Bug: || 1024 fallback silently activates deprecated budget-based thinking on every Anthropic API call

Description

The || 1024 fallback in pi-ai/dist/providers/anthropic.js silently activates deprecated budget-based thinking on every Anthropic API call when no valid thinking configuration is present. There is no log warning, no schema error, and no documentation indicating this will happen.

What happens

In anthropic.js, when thinkingBudgetTokens is undefined (i.e., no valid thinking config exists), the following code path executes on every API call:

params.thinking = {
  type: "enabled",
  budget_tokens: options.thinkingBudgetTokens || 1024,
};

The || 1024 fallback fires, sending type: "enabled" with budget_tokens: 1024 (the minimum non-zero value) to the Anthropic API on every request — including trivial actions such as file reads and short messages.

Why the thinking config ends up undefined

Two common configuration attempts fail silently:

1. params.thinking in the model entry — params is not a recognised key in the ModelDefinitionSchema. The schema is declared .strict(), so the block is silently discarded on gateway start. No error or warning is surfaced.

2. thinkingBudgets as a top-level key — does not exist in the schema. openclaw config validate catches this one, but users may not run it.

In both cases, the user believes they've configured thinking, but the config is inert, and the || 1024 fallback takes over.

Impact

• Every API call to claude-opus-4-6 (or any model where supportsAdaptiveThinking() returns false in the fallback path) gets budget_tokens: 1024 thinking overhead.
• This wastes tokens on trivial actions (reading files, short replies).
• The deprecated type: "enabled" / budget_tokens model is sent instead of the current type: "adaptive" / effort model for Opus 4.6.
• Affects any OpenClaw installation running claude-opus-4-6 without a correctly configured thinkingDefault.

Expected behavior

When thinkingBudgetTokens is not explicitly set (i.e., undefined), the fallback should either:

• Omit the thinking block entirely (no thinking unless explicitly configured), or
• Require an explicit opt-in rather than silently defaulting to budget_tokens: 1024.

At minimum, a warning log should be emitted when the fallback fires, so users know thinking is being activated by default.

Correct configuration (workaround)

The correct way to enable adaptive thinking for Opus 4.6 is:

{
  "agents": {
    "defaults": {
      "thinkingDefault": "medium"
    }
  }
}

This routes through mapThinkingLevelToEffort("medium") → supportsAdaptiveThinking("claude-opus-4-6") returns true → Anthropic API receives { thinking: { type: "adaptive" }, output_config: { effort: "medium" } }.

Valid thinkingDefault values from the Zod schema: off, minimal, low, medium, high, xhigh.

How we found it

1. Located pi-ai provider file at the installed path.
2. Read the supportsAdaptiveThinking(), mapThinkingLevelToEffort(), and streamSimpleAnthropic() call chain in anthropic.js.
3. Read the ModelDefinitionSchema Zod definition in daemon-cli.js — confirmed reasoning is a boolean, thinkingDefault is a valid string enum, and neither thinkingBudgets nor params.thinking exist in the schema.
4. Confirmed with openclaw config validate and clean gateway restart.

Environment

• OpenClaw 2026.4.5
• Model: anthropic/claude-opus-4-6
• macOS (arm64)