feat(MimoHandler): Missing reasoning token tracking in usage metrics

[proyectoauraorg]

Description

The mimoModels definition in packages/types/src/providers/mimo.ts lacks an input token count for reasoning_content tokens. When MiMo uses deep thinking (enabled via extra_body.thinking), the reasoning tokens produced by the model are streamed as reasoning_content deltas — but these tokens are not tracked in the usage/cost calculation.

Current Behavior

// src/api/providers/mimo.ts:130-135
if ("reasoning_content" in delta && delta.reasoning_content) {
    yield {
        type: "reasoning",
        text: (delta.reasoning_content as string) || "",
    }
}

The reasoning content is yielded as a stream event, but in the usage block:

// src/api/providers/mimo.ts:144-166
if (lastUsage) {
    const inputTokens = lastUsage?.prompt_tokens || 0
    const outputTokens = lastUsage?.completion_tokens || 0
    const cacheWriteTokens = (lastUsage?.prompt_tokens_details as any)?.cache_write_tokens || 0
    const cacheReadTokens = lastUsage?.prompt_tokens_details?.cached_tokens || 0
    // No reasoning_tokens tracked!
}

What other providers do

DeepSeek handler (src/api/providers/deepseek.ts:189) explicitly tracks reasoningTokens from completion_tokens_details.reasoning_tokens. The MiMo handler does not.

Proposed Fix

1. Add reasoningTokens field to ApiStreamUsageChunk if not already present
2. Extract reasoning token count from MiMo usage response:

// In MimoHandler.createMessage usage block:
const reasoningTokens = (lastUsage?.completion_tokens_details as any)?.reasoning_tokens || 0

yield {
    type: "usage",
    inputTokens,
    outputTokens,
    reasoningTokens,  // NEW: track thinking cost separately
    cacheWriteTokens: cacheWriteTokens || undefined,
    cacheReadTokens: cacheReadTokens || undefined,
    totalCost,
}

3. Update cost calculation in calculateApiCostOpenAI to account for reasoning tokens if MiMo charges differently for them

Impact

• Severity: Medium — Users see inaccurate cost tracking for thinking-heavy conversations
• Scope: src/api/providers/mimo.ts, potentially src/shared/cost.ts and packages/types/
• Risk of fix: Low — additive change, does not break existing behavior

Additional Context

MiMo V2.5 Pro documentation indicates thinking mode produces reasoning tokens that may be billed separately. Without tracking, users have no visibility into the cost of deep thinking vs. direct output.

[proyectoauraorg]

Auto-assigned for implementation. Will add reasoning_tokens tracking from completion_tokens_details.reasoning_tokens in the usage block, matching DeepSeek handler pattern. ETA: 2026-06-05.