🤖 fix: accurate cost estimation for multi-step tool usage

[ethanndickson]

Problem

The application was severely underestimating costs for conversations involving tool calls.

Root Cause

The Vercel AI SDK provides two usage metrics:

• streamResult.usage — Token usage from the last step only
• streamResult.totalUsage — Sum of token usage across all steps

The application was using usage instead of totalUsage. For a conversation with 10 tool calls, only ~1/10th of actual consumption was reported. A $5 conversation would display as $0.50.

The Complicating Factor

Two UI elements use token data with different semantic requirements:

Display	Needs	Why
Cost	Sum of all steps	If model read context 10 times, you paid for 10 reads
Context window	Last step only	Shows "how full is the conversation now" for the next request

Simply switching to totalUsage would fix costs but break context display (showing 500% utilization after many tool calls).

Cache Creation Tokens

Anthropic's cache creation tokens (cacheCreationInputTokens) are:

• Only in provider-specific metadata, not normalized usage
• Need to be summed across all steps
• Not automatically aggregated by the AI SDK

Even with totalUsage, cache creation costs were lost unless manually aggregated from each step's provider metadata.

Solution

Track both values with different semantic purposes:

For cost calculation:

• usage / cumulativeUsage — total across all steps
• providerMetadata / cumulativeProviderMetadata — aggregated cache creation tokens

For context window display:

• contextUsage / lastContextUsage — last step only
• contextProviderMetadata — last step only

Key Changes

1. Backend (streamManager.ts): Use totalUsage for cost, track lastStepUsage for context, aggregate provider metadata across steps
2. Types: Extended StreamEndEvent, MuxMetadata, UsageDeltaEvent with dual fields
3. Frontend: StreamingMessageAggregator tracks both cumulative and per-step usage
4. Store: WorkspaceUsageState provides usageHistory (cost) and lastContextUsage (context window)
5. UI: Components use appropriate field for their purpose

Also Fixed

• OpenAI cached token double-counting: Gateway models (mux-gateway:openai/gpt-5.1) weren't recognized as OpenAI, causing cached tokens to be counted in both "Cache Read" and "Input". Now normalizes gateway model strings before provider detection.

• Google/Gemini cached token double-counting: Google, like OpenAI, reports inputTokens inclusive of cachedInputTokens. Extended the subtraction logic to handle Google models.

---

Generated with mux

[ethanndickson]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep them coming!

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".