feat: make fallback base URL configurable via settings.json

[ashleywd]

Revised GitHub Issue: Configurable Fallback Base URL

Problem

Currently, the fallback model's API base URL is hardcoded to https://api.z.ai/api/anthropic in src/runner.ts. This restricts users to using only GLM as a fallback and prevents the use of other Anthropic-compatible providers like OpenRouter, LiteLLM, or custom local proxies.

Additionally, the current logic is gated behind the hardcoded string "glm". This means even if a user wants to use a different provider (like Gemini or DeepSeek via a proxy), they are currently forced to name their model "glm" in settings.json just to trigger the custom fallback address.

Proposed Solution

Add an optional baseUrl field to the configuration and refactor the environment building logic to be provider-agnostic. If a baseUrl is provided in the configuration, it should be used for the ANTHROPIC_BASE_URL regardless of what the model is named.

---

Proposed Changes

1. src/config.ts

Add an optional baseUrl property to the ModelConfig interface so both primary and fallback models can use custom endpoints:

export interface ModelConfig {
  model: string;
  api: string;
  baseUrl?: string; // New optional field
}

2. src/runner.ts

Update buildChildEnv to prioritize the user-defined baseUrl. This removes the rigid dependency on the "glm" string while maintaining backward compatibility:

function buildChildEnv(baseEnv: Record<string, string>, config: ModelConfig): Record<string, string> {
  const childEnv: Record<string, string> = { ...baseEnv };
  const normalizedModel = config.model.trim().toLowerCase();

  if (config.api.trim()) childEnv.ANTHROPIC_AUTH_TOKEN = config.api.trim();

  // Priority 1: User-defined baseUrl from settings.json
  // Priority 2: Legacy hardcoded fallback for "glm" model
  if (config.baseUrl) {
    childEnv.ANTHROPIC_BASE_URL = config.baseUrl;
    childEnv.API_TIMEOUT_MS = "3000000";
  } else if (normalizedModel === "glm") {
    childEnv.ANTHROPIC_BASE_URL = "https://api.z.ai/api/anthropic";
    childEnv.API_TIMEOUT_MS = "3000000";
  }

  return childEnv;
}

3. Execution Logic

Ensure that both the primary and fallback execution paths in execClaude and runClaudeStream pass their respective ModelConfig objects to the environment builder.

---

Benefits

• Flexibility: Support for any Anthropic-compatible API provider (OpenRouter, DeepSeek, Gemini-via-proxy, etc.).
• Improved DX: No need to "trick" the system by naming models "glm" to get a custom URL.
• Backward Compatibility: Existing GLM configurations remain fully functional.

[TerrysPOV]

@ashleywd thanks for flagging - this should be worked on here and will be picked up downstream on ClaudeClaw+. It has a small surface, broadly useful, removes a magic-string gate.

Two implementation notes for whoever picks it up so they don't get surprised in review:

1. Callers need updating, not just buildChildEnv

The proposed signature buildChildEnv(baseEnv, config: ModelConfig) is a refactor from today's buildChildEnv(baseEnv, model, api). Both call sites in src/runner.ts (the streaming path inside runClaudeStream around line 513 and the non-streaming runClaudeOnce around line 446) currently pass model and api as separate args. They'll need to either pass the full ModelConfig through, or the proposal needs a thin overload that keeps the old shape working. Either is fine, just call it out in the PR so the diff doesn't look incomplete.

2. API_TIMEOUT_MS = "3000000" extension is intentional but worth stating

Today the 3000000 ms (~50 min) timeout is set only on the GLM branch. The proposed code keeps it set whenever baseUrl is configured. That's the right call — OpenRouter/LiteLLM/proxied endpoints generally need the long timeout for the same streaming-tool-use reasons — but it does mean any user who switches to a custom provider via this new field inherits the longer client-side timeout. Should be explicit in the PR description so reviewers know it's intentional rather than accidentally extending GLM-only behaviour.

Do you want to open a PR for this? I'm happy to review once it's been pushed.