feat: reasoning effort control in Web UI (/reasoning command + settings)

[nesquena-hermes]

Summary

The Web UI has no way to control reasoning effort. The CLI offers /reasoning [none|minimal|low|medium|high|xhigh] which sets the effort level for the current session, and /reasoning [show|hide] which toggles whether the thinking block is visible in the response. Neither control exists in the Web UI — users get whatever the model defaults to and cannot change it without editing config.yaml on the server and restarting.

Current state

In the CLI, /reasoning high calls _parse_reasoning_config("high") which produces {"enabled": true, "effort": "high"} and stores it in agent.reasoning_effort in config.yaml. On the next turn the agent is re-initialized with reasoning_config set, which flows into the model API call as extra_body["reasoning"] (for OpenRouter), thinking kwargs (for Anthropic direct), or reasoning.effort (for Codex/Responses API).

In the Web UI, _run_agent_streaming() (api/streaming.py) constructs the AIAgent with no reasoning_config argument at all. The streaming endpoint POST /api/chat/start also accepts no reasoning_effort field. The Web UI does correctly stream and render reasoning SSE events (thinking blocks), but it cannot set the effort level that generates them.

Proposed solution

Two surfaces:

1. Per-session slash command — /reasoning [level] and /reasoning [show|hide] in the Web UI command list (static/commands.js). These pass through to the agent as regular messages (same pattern as skill slash commands — issue #460). The agent already handles these correctly when sent as user messages. The dropdown should show the valid levels as autocomplete hints.

2. Persistent setting in the Settings panel — a "Reasoning effort" selector (none / minimal / low / medium / high / xhigh) that writes to /api/settings and is forwarded to the agent config. This lets users set a default level that persists across sessions, not just the current one.

Backend change required: _run_agent_streaming() needs to accept a reasoning_effort parameter and pass it as reasoning_config to AIAgent. The /api/chat/start POST body should accept reasoning_effort. The /api/settings POST handler should accept and persist reasoning_effort to the agent's config.yaml (not just the WebUI settings store).

Reasoning visibility — the existing thinking-block toggle in the UI already handles show/hide at render time, so /reasoning show|hide could just toggle that display setting directly.

Valid effort levels

From hermes_cli/commands.py: none, minimal, low, medium, high, xhigh

• none → {"enabled": false} — disables reasoning entirely
• minimal–xhigh → {"enabled": true, "effort": "<level>"}
• Empty / unset → model default (currently medium on OpenRouter)

Files involved

• ~/.hermes/hermes-agent/hermes_constants.py — parse_reasoning_effort()
• ~/hermes-webui-public/api/streaming.py — _run_agent_streaming()
• ~/hermes-webui-public/api/routes.py — /api/chat/start, /api/settings
• ~/hermes-webui-public/static/commands.js — add /reasoning to COMMANDS
• ~/hermes-webui-public/static/panels.js — Settings panel reasoning selector

[qxxaa]

Regarding the reasoning visibility aspect of this issue — there is currently no thinking-block show/hide toggle in the Web UI. The issue description references one ("the existing thinking-block toggle in the UI already handles show/hide at render time") but it does not exist in the codebase.

I have prototyped and tested a working implementation locally. It is a small change (+19 / -1 lines across 7 files) that follows the exact same pattern as the existing bubble_layout and show_token_usage toggles:

api/config.py: Add show_thinking: true to _SETTINGS_DEFAULTS and _SETTINGS_BOOL_KEYS
static/boot.js: Initialise window._showThinking from settings on page load
static/index.html: Add a settingsShowThinking checkbox in the settings panel
static/panels.js: Wire into loadSettingsPanel(), saveSettings(), and _applySavedSettingsUi()
static/ui.js (~line 1135): Gate the thinking card render with if(window._showThinking !== false)
static/messages.js: Gate _renderLiveThinking() so live streaming cards are also hidden (parsing continues regardless, so data is always preserved)
static/i18n.js: Add English label and description strings

Default is true (preserving current behaviour). The toggle only controls visibility - thinking data is always preserved in message history, so toggling it back on reveals all prior thinking blocks instantly.

This would also give /reasoning show|hide something to actually toggle, as mentioned in the proposed solution.

[nesquena-hermes]

@qxxaa — your implementation sounds exactly right, and the approach (following the bubble_layout / show_token_usage toggle pattern) is the correct one for a new settings bool.

Would you like to open a PR with what you've built? A +19/-1 across 7 files following an established pattern is a fast review — and if the toggle is wired correctly with window._showThinking and the settings persistence, it should be straightforward to merge.

If you open a PR, include a short manual verification note (e.g. toggle off → thinking blocks hidden in new and existing messages; toggle on → all revealed immediately without a reload). That'll help us confirm the visibility gate is working at the right point in the render pipeline.