fix(reasoning): expose max effort for GPT-5.6

[PGAAUmN3yEL0O8g6Yn9J553D3Q8p6g7S7GW1i3U]

Summary

• expose the existing max reasoning effort in the gateway /reasoning command
• add max to CLI help, validation output, completion, and the interactive reasoning picker
• document the level in the gateway/config comments
• add regression coverage for gateway parsing, CLI parsing, menu ordering, and Codex transport passthrough

Testing

• uv run --extra dev pytest tests/test_hermes_constants.py tests/cli/test_reasoning_command.py tests/gateway/test_reasoning_command.py tests/hermes_cli/test_reasoning_effort_menu.py tests/run_agent/test_run_agent_codex_responses.py -q -o addopts=
• 245 passed

Fixes #61634

[alt-glitch]

This was generated by AI during triage.

Duplicate of #61638 (earlier open PR, created 2026-07-09, also Fixes #61634) — both add max to the /reasoning validators at the same code sites (gateway/slash_commands.py, hermes_cli/cli_commands_mixin.py, hermes_cli/commands.py). #61638 is earlier and refactors to reference VALID_REASONING_EFFORTS (plus locale/doc updates), so it's the canonical version. Marking duplicate; maintainer may prefer either. Related to the broader max-effort validator-gap cluster: #34199 / #25401 / #29248 / #49580.

[teknium1]

Thanks for addressing a real synchronization gap: current main accepts max in hermes_constants.py:794-819 and passes it through the Codex Responses transport at agent/transports/codex.py:155-166, while the gateway whitelist at gateway/slash_commands.py:2732 rejects it.

Problems

• The proposed gateway change continues a duplicated vocabulary. Please use VALID_REASONING_EFFORTS rather than another literal set; the current split is the defect being fixed.
• The change is incomplete across current UI surfaces. web/src/lib/reasoning-effort.ts:17-35 omits max and normalizes it to medium; desktop settings omit it at apps/desktop/src/app/settings/model-settings.tsx:85. The user guide also says max is not selectable at website/docs/user-guide/configuration.md:1284-1292.

Suggested changes

• Reuse the shared effort tuple in the gateway validator, then cover dashboard/desktop persistence or explicitly scope and document the supported surfaces.

Automated hermes-sweeper review.

[teknium1]

Please use hermes_constants.VALID_REASONING_EFFORTS here instead of extending this local set. max is already canonical at hermes_constants.py:794; retaining a second vocabulary recreates the drift this PR fixes.

[teknium1]

Closing — fully superseded. GPT-5.6 max effort acceptance landed in #62650, and the gateway/CLI exposure plus Codex Responses wire coverage this PR added is now on main via #65651 (salvage of #64012). Credit also to @SamuelZ12 (#12211) as the earliest max proposal. Thanks for the contribution!