openclaw - 💡(How to fix) Fix Bug: agents.defaults.models record-format allowlist stripped at gateway startup despite passing Zod validation (4.23/4.24) [1 comments, 2 participants]

Summary

Setting agents.defaults.models as a record (map keyed by provider/model, values are empty objects) per the documented schema results in the field being silently removed from the runtime config when the gateway starts. Reproducible 100% on OpenClaw 2026.4.23 and 2026.4.24.

Reproduction

1. Write /root/.openclaw/openclaw.json containing:

   {
     "agents": {
       "defaults": {
         "model": { "primary": "dos-ai/dos-ai" },
         "models": {
           "dos-ai/dos-ai": {},
           "dos-ai/dos-auto": {}
         }
       }
     }
   }

   (plus the rest of a valid config; dos-ai provider is loaded with a matching catalog.)

2. Confirm openclaw doctor accepts the record-format value:

   • Exit code 0
   • No Invalid input: expected record warning (the warning DOES fire if you instead pass an array - confirms the schema expects a record)
   • File hash unchanged after doctor run

3. Confirm openclaw config set agents.defaults.models '{...}' --strict-json --merge accepts and persists the field. CLI prints Updated agents.defaults.models. Restart the gateway to apply.

4. Restart the gateway: openclaw gateway run --bind lan (or docker restart for our setup)

5. After restart, observe that /root/.openclaw/openclaw.json no longer contains agents.defaults.models. The field is silently dropped. openclaw config get agents.defaults.models returns empty.

6. Multiple openclaw.json.clobbered.<timestamp> backups appear next to the active config, each one missing the field too.

Expected behavior

agents.defaults.models is documented as the model picker allowlist (docs/concepts/model-providers.md, docs/cli/configure.md, docs/cli/config.md). The Zod schema in src/config/zod-schema.agent-defaults.ts defines it as z.record(z.string(), z.object({alias?, params?, streaming?}).strict()).optional(). An empty inner object should be valid and persist.

Actual behavior

Gateway startup writes the file back without the field. Schema validation alone is fine; some later normalization step silently drops it. We have not identified which code path performs the strip.

Impact

The Control UI /model picker enumerates ~900 plugin-known model ids that we do not proxy. With models.mode: "replace" set, invocations of unsupported models still get a clean 400 from the gateway, so this is not a correctness or CPU-loop issue - but the picker is unusable.

Workaround / current state

models.mode: "replace" plus plugins.deny of all non-dos-ai providers prevents runtime invocation of unsupported models. The picker remains visually noisy.

Environment

• OpenClaw 2026.4.23 and 2026.4.24
• Linux Docker container, Alpine base
• Self-hosted on multi-tenant agent platform (DOS.AI)

Happy to add more detail or PR the schema docs if useful.