openclaw - 💡(How to fix) Fix DeepSeek V4 max thinking level degrades to high in cron isolated sessions [1 comments, 2 participants]

Bug: DeepSeek V4 "max" thinking level drops to "high" in cron isolated sessions

Description

When using DeepSeek V4 models (deepseek-v4-pro, deepseek-v4-flash) with thinking: "max" in cron tasks (isolated sessions), the thinking level silently degrades to "high". The model parameter reasoning_effort: "max" is not appended to the profile, causing potentially reduced reasoning quality in automated tasks.

This occurs because the thinking profile resolution has only one code path (Layer 1) that works in normal interactive sessions, but the DeepSeek extension's profile registration depends on activation: {onStartup: false} (line 43 of extensions/deepseek/index.js), which means the plugin profile is not reliably loaded in cron isolated sessions.

Root Cause

In dist/thinking-C7msqX62.js, the function resolveThinkingProfile builds a thinking profile and appends "max" level through catalogSupportsMax(context.compat):

// Original code - only Layer 1 exists:
if (binaryDecision !== true && catalogSupportsMax(context.compat)) appendProfileLevel(profile, "max");

catalogSupportsMax checks context.compat?.supportedReasoningEfforts for an array containing "max":

function catalogSupportsMax(compat) {
    return compat?.supportedReasoningEfforts?.includes?.("max");
}

However, the DeepSeek extension registers model compat in the provider plugin manifest using a boolean flag supportsReasoningEffort: true instead of the array format supportedReasoningEfforts: ["low", "high", "max"]. This means even when the plugin loads correctly, catalogSupportsMax returns false because the array doesn't exist.

In cron isolated sessions, the DeepSeek plugin profile is not loaded at all (due to onStartup: false), so even the supportsReasoningEffort flag is unavailable, and context.compat becomes undefined.

The Fix

Three-layer fallback chain added to dist/thinking-C7msqX62.js:

Layer 1 (existing): catalogSupportsMax(context.compat) — checks supportedReasoningEfforts array Layer 2: context.compat?.supportsReasoningEffort === true && context.reasoning === true && !catalogSupportsMax(context.compat) — handles boolean flag format Layer 3: context.reasoning !== false && /deepseek-v4/i.test(context.modelId) && !catalogSupportsMax(context.compat) — handles the case where compat is undefined (cron isolated session with missing plugin profile)

// After fix - three-layer fallback:
if (binaryDecision !== true && catalogSupportsMax(context.compat)) appendProfileLevel(profile, "max");
if (binaryDecision !== true && context.compat?.supportsReasoningEffort === true && context.reasoning === true && !catalogSupportsMax(context.compat)) appendProfileLevel(profile, "max");
if (binaryDecision !== true && context.reasoning !== false && context.modelId && /deepseek-v4/i.test(context.modelId) && !catalogSupportsMax(context.compat)) appendProfileLevel(profile, "max");

Additional Context

The catalogSupportsMax helper:

function catalogSupportsMax(compat) {
    // Only checks array format, fails on boolean format
    return compat?.supportedReasoningEfforts?.includes?.("max");
}

The DeepSeek plugin manifest registers compat as boolean:

// extensions/deepseek/index.js
compat: {
    supportsReasoningEffort: true,   // boolean, not array!
    // Missing: supportedReasoningEfforts: ["low", "high", "max"]
}

Steps to Reproduce

1. Configure a model route using deepseek-v4-pro or deepseek-v4-flash with thinking: "max"
2. Use it in a cron task (isolated session)
3. The thinking level will be "high" instead of "max"
4. Check Gateway logs — no error, silent degradation

Impact

Users relying on DeepSeek V4 models in automated/cron tasks get reduced reasoning quality without any error indication, potentially affecting trading decisions, report quality, or other automated reasoning tasks.