Models, thinking and speed

Each session has a model and up to four settings that shape how it runs: thinking level (effort), thinking mode, context window and speed. Which of these you see depends on the agent and on the model you picked. Weaver never shows a control the agent did not declare.

Change them for one session

1. Click the settings chip in the composer. It shows the current model and thinking level, for example Opus 5 · high.
2. Pick a row: Agent, Model, Thinking or Effort, Context, Speed.
3. Pick a value. The change applies to the session’s next turns.

You can also type /models (or /model) in the composer to open the model list.

The rows appear only when they apply:

• Model appears when the agent lets you switch models.
• Thinking or Effort appears when the selected model lists levels. When the model also has a separate thinking mode, the level row is called Effort and the mode row is called Thinking.
• Context appears when the model offers more than one context window.
• Speed appears when the model offers speed tiers. A faster tier that costs more usage is marked in the chip.

The lists come from the agent. Levels come from the selected model’s catalog entry, or from the agent’s declaration when the model gives none. Speed tiers come from the selected model’s catalog. When a model’s catalog allows only certain combinations, the picker greys out values that do not fit your other choices, and the daemon rejects combinations the model does not allow.

What each agent offers

	Claude Code	Codex	Pi	Cursor
Models	From Claude Code’s model list	From the Codex app-server catalog	Models from every Pi provider you have connected, grouped by provider	From Cursor’s catalog for your API key
Thinking levels	low, medium, high, xhigh, max, narrowed per model	The reasoning efforts each model lists	Up to off, minimal, low, medium, high, xhigh, per model	The levels each model lists
Thinking mode	No	No	No	Per model
Context window	No	No	No	Per model
Speed	Standard or Fast on models that support fast mode	The service tiers each model lists; the ordinary tier shows as Standard	No	The fast variants each model lists

Details that matter:

• Claude Code defaults. The default effort is the effortLevel in your Claude Code settings when the model supports it, otherwise high (or the model’s first listed level if it has no high). The default speed follows Claude Code’s fastMode setting. Fast gives faster responses and uses more of your plan.
• Claude Code’s default model. The “Default” entry is Claude Code’s recommended alias. Weaver shows which concrete model it resolves to.
• Codex speed. Codex keeps a chosen tier for later turns. Pick Standard to go back to the ordinary tier.
• Pi. Pi offers fast variants as separate models, so it has no speed control. If the model list is empty, connect a provider in Settings → Agents → Pi → Connections.
• Cursor. The model list needs a valid API key. A Free-plan Cursor key gets Cursor’s own plan-required error when the list refreshes.

Set defaults

New sessions start from defaults. You can change any of them per session afterwards.

• Per agent. Settings → Agents, pick the agent’s tab, then Session defaults. “Catalog default” and “Model default” mean you have not overridden the agent.
• Per project. Settings → Projects & New Tasks, pick a project. Its values override the agent defaults for new sessions in that project. “Inherit” means it uses the agent default.

If a default names a model the agent no longer lists, Settings warns you.

Legacy models

Claude Code lists one model per alias, so an older version drops out of the list when the alias moves on. Weaver keeps a short list of older Claude models the CLI still runs: Opus 4.8, 4.7 and 4.6 with 1M context, Opus 4.5, Sonnet 4.6 and Sonnet 4.5.

Turn one on in Settings → Agents → Claude → Legacy models. It then appears in every model picker. Each legacy model keeps its own thinking levels and fast-mode support. Opus 4.5 and Sonnet 4.5 have no effort control.

Summaries and session titles

Weaver runs small background turns to summarize diffs and title new sessions. These use their own model setting, separate from your chat. Configure them in Settings → Agents → Summaries & titles. You pick an agent and model there, and can add other agents as fallbacks.

Context and usage

• Every agent on today reports token counts per turn.
• Claude Code, Codex and Pi also report how full the context window is. Cursor does not report its window size, so Weaver shows no context percentage for Cursor.
• Account quota for Claude Code (marked Beta) and Codex shows in Settings → Usage. Codex users can redeem a ChatGPT rate-limit reset credit there.