API proxy token normalizer does not map OpenAI cached_tokens to cache_read_tokens

[lpcox]

Problem

The AWF API proxy / token usage parser does not correctly map OpenAI-style cached prompt tokens into the cache read tokens field. When OpenAI models report cached token usage via usage.prompt_tokens_details.cached_tokens, AWF fails to recognize these as cache reads.

Impact

• Token usage reports show 0 cache reads even when the underlying model heavily uses caching (e.g., 3.3M cached tokens over 33 turns)
• AI credits guard is affected: the max-ai-credits limit is calculated without accounting for cached reads, causing the reported AIC usage to be much higher than actual consumption. Users must set artificially high limits to avoid being rate-limited.

Expected Behavior

When usage.prompt_tokens_details.cached_tokens is present and explicit cache_read_tokens is absent, the normalizer should map cached_tokens → cache_read_tokens.

Root Cause

The token usage normalization path (in the API proxy token tracker / parse_token_usage.cjs) does not handle the OpenAI-specific prompt_tokens_details.cached_tokens field.

Reference

• Upstream issue: No cache read tokens used, run was expensive gh-aw#39837 (comment)

[github-actions]

👋 Hello! This issue appears to be closely related to a previously filed issue:

• [awf] api-proxy: cache_read_tokens reported as 0 despite cache hits in model usage metadata #4310 — [awf] api-proxy: cache_read_tokens reported as 0 despite cache hits in model usage metadata (closed as completed on 2026-06-04)

Both issues describe the same root cause: prompt_tokens_details.cached_tokens from the OpenAI Responses API wire format is not being mapped to cache_read_tokens in the API proxy token-usage parser, resulting in cache reads being reported as 0.

Issue #4310 was closed as completed — this may indicate a regression, or that the prior fix did not fully cover the parse_token_usage.cjs path described here. It may be worth reviewing the changes made for #4310 to determine whether they address this scenario.

This is an automated message from the issue duplication detector.

Generated by Issue Duplication Detector for issue #5203 · ◷