openclaw - 💡(How to fix) Fix Bundler: unhashed runtime-model-auth.runtime.js stub missing from dist/, breaks image.generate + any capability using runtime model auth [1 comments, 2 participants]

Summary

Builds from current main (as of 2026-04-23, tip f1ad5e27e0) no longer emit the unhashed dist/runtime-model-auth.runtime.{js,ts,mjs,...} stub that resolveRuntimeModelAuthModuleHref() in dist/provider-auth-runtime-*.js walks a candidate list to locate. Only hashed chunks exist, so the resolver exhausts its candidate list and throws:

``` Error: Unable to resolve runtime model auth module from file:///<repo>/dist/provider-auth-runtime-<hash>.js ```

Every capability that uses runtime model auth resolution fails: image generation, embeddings, probably audio / any provider-auth runtime path.

Reproduction

1. Check out main (verified on f1ad5e27e0).
2. pnpm install && pnpm build
3. openclaw infer image generate --provider openai-codex --model gpt-image-2 --prompt "test" (or any capability that touches runtime model auth).
4. See the error above. Attempt list also shows the same missing-module error for fallback candidates (openai/gpt-image-2, xai/grok-imagine-image), confirming the issue is resolver-level not provider-specific.

Root cause

The resolver at `dist/provider-auth-runtime-*.js`:

```js const RUNTIME_MODEL_AUTH_CANDIDATES = [ "./runtime-model-auth.runtime", "../plugins/runtime/runtime-model-auth.runtime", ]; const RUNTIME_MODEL_AUTH_EXTENSIONS = [".js", ".ts", ".mjs", ".mts", ".cjs", ".cts"];

function resolveRuntimeModelAuthModuleHref() { const baseDir = path.dirname(fileURLToPath(import.meta.url)); for (const relativeBase of RUNTIME_MODEL_AUTH_CANDIDATES) for (const ext of RUNTIME_MODEL_AUTH_EXTENSIONS) { const candidate = path.resolve(baseDir, `${relativeBase}${ext}`); if (fs.existsSync(candidate)) return pathToFileURL(candidate).href; } throw new Error(`Unable to resolve runtime model auth module from ${import.meta.url}`); } ```

It expects an unhashed `runtime-model-auth.runtime.js` sibling (or the ../plugins/runtime/ variant). On main only the hashed chunks exist:

``` dist/ runtime-model-auth.runtime-BnNsQZc0.js <- main bundle runtime-model-auth.runtime-C8pJ24BQ.js <- clean re-exporter ```

How v2026.4.22 worked

v2026.4.22 shipped a third file, a one-line re-export stub:

``` dist/runtime-model-auth.runtime.js: export * from "./runtime-model-auth.runtime-<hash>.js"; ```

The bundler config emitted that alongside the hashed chunks, so the resolver found it on the first candidate. That unhashed stub is not emitted in current main builds.

Suggested fix

Two paths:

1. Bundler config: re-emit the unhashed stub. Presumably `tsdown` config or a post-build script used to write `runtime-model-auth.runtime.js` as a re-export of the clean entry chunk. Restore that emission. Least surprise for the runtime resolver.
2. Resolver: use the bundle's import graph instead of filesystem probing. The clean entry chunk (`runtime-model-auth.runtime-<hash>.js`) is the intended entry; the resolver could import it via a statically-known identifier rather than fs.existsSync. Bigger change but removes the unhashed-stub requirement.

Option 1 is the minimal fix and restores parity with v2026.4.22.

Local workaround

Create the stub manually in the built dist:

``` cat > dist/runtime-model-auth.runtime.js <<'MJS' export * from "./runtime-model-auth.runtime-<whichever-C-hash>.js"; MJS ```

Confirmed: with this one-line file in place, `openclaw infer image generate` completes end-to-end on `openai-codex/gpt-image-2` via OAuth. Same fix unblocks other runtime-model-auth callers.

Discovered

Reproduced on a custom fork (wipcomputer/openclaw) built from upstream main during image-generation testing on Codex/gpt-5.5 provider. Discovery and hotfix by agents in a live debugging session; verified fix shipped an image end-to-end to iMessage via imsg --file on 2026-04-23 20:18 PDT.

Happy to submit a PR restoring the bundler-emitted stub if you can point me at the emission site.