Problem

Currently, the context token resolution in OpenClaw follows a strict priority order:

1. Agent config (agents.defaults.contextTokens or agents.list[].contextTokens)
2. Provider model config (models.providers.*.models[].contextWindow)
3. Model discovery cache (models.json)
4. Hardcoded default (200,000)

This means if agents.defaults.contextTokens is set to 500,000 but a model only supports 128K tokens, the agent config value is used, causing API errors like:

⚠️ Context overflow — prompt too large for this model.

Proposed Solution

Instead of strict priority, use the minimum of all available values:

effectiveContextTokens = min(
  agentConfig.contextTokens ?? Infinity,
  modelConfig.contextWindow ?? Infinity,
  discoveredContextWindow ?? Infinity,
  DEFAULT_CONTEXT_TOKENS
)

This would:

• Prevent accidental context overflow when agent config exceeds model limits
• Allow setting a global default via agents.defaults.contextTokens without worrying about per-model limits
• Still respect explicit per-agent overrides when they are smaller than model limits
• Make the system safer by default

Example

{
  "agents": {
    "defaults": {
      "contextTokens": 500000
    }
  },
  "models": {
    "providers": {
      "groq": {
        "models": [
          { "id": "llama-3.3-70b-versatile", "contextWindow": 131072 }
        ]
      }
    }
  }
}

Current behavior: Agent uses 500,000 → API error
Proposed behavior: Agent uses min(500000, 131072) = 131,072 → works correctly

Additional Context

The same logic should apply to maxTokens resolution.