Too slow to complete any meaningful request in under 60 minutes

**Describe the bug**
The model spends the entire 60 minute timeout period in high latency read/analysis loops, reading 200 lines of files at a time with significant delays in between, never even making it to implementation. Is anybody else dealing with this? The action is unusable as a remote/background agent with this level of latency - of course I expect the Github VMs to encounter more latency but it's not usable as a background/remote agent as it is. I'd like to avoid increasing the timeout if possible.

**To Reproduce**

The prompt I used to produce this behavior:

> @claude Please implement the calculation review and test end-to-end. Use the other reviews that are part of the workflow as examples: the document review, and the quantity review. This review, the calculation review, is responsible for figuring out if the final costs make sense, based on the quantities and rates used to calculate it. So if a listed cost is $10k, this review should determine if the math used to reach it is correct, based on the original labor hours and the rate, which could be specified in the claim documents, contract, or not at all. Use the same categorization scheme as the quantity review: [valid/invalid/unclear], based on whether the calculations are correct, incorrect, or can't be determined based on the information available.

And here's how far claude-code got before timing out:

**Expected behavior**
Running the same prompt in claude-code locally completes successfully in about 10 minutes. I'd accept 2-3x for the GH Action VM latency, but barely starting in 60 minutes seems like something could be wrong.

**API Provider**

- [x] Anthropic First-Party API (default)
- [ ] AWS Bedrock
- [ ] GCP Vertex