openclaw - 💡(How to fix) Fix Plugin-SDK `DEFAULT_PROVIDER` hardcoded to "openai" breaks plugin model resolution for users on `openai-codex`-only setups [1 comments, 2 participants]

Problem

DEFAULT_PROVIDER is exported from the plugin-SDK as a hardcoded string "openai". Plugins import this constant and use it as the fallback provider when parsing user-provided model refs (e.g., parseModelRef("gpt-5.5", DEFAULT_PROVIDER)).

For users on a Codex Pro OAuth setup (where the only configured provider is openai-codex, with no openai provider), the plugin-resolved model ref becomes openai/<model>, which fails downstream because openai isn't configured. There's no fallback to the user's actual configured providers.

Three bundled plugins are affected on a typical install:

• active-memory (recall lookups)
• skill-workshop (reviewer agent — workaround: set reviewMode: "heuristic" to bypass the LLM path)
• llm-slug-generator

Repro

Setup:

// ~/.openclaw/openclaw.json
{
  "agents": {
    "defaults": {
      "model": { "provider": "openai-codex", "id": "gpt-5.5" }
    }
  },
  "auth": {
    "profiles": {
      "openai-codex:default": { "type": "oauth" /* Codex Pro */ }
    }
  }
  // NO "openai" provider configured anywhere
}

Enable active-memory with default config (no model override). Send a DM that should trigger recall.

Observed: active-memory resolves model ref to openai/gpt-5.5 (or whatever the user's config returns through getModelRef paths) and the LLM call fails because the openai provider has no auth profile.

Workaround: every affected plugin's user-facing config must include an explicit model: "openai-codex/gpt-5.5" (or similar) field. Plugins without such a field (e.g., earlier skill-workshop revisions) have no recourse other than disabling the LLM path.

Root cause

DEFAULT_PROVIDER is a compile-time constant:

// dist/defaults-DM8yIn8C.js
const DEFAULT_PROVIDER = "openai";
const DEFAULT_MODEL = "gpt-5.5";
const DEFAULT_CONTEXT_TOKENS = 2e5;

Type signature:

// packages/plugin-sdk/dist/src/agents/defaults.d.ts
export declare const DEFAULT_PROVIDER = "openai";

Consumed by ~10 modules including plugin code:

// extensions/active-memory/index.ts
import { DEFAULT_PROVIDER, parseModelRef, ... } from "openclaw/plugin-sdk/agent-runtime";

function parseModelCandidate(modelRef) {
  if (!modelRef) return undefined;
  return (
    parseModelRef(modelRef, DEFAULT_PROVIDER) ??
    { provider: DEFAULT_PROVIDER, model: modelRef }
  );
}

parseModelRef (in dist/model-selection-shared-CGq2vKNt.js) only uses defaultProvider when the input has no / separator. So a user who passes a bare "gpt-5.5" (or whose plugin-resolved model is bare) ends up with openai/gpt-5.5 regardless of what providers are actually configured.

Proposed fix

Three options, in increasing scope:

Option A: Make DEFAULT_PROVIDER resolve from runtime config

Replace the constant with a function that reads agents.defaults.model.provider first, and only falls back to "openai" if nothing is configured:

export function resolveDefaultProvider(cfg: OpenClawConfig): string {
  return cfg.agents?.defaults?.model?.provider ?? "openai";
}

Plugins call api.runtime.resolveDefaultProvider() (or similar) instead of importing the constant. Backwards-compatible if the constant is kept as a fallback for plugins that haven't migrated.

Option B: Have parseModelRef consult configured providers

When the input is bare (no /), and the supplied defaultProvider isn't actually configured, fall through to the first configured provider (or the agent's primary). This silently fixes existing plugin code without requiring migration.

Option C: Plugin-SDK opinion: require explicit model in plugin config

Document that any plugin invoking an LLM must accept a user-provided model field, validate it against configured providers at config-load time, and reject the plugin enable if no usable provider is available. This is the most invasive but the most correct.

Impact

Anyone running OpenClaw against openai-codex OAuth (Codex Pro / ChatGPT subscription) without also configuring an openai API key hits this on active-memory, skill-workshop LLM reviewer, and llm-slug-generator. The failure mode is a stack trace in the gateway logs and the plugin path silently doing nothing user-visible (active-memory recall returns empty; skill-workshop reviewer noops).

Same impact applies to Anthropic-only setups using claude-cli or ACP backends, since openai likewise isn't configured.

Environment

• OpenClaw: 2026.4.24 (028b6c9)
• Symbol confirmed in current dist:
  • dist/defaults-DM8yIn8C.js:2
  • packages/plugin-sdk/dist/src/agents/defaults.d.ts
  • extensions/active-memory/index.ts (consumer)
• Reproducible setup: openai-codex OAuth (Codex Pro), no openai API key, default agents.defaults.model.provider: openai-codex.