Summary

The bundled deepseek provider plugin in OpenClaw v2026.4.25 stable does not register a resolveThinkingProfile() hook, which causes /think max and /think xhigh to be rejected for deepseek/deepseek-v4-pro and deepseek/deepseek-v4-flash, even though:

1. DeepSeek's official API supports both (docs): reasoning_effort: "high" | "max", with xhigh auto-mapped to max.
2. OpenClaw's stream wrapper already supports it: provider-stream-shared-CMqu78qu.js:157 correctly maps xhigh/max to max for DeepSeek payloads.
3. The thinking ranks already include them: THINKING_LEVEL_RANKS in thinking-enE1knZM.js defines xhigh: 60, max: 70.

The only missing piece is the provider plugin not exposing resolveThinkingProfile, which causes resolveThinkingProfile() (in thinking-enE1knZM.js) to fall back to buildBaseThinkingProfile() returning only ["off", "minimal", "low", "medium", "high"].

Related closed issue: #71584 covered the same root cause for ollama/deepseek-v4-flash:cloud, but the native DeepSeek provider plugin was not patched.

Reproduction

OpenClaw v2026.4.25 stable, native DeepSeek provider via API key:

/model deepseek/deepseek-v4-pro
/think max
→ Thinking level "max" is not supported for deepseek/deepseek-v4-pro. Use one of: off, minimal, low, medium, high.

/think xhigh
→ Thinking level "xhigh" is not supported for deepseek/deepseek-v4-pro. Use one of: off, minimal, low, medium, high.

Direct DeepSeek API call with reasoning_effort: "max" works without issue (confirmed against official docs).

Root Cause

dist/extensions/deepseek/index.js (48 lines total) is missing the resolveThinkingProfile hook that other reasoning-capable providers register. Compare:

• ✅ extensions/anthropic/register.runtime.js:388 — registers profile
• ✅ extensions/google/provider-hooks.js:7 — registers profile
• ✅ extensions/xai/index.js:134 — registers profile
• ✅ extensions/moonshot/index.js:51 — registers profile
• ✅ extensions/mistral/index.js:46 — registers profile
• ✅ extensions/zai/index.js:244 — registers profile
• ✅ extensions/kimi-coding/index.js:80 — registers profile
• ✅ extensions/codex/provider.js:39 — registers profile
• ✅ extensions/github-copilot/index.js:112 — registers profile
• ✅ extensions/amazon-bedrock/register.sync.runtime.js:247 — registers profile
• ❌ extensions/deepseek/index.js — missing

Suggested Fix

Add resolveThinkingProfile to the DeepSeek provider definition in extensions/deepseek/index.ts, between wrapStreamFn and isModernModelRef:

resolveThinkingProfile: ({ modelId }) => {
    const lower = modelId.toLowerCase();
    if (lower === "deepseek-v4-flash" || lower === "deepseek-v4-pro") {
        return {
            levels: [
                { id: "off", label: "off" },
                { id: "minimal", label: "minimal" },
                { id: "low", label: "low" },
                { id: "medium", label: "medium" },
                { id: "high", label: "high" },
                { id: "xhigh", label: "xhigh" },
                { id: "max", label: "max" }
            ],
            defaultLevel: "high"
        };
    }
},

This matches DeepSeek's documented behavior:

• low/medium → auto-mapped to high server-side
• xhigh → auto-mapped to max server-side
• high/max → sent through as-is

The existing createDeepSeekV4OpenAICompatibleThinkingWrapper already handles the payload mapping correctly; this change is purely about letting /think and agents.defaults.thinking accept xhigh/max as valid levels.

Environment

• OpenClaw: 2026.4.25 (aa36ee6) — latest stable
• OS: Ubuntu 24.04 LTS (aarch64), Oracle Cloud VPS
• Node: v24.14.1
• Provider: native DeepSeek (api-key auth, https://api.deepseek.com)
• Model: deepseek-v4-pro

Local Workaround

Patched dist/extensions/deepseek/index.js to add the missing resolveThinkingProfile and confirmed /think max and /think xhigh are accepted, the streamed payload reaches DeepSeek with reasoning_effort: "max", and reasoning_content is returned correctly.