[FEATURE] Discover and list models from custom base URL in `/model` picker

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When using ANTHROPIC_BASE_URL with a local model server that hosts multiple models simultaneously, the /model picker only shows a single custom model (set via ANTHROPIC_MODEL env var). There is no way to switch between multiple locally running models within an active session without exiting Claude Code and relaunching with a different ANTHROPIC_MODEL value.

I run multiple local LLMs concurrently behind a reverse proxy (e.g. one large reasoning model + one fast coding model). My server exposes a standard GET /v1/models endpoint that returns all currently active models — but Claude Code never queries it. The /model menu is limited to 5 hardcoded Anthropic models + exactly 1 custom model slot.

This means switching between local models requires restarting Claude Code, which breaks workflow and loses session context.

Proposed Solution

When ANTHROPIC_BASE_URL is set to a custom endpoint, Claude Code should query GET /v1/models (standard OpenAI-compatible endpoint) on that URL and list all returned models as selectable entries in the /model picker.

For example, if the server returns:

{
  "object": "list",
  "data": [
    { "id": "ornith-35b", "object": "model" },
    { "id": "qwopus-coder", "object": "model" }
  ]
}

Then /model should show:

1. Default (recommended)   Opus 5 (1M context)
2. Opus (1M context)       ...
3. Sonnet                  ...
4. Sonnet 5 (1M context)   ...
5. Haiku                   ...
6. ornith-35b              Custom model (local)
7. qwopus-coder            Custom model (local)

Selecting a different local model in /model would update the model field in subsequent API requests. No other protocol changes needed — the server already routes based on the model field in the JSON body.

Alternative Solutions

My current workaround is a .bashrc wrapper function (claude-with) that queries /v1/models before launch to find the specified model and sets ANTHROPIC_MODEL:

# Auto-detects the first active model:
claude-local() {
  export ANTHROPIC_BASE_URL=http://localhost:9080
  export ANTHROPIC_AUTH_TOKEN=local
  export ANTHROPIC_API_KEY=""
  local live_model
  live_model=$(curl -s http://localhost:9080/v1/models 2>/dev/null \
    | grep -o '"id":"[^"]*"' | head -1 | cut -d'"' -f4 | sed 's/^claude-//')
  export ANTHROPIC_MODEL="${live_model:-unknown}[1m]"
  export ANTHROPIC_SMALL_FAST_MODEL="$ANTHROPIC_MODEL"
  echo "🚀 Claude Code → $live_model"
  claude "$@"
}

# Target a specific model by name:
claude-with() {
  local target="$1"; shift
  export ANTHROPIC_BASE_URL=http://localhost:9080
  export ANTHROPIC_AUTH_TOKEN=local
  export ANTHROPIC_API_KEY=""
  local live_model
  live_model=$(curl -s http://localhost:9080/v1/models 2>/dev/null \
    | grep -o '"id":"[^"]*"' | cut -d'"' -f4 | grep -i "$target" | head -1 | sed 's/^claude-//')
  if [ -z "$live_model" ]; then
    echo "⚠️  No SWAI model matching '$target'"
    return 1
  fi
  export ANTHROPIC_MODEL="${live_model}[1m]"
  export ANTHROPIC_SMALL_FAST_MODEL="$ANTHROPIC_MODEL"
  echo "🚀 Claude Code → $live_model"
  claude "$@"
}

This lets me launch with claude-with qwopus or claude-with ornith, but it still only provides one custom model slot in the /model menu during a session. To switch, I have to exit and relaunch.

Priority

Medium - Would be very helpful

Feature Category

Configuration and settings

Use Case Example

1. I start two local LLMs on my machine: ornith-35b (a 35B general reasoning model) and qwopus-coder (a coding-specialized model with speculative decoding).
2. Both run concurrently, each on its own port, behind a reverse proxy on localhost:9080.
3. I launch Claude Code with claude-with ornith for architecture planning.
4. Once the plan is ready, I want to switch to qwopus-coder to generate code faster.
5. Today: I must exit Claude Code (/exit) and relaunch with claude-with qwopus, losing my conversational session context.
6. Desired: I simply type /model, choose qwopus-coder from the list of running models, and continue in the same session seamlessly.

Additional Context

• Industry standard pattern: Tools like Cursor, Continue.dev, Aider, and OpenCode already query GET /v1/models from the configured base URL and dynamically populate their model picker.
• Server is ready: The reverse proxy already implements GET /v1/models returning all active models, and dispatches requests to the right backend based on the "model" field in the JSON request body.
• Minimal client change: Only requires querying GET /v1/models on ANTHROPIC_BASE_URL and adding the returned model IDs to the /model interactive list.
• Environment: Claude Code v2.1.231, Linux (Fedora), local models via llama.cpp / llama-server.