Feature: Show agentModels from settings.json in /model picker instead of hardcoded GPT/Codex models

[suenot]

Problem

When using OpenClaude with a custom OpenAI-compatible provider (e.g. Kimi, GLM, DeepSeek via OpenRouter), the /model picker shows hardcoded GPT-5.x / Codex models that are irrelevant to the configured provider. Meanwhile, models defined in agentModels within ~/.openclaude/settings.json do not appear in the picker at all.

Current behavior

1. Set up a provider profile with Kimi (https://api.kimi.com/coding/, model kimi-k2.6)
2. Define additional models in settings.json under agentModels (GLM-5.1, DeepSeek V4 Flash, MiniMax)
3. Open /model picker

Result: The picker shows gpt-5.5, gpt-5.4, gpt-5.3-codex, gpt-5.3-codex-spark, gpt-5.2-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, gpt-5.5-mini, gpt-5.4-mini, plus Sonnet/Opus/Haiku — none of which are available through the configured provider. The custom agentModels (glm-5.1, deepseek-v4-flash, etc.) are nowhere in the list.

Expected behavior

• Models from agentModels in settings.json should appear in the /model picker so the user can switch between them for the main session (not just subagents)
• Hardcoded GPT/Codex models should not be injected when getAPIProvider() === "openai" and the base URL is clearly not OpenAI/Codex (e.g. api.kimi.com, api.z.ai, openrouter.ai)

Root cause (source analysis)

In src/utils/model/modelOptions.ts, lines 550-551:

if (getAPIProvider() === 'openai' || getAPIProvider() === 'codex') {
    payg3pOptions.push(...getCodexModelOptions())
}

This unconditionally injects 10 GPT/Codex model options whenever the provider is openai, regardless of the actual base URL or configured models.

Meanwhile, agentModels from settings.json are only consumed by resolveAgentProvider() in src/services/api/agentRouting.ts for subagent routing — they are never surfaced in the model picker (getModelOptionsBase()).

Proposed solution

1. Skip Codex injection when OPENAI_BASE_URL does not point to an OpenAI/Codex endpoint
2. Surface agentModels from settings.json in the /model picker, so users can switch the main session model to any of their configured providers
3. Alternatively, allow providerProfiles to define multiple profiles with different base URLs, and let the user switch between them from /model (currently only the active profile's models appear)

Use case

I use a tiered multi-provider setup:

• GLM-5.1 (Z.AI) — for planning, 10-20 tps
• Kimi K2.6 (Moonshot) — for routine coding, 20-60 tps
• DeepSeek V4 Flash (OpenRouter) — for exploration, high throughput
• MiniMax (SambaNova) — for simple tasks, 200-400 tps

All are defined in agentModels + agentRouting in settings.json. The routing works for subagents, but I cannot switch the main session model to any of these without manually editing env vars and restarting.

Environment

• OpenClaude v0.10.0
• macOS
• Provider: Moonshot AI (Kimi Code) via OpenAI-compatible profile

[suenot]

Configuration details

~/.openclaude/settings.json

This is where agentModels and agentRouting are defined — they work for subagent routing but are invisible in /model picker:

{
  "permissions": {
    "allow": [
      "Bash(find:*)",
      "Bash(ls:*)",
      "Bash(cat:*)"
    ]
  },
  "agentModels": {
    "glm-5.1": {
      "base_url": "https://api.z.ai/api/anthropic",
      "api_key": "your-zai-key"
    },
    "kimi-k2.6": {
      "base_url": "https://api.kimi.com/coding/",
      "api_key": "your-kimi-key"
    },
    "deepseek-v4-flash": {
      "base_url": "https://openrouter.ai/api/v1",
      "api_key": "your-openrouter-key"
    },
    "MiniMax-M2.5-sambanova": {
      "base_url": "https://api.sambanova.ai",
      "api_key": "your-sambanova-key"
    }
  },
  "agentRouting": {
    "Plan": "glm-5.1",
    "code-review": "glm-5.1",
    "general-purpose": "kimi-k2.6",
    "frontend-dev": "kimi-k2.6",
    "Explore": "deepseek-v4-flash",
    "default": "kimi-k2.6"
  }
}

~/.openclaude.json (relevant section)

The main provider profile for the session — this model (kimi-k2.6) correctly appears in the picker, but the other agentModels do not:

{
  "providerProfiles": [
    {
      "id": "provider_kimi_k26",
      "name": "Kimi",
      "provider": "openai",
      "baseUrl": "https://api.kimi.com/coding/",
      "model": "kimi-k2.6"
    }
  ],
  "activeProviderProfileId": "provider_kimi_k26"
}

~/.openclaude/.openclaude-profile.json

Startup env profile persisted by /provider:

{
  "profile": "openai",
  "env": {
    "OPENAI_BASE_URL": "https://api.kimi.com/coding/",
    "OPENAI_MODEL": "kimi-k2.6",
    "OPENAI_API_KEY": "your-kimi-key"
  }
}

What happens

The /model picker shows:

 1. Default (recommended)     Use the default model (currently kimi-k2.6)
 2. gpt-5.5                   GPT-5.5 with high reasoning
 3. gpt-5.4                   GPT-5.4 with high reasoning
 4. gpt-5.3-codex             GPT-5.3 Codex with high reasoning
 ...
12. Sonnet                    Sonnet 4.6
13. Sonnet (1M context)       Sonnet 4.6 for long sessions
14. Opus 4.1                  Opus 4.1 · Legacy
15. Haiku                     Haiku 3.5 for simple tasks
16. kimi-k2.6 ✓               Provider: Kimi

Missing from the list: glm-5.1, deepseek-v4-flash, MiniMax-M2.5-sambanova — all defined in agentModels in settings.json.

Should not be in the list: GPT-5.x/Codex models (items 2-11) — injected by getCodexModelOptions() because getAPIProvider() === "openai", even though OPENAI_BASE_URL points to api.kimi.com, not OpenAI.

Desired behavior

The picker should show:

 1. Default (recommended)     Use the default model (currently kimi-k2.6)
 2. kimi-k2.6 ✓               Provider: Kimi
 3. glm-5.1                   Provider: Z.AI (from agentModels)
 4. deepseek-v4-flash         Provider: OpenRouter (from agentModels)
 5. MiniMax-M2.5-sambanova    Provider: SambaNova (from agentModels)

Selecting one of these should switch the main session to use that model's base_url and api_key from the agentModels config, not just route subagents.

[suenot]

Update: correct multi-profile config

I realized the proper approach is to create separate provider profiles for each model (since they have different baseUrl). Updated ~/.openclaude.json:

{
  "providerProfiles": [
    {
      "id": "provider_kimi_k26",
      "name": "Kimi",
      "provider": "openai",
      "baseUrl": "https://api.kimi.com/coding/",
      "model": "kimi-k2.6",
      "apiKey": "your-kimi-key"
    },
    {
      "id": "provider_glm_51",
      "name": "GLM-5.1 (Z.AI)",
      "provider": "openai",
      "baseUrl": "https://api.z.ai/api/anthropic",
      "model": "glm-5.1",
      "apiKey": "your-zai-key"
    },
    {
      "id": "provider_deepseek_v4",
      "name": "DeepSeek V4 Flash (OpenRouter)",
      "provider": "openai",
      "baseUrl": "https://openrouter.ai/api/v1",
      "model": "deepseek/deepseek-v4-flash:nitro",
      "apiKey": "your-openrouter-key"
    },
    {
      "id": "provider_minimax_samba",
      "name": "MiniMax (SambaNova)",
      "provider": "openai",
      "baseUrl": "https://api.sambanova.ai",
      "model": "MiniMax-M2.5",
      "apiKey": "your-sambanova-key"
    }
  ],
  "activeProviderProfileId": "provider_kimi_k26"
}

Now all 4 profiles exist, but the /model picker still only shows models from the active profile. Switching between providers requires /provider, not /model.

The core feature request remains: the /model picker should surface models from all configured provider profiles (or at least from agentModels in settings.json), so the user can switch between providers and models from a single menu without restarting or using /provider.

[suenot]

Test results (v0.10.0)

Confirmed behavior after setting up 4 provider profiles in ~/.openclaude.json:

/provider — ✅ all 4 profiles visible

Edit provider
 1. Kimi (active)                  OpenAI-compatible API · https://api.kimi.com/coding · kimi-k2.6
 2. GLM-5.1 (Z.AI)                 OpenAI-compatible API · https://api.z.ai/api/anthropic · glm-5.1
 3. DeepSeek V4 Flash (OpenRouter)  OpenAI-compatible API · https://openrouter.ai/api/v1 · deepseek/deepseek-v4-flash:nitro
 4. MiniMax (SambaNova)             OpenAI-compatible API · https://api.sambanova.ai · MiniMax-M2.5

/model — ❌ only active profile model + hardcoded GPT/Sonnet/Opus

Select model
 1. Default (recommended)     Use the default model (currently kimi-k2.6)
 2. gpt-5.5                   GPT-5.5 with high reasoning
 3. gpt-5.4                   GPT-5.4 with high reasoning
 4. gpt-5.3-codex             GPT-5.3 Codex with high reasoning
 ...
12. Sonnet                    Sonnet 4.6
13. Sonnet (1M context)       Sonnet 4.6 for long sessions
14. Opus 4.1                  Opus 4.1 · Legacy
15. Haiku                     Haiku 3.5 for simple tasks
16. kimi-k2.6 ✓               Provider: Kimi

Missing: glm-5.1, deepseek/deepseek-v4-flash:nitro, MiniMax-M2.5 — all defined as provider profiles but not shown in /model.

Should not be there: GPT-5.x / Codex models (items 2-11) — the base URL is api.kimi.com, not OpenAI.

Summary

• Workaround: use /provider to switch between providers instead of /model
• Issue: switching providers via /provider requires multiple keystrokes and menu navigation. Would be much better to have all configured models in one /model list, or at minimum, skip the hardcoded GPT/Codex injection when OPENAI_BASE_URL is not an OpenAI endpoint.

[Vasanthdev2004]

Thanks for the detailed write-up. I think there are two separate pieces here:

1. Bug/polish: /model should not inject Codex/GPT options when the active OpenAI-compatible profile clearly points at a non-OpenAI endpoint like Kimi, Z.AI, OpenRouter, etc.
2. Larger UX feature: /model could become a unified switcher across configured provider profiles, not just models from the active profile.

The first piece feels like a good targeted fix. The second is useful, but it touches the boundary between /model and /provider, so it probably needs a small design pass to avoid confusing “change model only” with “change provider/base URL/API key”.

Keeping this open as a feature/UX issue. A clean first PR would be the narrower Codex/GPT option suppression for non-OpenAI endpoints, with tests around getModelOptionsBase().

[jatmn]

Yea this would be 2 independent issues/features

[0xfandom]

Would like to take the second piece — surfacing all configured providerProfiles (not just the active one) in the /model picker. PR #1146 already covers the Codex/GPT suppression piece, this would close the larger UX gap.

Design plan, kept narrow to match @Vasanthdev2004's ask for a /model vs /provider boundary that doesn't confuse "change model only" with "change provider/base URL/API key":

1. ModelOption gains an optional switchToProfileId?: string field. Existing options leave it unset and behave exactly as today.
2. getModelOptionsBase() appends a clearly-labeled section for inactive providerProfiles. Each entry is rendered as <profile.name> · <profile.model> so the user can see they're switching providers, not just models. Marked with switchToProfileId.
3. The /model select handler short-circuits when switchToProfileId is set: call setActiveProviderProfile(id), then applyProfileEnvToProcessEnv so subsequent requests use the new OPENAI_BASE_URL / OPENAI_API_KEY. Same code paths /provider already uses — no new env-management surface.
4. Tests around getModelOptionsBase() covering: active-only profile (today's behavior), 4-profile config from the repro, and the firstParty path (which should not surface inactive openai-compatible profiles).

Opening within a few hours.

[0xghost42]

Following up on the #1146 close — agree the URL-substring shape was wrong. For Piece 1 (suppress hardcoded GPT/Codex options when the active profile isn't actually OpenAI/Codex), here's a descriptor-backed approach that drops the URL sniff entirely.

In `src/utils/model/modelOptions.ts` around line 561, replace:

```ts
if (getAPIProvider() === 'openai' || getAPIProvider() === 'codex') {
payg3pOptions.push(...getCodexModelOptions())
}
```

with a check against the active profile's resolved `routeId`:

```ts
const profile = getActiveProviderProfile()
const routeId = profile ? resolveProfileRoute(profile.provider).routeId : null
const isActualOpenAIOrCodexRoute = routeId === 'openai' || routeId === 'codex'

if (isActualOpenAIOrCodexRoute) {
payg3pOptions.push(...getCodexModelOptions())
} else if (!profile && (getAPIProvider() === 'openai' || getAPIProvider() === 'codex')) {
// No profiles configured — preserve today's env-only behaviour for users
// who haven't migrated to provider profiles yet.
payg3pOptions.push(...getCodexModelOptions())
}
```

This drives off `PROVIDER_PRESET_MANIFEST` (the same source `/provider` already uses), so Kimi / Moonshot / OpenRouter / Z.AI / etc. — all `vendorId: 'openai'` but each with a distinct `routeId` (kimi-code, moonshot, openrouter, ...) — correctly skip the GPT/Codex catalog without any URL string matching, regex, or `OPENAI_API_BASE` fallback gap.

Three questions before I draft a PR:

1. @jatmn — does the descriptor/route approach match the redesign you flagged in fix(model picker): skip Codex/GPT options when OPENAI_BASE_URL is non-OpenAI (#1119) #1146? Or should Piece 1 hold until Add full access mode and fix bypass commit prompts (Issue 1097) #1110's permission-mode work lands first to avoid stacking conflicts?
2. Should the new behaviour gate on `CLAUDE_CODE_PROVIDER_PROFILE_ENV_APPLIED === '1'` like feat(model-picker): surface inactive provider profiles in /model (#1119 piece 2) #1164 (Piece 2) does, to stay consistent with that PR's opt-in shape? Or should Piece 1 be unconditional since the manifest is authoritative regardless of profile-env application?
3. Anyone already working on Piece 1? Happy to drop it if there's an in-flight effort I haven't found.

If the direction looks right, the test surface is `getModelOptionsBase()` (per @Vasanthdev2004's note above) with cases for openai preset, kimi-code preset, openrouter gateway, unknown-fallback, and no-profile env-only fallback.