Feature: Add reasoningDefault config key to agents.defaults

[marcbal77]

Problem

When using a model with reasoning: true in the catalog (e.g., claude-opus-4-6 which inherits from the claude-opus-4-5 template), resolveReasoningDefault() auto-enables reasoning ("on") for new sessions. This causes the model's thinking/reasoning block to be included in outbound messages on channels like Discord, prefixed with Reasoning:\n.

There is no config-level way to override this. The agents.defaults schema validates with .strict() and only recognizes thinkingDefault, verboseDefault, and elevatedDefault — but not reasoningDefault.

Current behavior

// dispatch line ~100301
if (no explicit directive && no session reasoning level && resolvedReasoningLevel === "off" && thinking is off)
    resolvedReasoningLevel = await modelState.resolveDefaultReasoningLevel();
    // → returns "on" for models with reasoning: true in catalog

The isReasoning block-reply filter works correctly for streaming block replies, but when reasoning is "on" and there's no onBlockReply callback, the reasoning text gets inlined into the assistant message text (line ~51797).

Affected channels

Discord (confirmed), likely any channel where reasoning text leaks into the final message.

Expected behavior

Allow agents.defaults.reasoningDefault: "off" (or per-agent reasoningDefault) to override the catalog-derived default reasoning level, consistent with how thinkingDefault, verboseDefault, and elevatedDefault work.

Workaround

1. Manually set reasoningLevel: "off" on each session entry in the session store JSON
2. Patch resolveReasoningDefault() in dist to always return "off"

Both are fragile — session fix doesn't prevent new sessions from auto-enabling, and dist patch is overwritten on upgrade.

Environment

• OpenClaw 2026.3.11
• Model: anthropic/claude-opus-4-6 (inherits reasoning: true from opus-4-5 template)
• Channel: Discord

🤖 Generated with Claude Code