#21614: fix: warn when thinking level xhigh falls back for unsupported models

Fixes #21582 ## Summary When `thinkingDefault: "xhigh"` is set in config for models that don't support it (e.g., Anthropic Claude), the level silently falls back to `"high"` with no warning in logs or status output. This leaves users confused when their thinking budget is lower than expected — only discoverable by manually checking `/status` after an upgrade. ## Root Cause In `src/commands/agent.ts`, config-level `xhigh` settings for unsupported models fall back to `high` without emitting any diagnostic: ```typescript if (resolvedThinkLevel === "xhigh" && !supportsXHighThinking(provider, model)) { const explicitThink = Boolean(thinkOnce || thinkOverride); if (explicitThink) { throw new Error(...); // Explicit /thinking xhigh → error } resolvedThinkLevel = "high"; // Config xhigh → silent fallback ⚠️ } ``` ## Changes - Added `console.warn` before fallback to `"high"` for config-level `xhigh` - Warning message includes: - Unsupported model (e.g., `anthropic/claude-opus-4-6`) - List of supported models (GPT-5.2, Codex variants) - Fallback action ("Falling back to 'high'") - Explicit `/thinking xhigh` commands continue to throw errors (unchanged) **Example warning:** ``` [openclaw] Thinking level "xhigh" is not supported for anthropic/claude-opus-4-6. Supported models: openai/gpt-5.2, openai-codex/gpt-5.3-codex, ... Falling back to "high". ``` ## Impact - ✅ Users see clear warning at startup instead of silent runtime downgrade - ✅ Helps identify config issues immediately (no need to check `/status`) - ✅ No behavior change: fallback logic remains `high` (just visible now) - ✅ Existing tests unaffected (only added logging, no logic change) ## Testing - All existing tests pass (`thinking.test.ts`: 14/14 ✅) - Fallback behavior unchanged (still `"high"`) - Warning only emitted for config-level `xhigh`, not explicit commands ## Closes #21582 <!-- greptile_comment --> <h3>Greptile Summary</h3> This PR adds a warning message when the `thinkingDefault: "xhigh"` config setting silently falls back to `"high"` for models that don't support extra-high thinking levels (e.g., Anthropic Claude). Previously, users would only discover this downgrade by checking `/status` output, leading to confusion about lower-than-expected thinking budgets. The change adds a single `console.warn` call that clearly identifies the unsupported model, lists which models do support `xhigh` (GPT-5.2, Codex variants), and notifies users of the fallback to `"high"`. The warning only appears for config-level `xhigh` settings—explicit `/thinking xhigh` commands continue to throw errors as before. The implementation is minimal, non-breaking, and improves user experience by making silent configuration issues visible at startup. <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with minimal risk - The change adds only a single `console.warn` statement without modifying any logic or control flow. The fallback behavior remains identical (still uses `"high"`), and the warning provides clear, actionable information. No test changes are needed since the functionality is unchanged. The implementation follows repository conventions (uses `[openclaw]` prefix), and the message accurately reflects the helper functions' output. - No files require special attention _{Last reviewed commit: 25875d2} <!-- greptile_other_comments_section --> <!-- /greptile_comment -->