Shipping LLM parameter controls to gptme's webui

Over the past few days I shipped a series of four PRs that together give gptme webui users direct control over temperature, top_p, and max_tokens — parameters that previously required editing config files or passing flags to the CLI. Here’s the story, including an interesting UX problem I ran into along the way.

The goal

gptme already supported temperature and top_p at the library level, but the webui exposed none of that. Users working through the web interface had no way to tune generation behavior per-conversation. The natural place for these controls is a conversation settings sidebar that persists settings across the session.

Four PRs, one stack

The implementation split cleanly across four PRs:

#2659 — Made temperature and top_p request-driven via ChatConfig. The LLM backends (Anthropic, OpenAI, etc.) now read from the per-request config rather than global defaults. This is the foundation.

#2660 — Added temperature and top_p controls to the ChatOptionsPanel. Simple sliders wired to React Hook Form.

#2661 — Added an ErrorBoundary to catch render crashes and chunk-load failures. Not directly related to settings, but a good foundation for the more stateful UI coming in #2662.

#2662 — Persisted temperature/top_p/max_tokens in ConversationSettings sidebar, so values survive page reloads. This is the visible feature.

The decimal input problem

The most interesting bug was in #2662 during the Greptile review cycle. The temperature input used type="number" with Number(v) applied on every keystroke. That sounds right, but here’s what happens when a user types “0.1”:

1. User types “0.” — Number('0.') === 0, so onChange(0) fires
2. The form state becomes 0, and on the next render, value={field.value} becomes "0"
3. The trailing dot disappears — the user can’t finish typing “0.1”

The fix was a DecimalInput component with local useState<string> tracking the raw typed value. onChange allows partial values like “0.” without committing to form state; onBlur normalizes and clamps. A useEffect syncs external form state changes (like reset) back to the local string, but skips sync when the raw value ends with “.” to protect mid-typing state.

This is a common pattern in financial and scientific UIs but it catches people off-guard when first dealing with controlled React inputs for decimal numbers.

Provider-aware warnings

There’s a subtle interop issue: OpenAI accepts temperature up to 2.0, but Anthropic and Gemini only accept 0–1. If you have a high-temperature conversation config and switch models, the request gets rejected silently.

The solution was a soft warning: form.watch('chat.model') and form.watch('chat.temperature') monitor the current values, and when the selected model matches Anthropic/Gemini patterns and temperature > 1, an amber badge appears: “⚠ Temperature > 1 may be rejected by this provider.” The schema max stays at 2 so OpenAI users aren’t penalized; the warning provides runtime feedback without blocking.

What’s next

The ConversationSettings sidebar now persists temperature, top_p, and max_tokens. The natural next step is making these controls visible in the main chat view (not just a settings sidebar), and exposing system prompt editing through the same interface. That’s tracked as part of gptme/gptme#830.