feat: add per-message reasoning control

[jacksonriding]

Summary

• add a compact reasoning control to the primary chat composer for the next message
• resolve vendor-native reasoning parameters for OpenAI, Azure, Anthropic, Google, Bedrock, and configured custom/OpenRouter models such as Qwen
• normalize categorical reasoning effort and numeric thinking budgets into a consistent scale-based interaction
• keep overrides request-scoped while preserving them for regenerate, edit, continue, durable queue, and HITL resume flows
• validate provider/model compatibility server-side and prevent the override from changing saved conversation defaults

Composer experience

• replace the settings-style popover and long provider descriptions with a compact, focused control
• use a draggable effort scale for categorical settings such as OpenAI reasoning effort
• give token-budget providers the same scale treatment while retaining exact numeric entry and an explicit Auto state
• derive numeric scale floors, ceilings, and steps from the active model so Gemini and Claude budgets stay within their supported ranges
• show compact token values in the composer and at the ends of the scale
• contain focus and pointer interaction inside the open modal popover so selecting or dragging does not feel like an outside click
• use semantic surface, text, border, hover, and inverted roles throughout for light, dark, and custom-theme compatibility
• preserve keyboard control and expose human-readable slider values such as Medium, High, Auto, and formatted token counts to assistive technology

Provider mapping

• OpenAI and Azure reasoning models: reasoning_effort
• adaptive Claude models: effort
• manual-thinking Claude models: thinkingBudget
• Gemini 3 and newer: thinkingLevel
• Gemini 2.5: thinkingBudget
• Bedrock: provider-family setting
• custom and OpenRouter models: configured parameter definition

Test plan

• 814 focused tests passed across client, API, data-provider, HITL, and queue paths
• added focused composer tests for compact rendering, enum scales, numeric token scales, exact entry, Auto, focus containment, and accessible value text
• Mongo-backed durable queue suite passed
• client TypeScript check passed
• shared-client build passed after extending the Slider primitive with aria-valuetext
• semantic-token and contrast guardrails passed for shared light/dark theme roles
• data-provider, data-schemas, and API package builds passed
• staged ESLint, Prettier, import sorting, package validation, and circular-dependency checks passed
• production npm audit found no high or critical advisories
• manually started the modified frontend and API and verified both the UI and proxied config endpoint return HTTP 200

Closes #15242

[jacksonriding]

Mapped appropriately for all models depending on what is supported

[github-actions]

👋 Thanks for the contribution! LibreChat merges all changes into dev first — main only moves at release time — so this pull request's base branch was switched from main to dev automatically.

Nothing is needed from you; your commits, reviews and discussion are unchanged. If the diff now shows files you did not touch, rebase onto dev:

git remote add upstream https://github.com/danny-avila/LibreChat.git
git fetch upstream dev
git rebase upstream/dev
git push --force-with-lease

Maintainers: apply the target: main label and restore the base branch if this one genuinely belongs on main.

[berry-13]

Hey @jacksonriding, thanks for the thorough implementation. I compared #15406 with #14473 and confirmed substantial overlap. All 51 feature files changed here are represented in #14473. Your request-scoped reasoning design, provider-aware numeric budgets, and propagation through regenerate, queue, edit, continue, and HITL were stronger than the persistent-setting approach previously used in #14473. I integrated those parts into #14473 while preserving your authorship, adapting them to the unified Composer/Bar, and retaining its Bedrock, Azure, durable queue, and feature-owned state improvements.