No cache read tokens used, run was expensive

[cknight]

First time user of gh-aw. I've finally got my fist workflow running, an automation which reads open dependabot alerts, does analysis on them and suggests remediation actions. I chose GPT5.4 model for my task. During the agent job, it iterated over 34 turns in approx 7 minutes to complete the session (successfully), but consumed 956 AIC. No turns used cached read tokens. Why didn't my session use any cache?

[pelikhan]

@cknight could you ask our agent or look in the run artifacts for the full log of firewall logs. You should be able to see the raw JSON message to make sure that this is not a miscounting where the cache token field was renamed for this particular model.

[pelikhan]

Adding a guardrail for this github/gh-aw-firewall#5195

[cknight]

Thanks for the quick response @pelikhan. The agent has identified that you are correct, the source of the issue appears to be an incorrectly mapped cached tokens field and my run did actually properly use cached tokens (3.3 million over 33 of the 34 turns) under the covers. It reports:

So, the bug is in the gh-aw token usage normalization/parser path, not in the model usage itself. Concrete fix: update parse_token_usage.cjs / the firewall API-proxy token normalizer to map OpenAI-style cached prompt tokens into cache reads, e.g. usage.prompt_tokens_details.cached_tokens -> cache_read_tokens when explicit cache_read_tokens is absent.

Note that this isn't just an annoyance, to work around this issue I have to set a very high max AIC usage amount since this limit appears to be based on the gh-aw reported AIC usage (without cached reads taken into the AIC used calculation), not the actual underlying model consumption.

[pelikhan]

We are working on a fix and getting cached metrics is very important. We are also adding a guardrail to detect catastrophic cache misses.

[pelikhan]

@lpcox see #39837 (comment)