fix: accept responses as alias for codex_responses in api_mode config

[Kailigithub]

Problem

When api_mode: responses is set in config.yaml for a custom provider, the value is silently ignored. Hermes always uses /chat/completions regardless of the setting.

The root cause is that _parse_api_mode() in runtime_provider.py only accepts values from _VALID_API_MODES, which does not include "responses". The setup wizard (hermes setup option 3) correctly maps "responses" → "codex_responses", but anyone who writes api_mode: responses directly in config.yaml hits this gap.

Fix

Add a "responses" → "codex_responses" alias in _parse_api_mode() so the config file behaves identically to the setup wizard selection. This is a 1-line functional change (plus comments).

Testing

• python3 -m py_compile hermes_cli/runtime_provider.py — passes
• Unit verification: _parse_api_mode("responses") now returns "codex_responses"
• Existing provider tests (5 selected) pass without modification

Closes #33600

[alt-glitch]

Duplicate of #33604 — both normalize api_mode: responses → codex_responses in runtime_provider.py. #33604 also touches agent_init.py for URL routing.

[teknium1]

Thanks for the focused compatibility fix. Current main still lacks the responses alias in hermes_cli/runtime_provider.py:352-358, and named custom-provider resolution feeds entry values through that parser at hermes_cli/runtime_provider.py:695, :718, and :766.

Problems

• The PR adds no regression test. Please cover a real named providers: or custom_providers: entry with api_mode: responses and assert that resolve_runtime_provider() returns codex_responses.
• The description should not imply that every bare provider: custom endpoint will use /responses. Current main deliberately drops codex_responses for non-direct OpenAI/xAI bare custom relays in hermes_cli/runtime_provider.py:154-161; tests/hermes_cli/test_runtime_provider_resolution.py:1432-1447 protects that behavior.

Suggested changes

• Limit the documented scope to named custom-provider configuration and add the focused resolver regression above, without weakening the bare-custom relay guard.

Automated hermes-sweeper review.

[teknium1]

Please add a resolver-level regression for a named providers: or custom_providers: entry using api_mode: responses. In current main, bare non-OpenAI provider: custom routes deliberately discard the resulting codex_responses mode (_resolve_plain_custom_api_mode()), so a parser-only check would not verify the supported configuration path.

[Kailigithub]

Closing as superseded by #33604 (LeonSGP43), which makes the same responses → codex_responses alias fix in hermes_cli/runtime_provider.py AND additionally patches agent/agent_init.py for the same alias normalization plus adds 46 lines of regression coverage across two test files — a strict superset (53/2 lines vs this PR's 5/0). The triage bot's duplicate marker on this PR is correct on intent; #33604 was opened ~22 minutes before this PR and is strictly broader in scope (more files, more cases, full test coverage).

(Autopilot housekeeping takeover per v1.18 / v1.59 strict-supersede close protocol.)