Custom model session starts with a 200K window (99% full) until /model → Default switches it to [1m]

**Versions:** Claude Code v2.1.272, ollama 0.34.0, macOS 26

**Steps**
1. `cd && ollama launch claude`
2. In the launcher pick `glm-5.3:cloud`, listed as "Long-horizon coding and agentic engineering with deep reasoning and a 1M context" (real context: `ollama show glm-5.3:cloud` → `context length 1048576`).
3. `/context` → `glm-5.3:cloud · 197.2k/200k tokens (99%)`, `Auto-compact window: 200k tokens`, "Context is 99% full" on an empty session (messages: 69 tokens).
4. `/model` → the list shows `1. Default (currently glm-5.3:cloud[1m])`, `2. glm-5.3:cloud ✔ Custom Opus model`, `4. glm-5.3:cloud Custom Sonnet model`, `5. glm-5.3:cloud Custom Haiku model`.
5. Select `1. Default` → "Set model to glm-5.3:cloud[1m]".
6. `/context` → `glm-5.3:cloud[1m] · 215.2k/1m tokens (22%)`, `Auto-compact window: 1m tokens`.

**Expected:** the session starts with the 1M window, or the picker makes clear that the entries named `glm-5.3:cloud` give 200K while Default gives 1M.

**Actual:** the launched session runs on a 200K window and is at 99% before the first message. The same model gets 1M only after re-selecting Default in `/model`.

**Verbatim output**

Launcher — I pick `glm-5.3:cloud`, described as "a 1M context", and expect a 1M window, not 200K:
```
➜ cloudshot git:(main) ollama launch claude
Select model for Claude Code: Type to filter...

Recommended
glm-5.3-flash:cloud
Fast reasoning for coding and agentic workloads with 1M context and native image understanding
▸ glm-5.3:cloud
Long-horizon coding and agentic engineering with deep reasoning and a 1M context
deepseek-v4.1-flash:cloud
Fast coding and agentic tool use with 1M context and native image understanding
gemma4:31b-cloud
Agentic workflows and multimodal reasoning
gemma4:26b
Agentic workflows and multimodal reasoning, ~19GB, (not downloaded)
```

Right after `ollama launch claude` → `glm-5.3:cloud`:
```
❯ /context
⎿ Context Usage
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ glm-5.3:cloud
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ 197.2k/200k tokens (99%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ Estimated usage by category
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ System prompt: 2k tokens (1.0%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ System tools: 41k tokens (20.5%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ MCP tools: 142.4k tokens (71.2%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ Custom agents: 173 tokens (0.1%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ Memory files: 9.5k tokens (4.8%)
⛁ ⛁ ⛀ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛀ ⛁ Skills: 2.1k tokens (1.0%)
⛁ Messages: 69 tokens (0.0%)
⛝ Autocompact buffer: 33k tokens (16.5%)

Auto-compact window: 200k tokens
```

It switches to 1M **only** if I open `/model` and select `1. Default`:
```
Select model
Switch between Claude models. Your pick becomes the default for new sessions. For other/previous model names, specify with --model.

❯ 1. Default (recommended) Use the default model (currently glm-5.3:cloud[1m])
2. glm-5.3:cloud ✔ Custom Opus model
3. Fable Fable 5.1 · Most capable for your hardest and longest-running tasks · $10/$50 per Mtok
4. glm-5.3:cloud Custom Sonnet model
5. glm-5.3:cloud Custom Haiku model

◐ Medium effort ←/→ to adjust

Enter to set as default · s to use this session only · Esc to cancel
```
After that:

```
❯ /model
⎿ Set model to glm-5.3:cloud[1m] (default) and saved as your default for new sessions

❯ /context
⎿ Context Usage
⛀ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ glm-5.3:cloud[1m]
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛀ ⛁ ⛁ ⛁ 215.2k/1m tokens (22%)
⛀ ⛁ ⛁ ⛁ ⛀ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ Estimated usage by category
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System prompt: 2k tokens (0.2%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System tools: 35.2k tokens (3.5%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ MCP tools: 142.4k tokens (14.2%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Custom agents: 173 tokens (0.0%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Memory files: 9.5k tokens (1.0%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛁ Skills: 7.8k tokens (0.8%)
⛁ Messages: 18.1k tokens (1.8%)
⛶ Free space: 751.8k (75.2%)
⛝ Autocompact buffer: 33k tokens (3.3%)

Auto-compact window: 1m tokens
```

No source files or tests are named. Start by tracing model selection from `ollama launch claude`, then compare how that path initializes the model with the `/model` Default selection; done means a new session launched with the 1M-context model uses the 1M window, with the picker clearly distinguishing the alternatives.

Written by the indexing model from the issue text.