feat(web): pick Codex and Claude effort with an effort dial

[incognitojam]

Note

For Codex and Claude, the composer's model pill now opens an effort dial: a slider over four levels, Light, Standard, Deep and Ultra. Each level picks a model and reasoning effort from a fixed table, across both providers when both are set up. The separate effort menu goes away for those providers.

Problem

The effort menu listed every reasoning level each model offers, so it was easy to run Opus 5.5 at xhigh or max, or Fable when a cheaper model would do. Benchmarks don't back those choices. On FrontierCode, Opus 5.5 at medium matches max at about an eighth of the cost. Choosing well also meant knowing which provider and model suit each kind of task, and for users with both subscriptions that choice decides whose usage a thread spends.

Change

Levels. packages/shared/src/effortDial.ts resolves each level to a provider, model and effort:

Level	Both providers, new thread	Claude only, or a Claude thread	Codex only, or a Codex thread
Light	GPT-6.1-Sol, medium	Sonnet 5.5, high	GPT-6.1-Sol, low
Standard	Opus 5.5, medium	Opus 5.5, medium	GPT-6.1-Sol, medium
Deep	Opus 5.5, high	Opus 5.5, high	GPT-6.1-Sol, xhigh
Ultra	Fable 5.1, high	Fable 5.1, high	GPT-6-Astra, xhigh

• Each cell is an ordered list, so a CLI without the newest model falls back. With Codex before 0.159, for example, GPT-6.1-Sol levels move to GPT-6-Sol or GPT-6-Astra.
• A model chosen from the model list that the tables don't cover stays selected; the levels change only its effort (low, medium, high, xhigh, or the nearest it offers).
• The choices follow benchmarks run through the real Codex and Claude Code harnesses where available.

Composer. This follows the Codex app's picker:

• The pill shows the model and level, for example "Claude Opus 5.5 Deep", or the raw effort when the selection matches no level. Its tooltip names the model and its effort.
• The popover follows the ChatGPT picker: a model row showing the effort on top, then the level names over a chunky slider, then a speed toggle. Clicking a name also selects its level. The thumb and fill animate between levels with a 150ms transform transition, which is off when reduced motion is on.
• The model row opens a short list: Default, then the current built-in Codex and Claude models (no legacy or custom models, and only the thread's provider once it has started), then All models, which opens the existing full picker.
  • Choosing a model switches the slider to that model's own efforts (Low to Max or Ultracode), and the pill shows the raw effort. This is how to run, say, Opus at xhigh.
  • Choosing Default returns to the levels. If the current selection matches no level, it moves to Standard.
  • The mode is a client setting (effortControl, levels or custom).
• Speed is Claude's fast mode or Codex's service tier, which previously lived in the effort menu. Its usual setting is labelled "Normal" so it doesn't read as the Standard level. A cost icon on Fast shows the provider's description on hover:
  • Codex: from its catalog, for example "2x speed, increased usage".
  • Claude: the manifest's fast mode option now carries a description taken from Claude Code's own wording: "Same model, faster output. Billed to usage credits rather than your plan's included usage."
  • Neither provider gives a numeric price ratio.
• Keyboard: the model picker keybinding and /model open the full model list directly.
• Context window: each model's default applies (1M for Opus 5.5 and Fable 5.1, 200k for Sonnet 5.5). A thread that already chose a size keeps it.

Started threads. Levels stay on the thread's provider. Stops that would switch its model have a dashed ring. Choosing one says the next turn re-reads the whole conversation. Effort-only changes apply without a restart (Claude since #541).

Unchanged.

• Other providers keep their effort menu.
• The settings panels (project default model, text generation model) keep the full pickers.
• The server is untouched: a level writes an ordinary model selection.

Docs: a new "Effort levels" section in the composer guide, a bullet in the styal differences page, and the effort-dial fork ledger entry.

Not in this PR:

• Mobile: its thread settings sheet still shows the raw effort choices.
• Usage notes and pacing: for example Fable's weekly allowance on Ultra.
• Keybindings to step between levels.

Evidence

The screenshots come from a dev server with isolated state and real Codex 0.159 and Claude Code providers. Thread names come from earlier testing.

Before: the effort menu	After: new thread at Light

Started Claude thread, light theme	Narrow footer

Dial at Standard	Model list

A chosen model's own efforts	Fast cost note

Validation

• New thread, both providers ready (browser automation against the dev server): the slider moved the pill through Light (GPT-6.1-Sol medium), Standard (Opus 5.5 medium), Deep (Opus 5.5 high) and Ultra (Fable 5.1 high). The speed toggle appears only for models that offer it. The model row opens the model list, and reopening the pill returns to the dial. No console or page errors.
• Started Claude thread on Sonnet 5.5 at low: the pill showed "Low" with no level selected. Every level stayed on Claude, and Standard, Deep and Ultra carried the model-switch note.
• Selection matching no level: choosing Light worked. It initially didn't, because the hidden input already rested on that stop; the panel now also applies stops on click.
• Narrow footer and light theme: checked in screenshots.
• Animation: sampled the thumb's transform every ~30ms after moving from Standard to Deep. It eased from 81px to 163px and settled within about 160ms.
• Tooltip: hovering the model name shows "Claude Opus 5.5, high effort".
• Model list: the list showed Default plus GPT-6.1-Sol, GPT-6-Astra, GPT-6-Luna, Opus 5.5, Fable 5.1 and Sonnet 5.5, with no legacy models.
  • Choosing GPT-6.1-Sol showed "GPT-6.1-Sol Medium" and a slider over its own efforts. Its top effort, Ultra, gave "GPT-6.1-Sol Ultra".
  • The list then checked GPT-6.1-Sol, not Default.
  • All models opened the full picker.
  • Choosing Default returned to "Claude Opus 5.5 Standard".
  • An earlier build confirmed that the mode survives a reload.
• Fast cost note: hovering the icon showed Codex's catalog text for GPT-6.1-Sol and the manifest text for Opus 5.5. The Claude text shows only with this branch's manifest; the dev check ran with provider update checks off, because the server otherwise prefers the published manifest from main.
• Tests: packages/shared/src/effortDial.test.ts covers the tables, provider locking, fallbacks, models off the tables, and instance choice. apps/web/src/components/chat/effortDial.logic.test.ts covers which models the dial offers and the speed control. The existing ProviderModelPicker and source-control writing settings suites pass.
• Checks: web typecheck, lint (no new warnings), format, and ledger check.
• Not verified: mobile, which is unchanged, and the desktop shell, which uses the same web code.

---

Written by an agent (Claude Code, claude-opus-5-5).

[github-actions]

Web preview

https://styal-web-pr-551.cameron-clough.workers.dev (for de952a6)

Open this exact URL; the hosted-app origin is baked in at build time. Pair a server into it with styal pair --tailscale, or paste a host and pairing code under Settings → Connections.

The preview is deleted when this PR closes or the preview label is removed.