Model editor (/model)

Use /model to select a built-in provider preset or manage named custom model profiles. This document covers the editor workflow and saved fields. Wire-level adapter behavior lives in Providers.

Fast path: built-in preset

1. Run /model.
2. Choose a vendor, then a preset.
3. Open the preset editor, paste the API key, and save with Ctrl+S or the [ Save ] row.

The short preset form contains only Preset, Model, Endpoint, and API Key. Changing Model or Endpoint intentionally converts the configuration to Custom; the preset label never remains attached to modified preset-owned fields.

API keys are remembered per vendor. Switching among presets of the same vendor does not erase the key, and choosing a preset never copies credentials into a session journal.

Built-in presets

The shipped catalog is the source used by both the picker and runtime:

Preset	Model id	Endpoint	Context	Max output
DeepSeek V4.0 Flash	deepseek-v4-flash	https://api.deepseek.com	1M	384K
DeepSeek V4.0 Pro	deepseek-v4-pro	https://api.deepseek.com	1M	384K
DeepSeek V4.0 Flash Vision (Exp)	deepseek-v4-flash-vision-exp	https://api.deepseek.com	1M	384K
GLM 5.3	glm-5.3	https://open.bigmodel.cn/api/coding/paas/v4	1M	128K
GLM 5.3 Flash	glm-5.3-flash	https://open.bigmodel.cn/api/coding/paas/v4	1M	128K
Qwen3.8 Max	qwen3.8-max	https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1	1M	128K
Kimi K3	kimi-k3	https://api.kimi.com/coding/v1	1M	128K

All use the OpenAI-compatible protocol. The catalog also owns request paths, reasoning parameters, usage streaming, context-window seeds, and vendor-specific headers. For example, GLM preserves thinking, Qwen uses its low/medium/xhigh ladder, and the Kimi Coding endpoint requires a whitelisted coding-agent User-Agent. See Providers before overriding those fields.

GLM 5.3 Flash is currently the only built-in preset with probe-verified native image input, and therefore the only picker route that enables the image attachments described in Using CLAT. The DeepSeek Vision experimental preset and all other presets remain text-only in CLAT until their exact route has equivalent evidence. This does not restrict images handled wholly inside configured MCP tools.

Custom profiles

Custom models are named profiles. Each row keeps a complete model configuration and its own optional API key, so switching profiles does not destroy the inactive configuration.

• With no profiles, selecting Custom opens a blank profile template.
• With profiles, Custom opens a list showing name, endpoint · model, and a ● beside the active saved row.
• Enter activates a row, e edits, and d twice confirms deletion.
• New… opens the blank template.

Deleting the active profile activates the first remaining profile. If none remain, CLAT returns to factory custom defaults. Existing installations with a legacy single custom slot are adopted into the first named profile.

The required profile fields are:

• Name — unique display identity;
• Model — provider model id;
• Endpoint — base URL;
• API Key — optional, so local gateways can omit it.

Save activates the profile immediately. Esc cancels the current editor and returns to the exact picker row it came from; it does not collapse the entire dialog chain.

Editor controls

Press Enter, click, or type on a text field to open its input. Enter confirms the field, Esc cancels it. Ctrl+S saves the whole editor. On Max Output, Context Window, Temperature, Parallel Tool Calls, and the profile Thinking row, Ctrl+D toggles an explicit Clear tombstone. A cleared row says cleared (field omitted); this differs from an empty value, which means Inherit. Editing or cycling the row replaces the tombstone.

Numeric settings use bounded choice rows before a custom input:

Field	Choices
Context Window	128K (default template), 256K, 1M, Custom…
Max Output	8K, 32K (default template), 128K, Custom…
Spend Budget	1M, 10M (default), 50M, off, Custom…

Enter or ←/→ cycles a choice. Selecting Custom opens a numeric input.

Reasoning level

The profile Thinking row offers Low, High, Max, and Off. Off sends no inferred reasoning parameter and follows the endpoint's own behavior. Known vendor domains receive the vendor-native mapping; unknown domains retain the saved selection for display but do not receive an injected parameter because strict gateways may reject it. Configure unknown endpoints through Extra Body.

Preset defaults are pinned to a practical middle tier:

• DeepSeek, GLM, and Kimi store high;
• Qwen stores medium, which is CLAT's High tier;
• all presets keep a context-window value so automatic compaction works on the first long conversation.

Shift+Tab changes the active runtime configuration for the next run. For a named profile this deliberately does not rewrite the saved row: the ● marker drops until you switch away and reactivate the profile. To persist the change inside that profile, edit its Thinking row.

DeepSeek also supports a raw non-thinking mode, but the shortcut ladder stays Low → High → Max. GLM 5.3 rejects disabled thinking. Hand-written Extra Body is the escape hatch for any provider-specific mode not represented by the editor.

Advanced fields

[ Advanced ] exposes transport and request controls:

• Protocol — OpenAI Compatible (default) or OpenAI Responses;
• Request Path — /chat/completions or /responses by default;
• Auth Header and Auth Prefix — default Authorization and Bearer ;
• Extra Headers JSON — for example {"X-Tenant":"acme"};
• Extra Body JSON — for example {"top_p":0.9};
• Max Output Tokens;
• Context Window;
• Spend Budget;
• Temperature;
• Parallel Tool Calls.

The provider adapter protects managed fields such as model, messages, tools, stream, and instructions. Trying to override them through Extra Body is an error rather than an ambiguous merge.

Editing any preset-controlled field—Model, Endpoint, Protocol, Request Path, Extra Body, Max Output, Temperature, or Parallel Tool Calls—clears the preset identity. Editing Extra Body also clears the shortcut-managed thinking value so the raw body becomes the single source of truth.

Context and spend budgets

Context Window is the automatic compaction budget. Presets seed it with the catalog value; a user-entered value wins. CLAT estimates context before the next run, triggers around 80% of the window, and compacts toward the same ratio while preserving the original journal.

Spend Budget limits total input+output tokens for one run. The default is 10M and 0 disables it. Despite the label, its unit is tokens, not money. Every model or extension-sampling request reserves a conservative amount before it starts; provider usage reconciles the reservation when available. Malformed or absurd usage cannot wrap or silently reset the ledger. Crossing the cap stops the run with used/cap values and a pointer back to /model.

The spend value belongs to the model configuration, so different profiles can carry different risk/cost limits.

Provider replay data

Some providers require hidden state on later turns. The editor does not expose or persist that state as configuration:

• OpenAI Responses reasoning items travel as opaque provider replay state;
• DeepSeek reasoning_content is replayed only on assistant turns that made tool calls, which is the wire shape required for multi-turn tool use.

This data belongs to the session journal and provider adapter. Switching a profile changes future requests but does not rewrite historical events.