No problem! It's an area I'm recently quite interested in but don't really have the background necessary to approach it.
I am also quite biased against any system that reduces the accountability of the care provider - and so much of modern application development is about being able to scale essentially everything _but_ accountability.
I like the idea of "ThoughtCoach" because you are framing it as empowering the user to improve their own abilities to manage their mental health. In this regard, it's kind of in the genre of mindfulness meditation (and the various apps/integrations that have shown up to help people with that). You could go with a gamified option like Noom does.
This requires you to scope out a progression for the user to go through - and to be informed (at least hopefully) by modern psychology and proven to work statistically before being implemented. This also requires you to significantly narrow the LLM's possible outputs/modes at each level - which alleviates many issues with overconfident outputs and unreliable patients (you can even force the user to choose from a list of words, rather than type anything at all).
I haven't checked this space out enough yet but it's rapidly evolving. A lot of chatGPT-based start-ups are in this learning-with-a-smart-tutor-available regime. I think it could be very powerful but a lot of clever engineering is still required. On the other hand, maybe the limits there prevent it from being nearly as useful as what you're imagining.
Food for thought.
edit: I guess my final thought is that, from personal experience, a good part of why therapy can be successful is that some amount of positive human interaction alone can be enough to improve one's general outlook. The context of "using an app" requires initiative and perseverance. The comfort from consistent human contact can inspire that but apps just feel like "thumbing through your phone". Reducing screentime in general is good for mental health. Thus, the app shouldn't be gamified really; because the only targets you can set in an app are LLM-based and will require the user to type or tap. Anything other than that is just what's already out there - Apple's Screen Time, step tracking, calorie counters, mindfulness guiding. Adding an LLM doesn't obviously help there.