[Bug] Claude Code loses context after API failures and denies error occurrence

**Bug Description**
over the past week, it has been increasingly more and more impossible for me to do regular development work with the FABL model. I am creating a simple workflow in Jira and trying to harden that workflow so that it is usable and nonbrittle and I continually get a cyber failure on Fable. It makes no sense. There is no detail. There is no understanding of why, what. And then on top of that, once it fails, it loses all capability of understanding, one, that it did fail. Two, why it failed. So then I have to spend tokens going back through everything that happened so that it can try to piece together once again what happened, why it happened, and what to do about it. And on top of that, it denies that anything ever happened at all, which means now I'm in a state where I'm spending an hour, especially when I'm working and sending messages back and forth between multiple sessions, an hour trying to repiece together everything that happened that led to the failure, which should never have happened in the first place. The guardrails are so loose that it becomes goddamn impossible to do anything that lasts longer than a few hours before the session gets so steeped in what it's doing that it triggers all these failures over and over and over again. Why the hell would I even use Fable at this point? I get me deep in a project, and then it starts blowing up on my face.

**Environment Info**
- Platform: darwin
- Terminal: iTerm.app
- Version: 2.1.238
- Feedback ID: 08475679-b783-45ab-8e0b-a61283ed7dc1

**Errors**
```json
[]
```

Start by reviewing feedback ID 08475679-b783-45ab-8e0b-a61283ed7dc1 and reproducing the reported Jira workflow on darwin in iTerm.app. Compare the session behavior around the API failure; done should include a captured error and preserved context that acknowledges the failure and explains its cause.

索引モデルが issue の本文から書いたものです。