Summary

DeepSeek V4 Pro (deepseek-v4-pro) officially supports reasoning_effort: "max", and OpenClaw's bundled DeepSeek plugin defines a 7-level V4_THINKING_LEVEL_IDS array including "xhigh" and "max". However, the TUI / Control UI /think command only exposes 5 levels — xhigh and max are invisible and unselectable.

Environment

• OpenClaw: v2026.5.2 (8b2a6e5)
• Model: deepseek/deepseek-v4-pro
• Agent config: thinkingDefault: "max"
• OS: Darwin 25.4.0 (arm64)
• Node.js: v22.22.2

Expected Behavior

/think should display all 7 levels for DeepSeek V4 Pro:

off, minimal, low, medium, high, xhigh, max

Actual Behavior

/think only shows 5 levels:

off, minimal, low, medium, high

xhigh and max are not visible and not selectable.

Inconsistent state

• session_status reports Think: max ✅ (the API layer works correctly)
• The /think selector in the UI cannot see max ❌
• Manually attempting to set max in an old session reverts to high

Technical Analysis

1. DeepSeek API officially supports max

DeepSeek's Thinking Mode docs state:

Thinking Effort Control: reasoning_effort: "high/max"

In thinking mode, for compatibility, low and medium are mapped to high, and xhigh is mapped to max

2. OpenClaw's DeepSeek plugin already defines all 7 levels

File: dist/extensions/deepseek/index.js

const V4_THINKING_LEVEL_IDS = [
    "off",
    "minimal",
    "low",
    "medium",
    "high",
    "xhigh",   // ← defined but not surfaced in UI
    "max"      // ← defined but not surfaced in UI
];

const DEEPSEEK_V4_THINKING_PROFILE = {
    levels: V4_THINKING_LEVEL_IDS.map(buildDeepSeekV4ThinkingLevel),
    defaultLevel: "high"
};

function resolveDeepSeekV4ThinkingProfile(modelId) {
    return isDeepSeekV4ModelId(modelId) ? DEEPSEEK_V4_THINKING_PROFILE : void 0;
}

The plugin entry registers it correctly:

resolveThinkingProfile: ({ modelId }) => resolveDeepSeekV4ThinkingProfile(modelId),

3. The API mapping layer works correctly

File: dist/provider-stream-shared-BF5MzTUo.js

function resolveDeepSeekV4ReasoningEffort(thinkingLevel) {
    return thinkingLevel === "xhigh" || thinkingLevel === "max" ? "max" : "high";
}

When thinkingLevel is "max", the API request correctly sends reasoning_effort: "max" ✅

4. Suspected root cause: thinking profile resolution chain

File: dist/thinking-Bi9AZdtM.js

const BASE_THINKING_LEVELS = ["off", "minimal", "low", "medium", "high"];
// ← "xhigh" and "max" are NOT in the base list

function resolveThinkingProfile(params) {
    const context = resolveThinkingPolicyContext(params);
    // ...
    const pluginProfile = resolveProviderThinkingProfile({...});
    if (pluginProfile) {
        const normalized = normalizeThinkingProfile(pluginProfile);
        if (normalized.levels.length > 0) return normalized;  // ← should return all 7 levels here
    }
    // fallback: returns BASE_THINKING_LEVELS (5 levels)
    const profile = buildBaseThinkingProfile(defaultLevel);
    if (catalogSupportsXHigh(context.compat)) appendProfileLevel(profile, "xhigh");
    return profile;
}

Call chain:

1. resolveProviderThinkingProfile() → resolveProviderRuntimePlugin() → looks up DeepSeek plugin
2. If plugin found → calls plugin.resolveThinkingProfile(context) → resolveDeepSeekV4ThinkingProfile(modelId)
3. If modelId matches → returns DEEPSEEK_V4_THINKING_PROFILE (7 levels)
4. If modelId doesn't match → returns undefined → falls back to BASE_THINKING_LEVELS (5 levels)

Suspicion:

• resolveProviderRuntimePlugin may fail to match the DeepSeek plugin (provider ID mismatch, plugin registration failure, etc.)
• Or the modelId passed in doesn't match isDeepSeekV4ModelId's check (which only matches "deepseek-v4-flash" and "deepseek-v4-pro")

5. Session cache (secondary issue)

The thinkingLevel in old sessions' sessions.json is hardcoded to "high", even when the global thinkingDefault: "max" is set. Manually editing sessions.json works around this, but the UI-level problem still remains.

Reproduction Steps

1. Configure agent main to use deepseek/deepseek-v4-pro
2. Run /think in TUI or Control UI
3. Observe the available thinking-level options
4. Expected: off, minimal, low, medium, high, xhigh, max
5. Actual: only off, minimal, low, medium, high

Impact

• Users cannot select max reasoning effort through the UI (workaround: set via thinkingDefault)
• The xhigh tier is completely unusable (invisible in UI and no thinkingDefault workaround for it specifically)
• Inconsistent with DeepSeek's officially documented capabilities
• Old session thinkingLevel cache prevents config changes from taking effect

Workarounds

1. New sessions: setting thinkingDefault: "max" globally works at the API level
2. Old sessions: manually edit ~/.openclaw/agents/main/sessions/sessions.json and change the session's thinkingLevel to "max"

Neither workaround addresses the UI invisibility of xhigh/max.