Feature Request: pass-through extra_body for openai-completions adapter

Summary

Allow users to configure arbitrary provider-specific request-body fields via agents.defaults.models["<prov>/<id>"].params.extra_body (or extraBody). These fields should be merged into the outgoing JSON payload for models using the openai-completions adapter.

Motivation

Many OpenAI-compatible providers expose non-standard parameters that must be set in the request body. Examples:

Currently, createStreamFnWithExtraParams only whitelists a fixed set (temperature, maxTokens, transport, openaiWsWarmup, cachedContent, cacheRetention), and applyPostPluginStreamWrappers adds a single parallel_tool_calls override. Users cannot add new provider-specific params without modifying the OpenClaw source.

Proposed change

Add a generic wrapper analogous to createParallelToolCallsWrapper:

function createExtraBodyWrapper(baseStreamFn, extraBody) {
  const underlying = baseStreamFn ?? streamSimple;
  return (model, context, options) => {
    if (model.api !== "openai-completions") return underlying(model, context, options);
    if (!extraBody || typeof extraBody !== "object") return underlying(model, context, options);
    return streamWithPayloadPatch(underlying, model, context, options, (payloadObj) => {
      for (const [k, v] of Object.entries(extraBody)) {
        if (v !== undefined) payloadObj[k] = v;
      }
    });
  };
}

Wire it into applyPostPluginStreamWrappers by reading resolveAliasedParamValue(sources, "extra_body", "extraBody").

Scope initially to openai-completions (where the need is greatest). Future extension: allow on openai-responses and anthropic-messages behind per-adapter validation.

Config shape

{
  "agents": {
    "defaults": {
      "models": {
        "dashscope/qwen3.6-max-preview": {
          "alias": "Qwen 3.6 Max Preview",
          "params": {
            "extra_body": {
              "preserve_thinking": true,
              "enable_thinking": true
            }
          }
        }
      }
    }
  }
}

Backward compatible — only activates when extra_body is present.

Safety considerations

• Collisions with core fields (e.g. messages, model, tools) should be allowed but logged at warn level, so power users can override if needed but accidental shadowing is visible.
• No value sanitization required beyond basic type check: users explicitly opt in via config; schema validation should only check object-ness.
• Keys with undefined values are skipped.

Reference implementation

A working out-of-tree patch using streamWithPayloadPatch is already deployed locally in ~/.openclaw/patches/extra-body-passthrough.patch.js — happy to port it into a PR if useful.

Tests

• Unit: createExtraBodyWrapper merges keys, skips on non-openai-completions, skips undefined values, passes through when extra_body is empty.
• Integration: end-to-end stream against a mock openai-compatible endpoint asserts payload contains the extra keys.