yo, i feel u on this. after coding for way too many years, i've seen tools come and go, but the current state of ai is wild. honestly, for those nasty async race conditions and pydantic v2 schemas, Claude 3.5 Sonnet is absolutely the king right now. it's way less 'lazy' than gpt-4o and actually follows complex logic across files without hallucinating every two seconds. if your looking for that massive context window though, Gemini 1.5 Pro is pretty much unbeatable. i use it through Google AI Studio when i need to dump an entire fastapi codebase in to find a bug. also, if ur tired of cursor, check out Sourcegraph Cody—its context awareness is seriously top-tier for repo-wide stuff. basically, sonnet for logic and gemini for context is the winning combo imo. it really makes a difference when ur 15 prompts deep and dont wanna start over. gl with the project! peace

oh man, i feel u. honestly, i’ve had some pretty frustrating experiences with OpenAI GPT-4o lately. it’s been failing on complex refactors and just making up methods that don't exist, which is pretty disappointing when you're trying to move fast. for your situation, i would suggest these specific tools, though you still gotta be cautious about what they spit out: * Cursor Code Editor - this is basically the best way to keep the 'thread' alive. it indexes your whole repo so the Claude 3.5 Sonnet model actually knows the context of your other files.
* Claude 3.5 Sonnet - this is highkey the current king for python type hinting and pydantic. its reasoning is much more reliable than gpt's right now, i think.
* Gemini 1.5 Pro - i only use this in Google AI Studio for massive context needs, but it can be hit or miss for actual logic compared to claude. tbh, i'm still wary about using any of these for deep architectural stuff without double-checking everything. i had issues with Claude 3.5 Sonnet suggesting unsafe async patterns for database sessions before, and it wasn't as good as expected. it's still just a tool, so maybe check the official docs for those race conditions to see what it's reaching for. gl!

Similar situation here - i went through this last year when migrating a massive backend to Pydantic v2. Honestly, when ur dealing with complex async logic and session decorators, safety is highkey the biggest concern. I noticed that most models would hallucinate "thread-safe" patterns that actually caused deadlocks in my FastAPI sessions cuz they didnt respect the event loop correctly. Sooo, my workflow basically became a safety-first sandwich:
1. Use Claude 3.5 Sonnet specifically for the architectural refactoring cuz its reasoning feels more "grounded" than others.
2. Strictly pipe everything through Pyright and mypy 1.11.2 to catch the AI's creative interpretations of Pydantic schemas.
3. Never trust its async suggestions without a manual trace of the `await` points. It’s definitely a bit of a chore compared to the old "one-prompt" days, but i found that treating the AI as a brilliant but reckless intern is the only way to keep the codebase stable. Be careful with those custom decorators too, they're usually where the AI trips up the most! Peace.

Ok so, for your situation, I've actually found a way to handle those nasty Pydantic v2.0 schemas and async logic without breaking the bank. If you want the best ROI, honestly, I'd suggest skipping the $20/mo subscriptions and using Continue.dev (it's open-source!!) inside VS Code. You can plug in the DeepSeek-V2 API which is literally pennies compared to the big players but it punches way above its weight class for Python refactoring. For the massive context you need, check out:
* Google Gemini 1.5 Pro via Google AI Studio - The free tier is amazing for that 2M context window. * Claude 3.5 Sonnet via the Anthropic API - Use pay-as-you-go for those long sessions so you dont pay for what you dont use. It's sooo much cheaper than a fixed seat if youre just one person. Plus, DeepSeek-Coder-V2-Instruct is highkey one of the best models for complex logic right now. Seriously, the value you get for like $2 total a month is fantastic! gl with the async stuff!

Regarding what #5 said about "Saw this earlier but just getting a chance..."

• I totally agree that mapping the repo is the way to go. I just found this thread and wanted to add that you gotta be careful with those async database sessions. Even the best models will try to use old Pydantic v1 syntax or ignore the event loop if you don't keep them on a short leash. I would suggest checking out Double.bot Extension for VS Code if you want something that feels a bit more robust for long debugging sessions than the standard extensions. But honestly, if you want a real DIY fix, just start keeping a scratchpad file in your repo that lists your current architecture and session patterns. I feed that to the AI every few prompts and it stops the hallucinations cold. Also, give Phind-70B Model a look through their web interface... it's surprisingly good at reasoning through multi-threaded Python stuff without getting as lazy as GPT-4 does these days. Just make sure to double-check those session closes, these bots still love to leak connections if you let them.

Saw this earlier but just getting a chance to chime in. I've been through the same ringer with async database sessions... honestly a nightmare when the AI loses the script. Quick question tho: are you using a monolithic structure or are these microservices? It makes a huge difference for how the AI handles the dependency graph. Two quick tips if you're hitting walls:

• Try Aider AI coding assistant. It works via terminal and actually maps your whole repo, so it doesnt hallucinate file structures as much as a chat window.
• If you want a model that's tuned specifically for dev work without the ChatGPT fluff, check out Phind-70B Model. I've found Phind much better at sticking to the docs for Pydantic v2 than GPT-4o lately.