establish that the route accepts OpenAI's text field. The issue specifically asks not to send it to unsupported providers.

Suggested changes

• Merge configured verbosity into an existing request_overrides.text dict and test preservation of text.format.
• Gate injection on an explicit supported OpenAI GPT/Responses capability, with a custom non-OpenAI Responses regression test.

Automated hermes-sweeper review.

[teknium1]

request_overrides is merged immediately after this block, so an override containing text replaces this entire dict and silently drops configured verbosity. Apply overrides first, then copy/merge the existing text dict so fields such as text.format survive alongside text.verbosity.

[teknium1]

This is an exclusion check rather than a capability check. Named custom providers can opt into codex_responses on current main (hermes_cli/runtime_provider.py:695-697), so non-xAI/non-GitHub routes can still receive this OpenAI-only field. Please gate on a known supported OpenAI GPT/Responses route or capability.

[JiehoonKwak]

Updated the PR on current main in 44d3b1b1b and addressed both review points:

• request_overrides is applied first, then configured verbosity is nested into text with setdefault. This preserves text.format, keeps an explicit override authoritative, and does not mutate the caller's override dict.
• Capability is resolved at the agent request boundary from the resolved model and endpoint: GPT-5 family plus either Codex OAuth or exact hostname api.openai.com. The transport receives only that boolean. GPT-5 on a custom endpoint, xAI, GitHub Models, and non-GPT-5 models all fail closed.

The regression coverage includes the nested merge, explicit override precedence, the endpoint/model matrix, and an AIAgent configured as a custom codex_responses provider. Focused suite: 581 passed; changed-file Ruff, compileall, and git diff --check also pass.