Codex、API 或模型还有哪些“明明做得到却没做”的功能？

What it does:Reads the failing GitHub Actions output directly, identifies the root cause, commits the fix. Handles flaky imports, missing mocks, test ordering issues, lint rules, and environment variable mismatches. How to install: $skill-installer gh-fix-ci The real value: Most CI failures are the same five categories in different clothing. “gh-fix-ci” turns them into a background task. ## 4. Valyu. Web Search + Deep Research + Access to Specialised Sources The problem:Agents like OpenAI Codex struggle when tasks require real-world data. Paper search, academic research, GitHub search, economic indicators, or academic papers. Without integrations, they either hallucinate or defer. [...] Use Codex for: Terminal-heavy work (GPT-5.3-Codex leads Terminal-Bench 2.0 at 77.3% and GPT-5.4 at 75.1%, vs Opus 4.7 at 69.4% — Opus 4.7 narrowed the gap but Codex still leads) Background tasks. Fire a cloud task via Codex Cloud, review the PR an hour later Very very complex features. I h

In a 2025 study benchmarking LLMs on web API integration tasks, models hallucinated endpoint URLs up to 39% of the time and parameter names up to 31% of the time, and none of the open-source models tested solved more than 40% of the tasks. The same pattern shows up in dependencies: a 2025 USENIX Security study of 576,000 generated code samples found that 19.7% of the packages the models recommended did not exist. [...] One skill, 18+ agents: A single install gives Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and others the same context to research an API, write the integration, and test it. Tests against the real API: `nango dryrun` runs generated code against a real connection, returns the real response, and lets the agent fix real errors. This is the test loop most agents lack, and it is what closes the endpoint-hallucination gap. Deploys to a managed runtime: The same code runs unmodified in a tenant-isolated runtime with managed auth (OAuth, API keys, JWT, basic auth, MCP Au

### Codex capabilities in 2025: Multi-file code generation Code refactoring (including OpenAI GPT-5 Codex code refactoring) Legacy code transformation (Python2 → 3, AngularJS → React, etc.) High-quality documentation and tests Multi-agent workflows with tools Local CLI-powered coding sessions Repository-wide reasoning If you've seen "crisp code diffs," “intent-aware refactors,” or “explain & fix” flows - those are all powered by the Codex code generation model. ## How to Use Codex: includes all official methods There are now five main ways to work with Codex: 1. ChatGPT Codex inside the ChatGPT interface 2. Codex CLI for local file editing 3. Codex inside Cursor AI IDE 4. Codex API usage via the OpenAI API 5. Codex-powered GitHub review apps

provide comments and find critical bugs and issues before they get merged into production. You can also use it with slack mention. So you can integrate Codex into slack, at mention it and then it picks up the whole conversation thread that you've maybe been chatting with your team about and then implements it and gives you a PR. You can also build your own integrations with a Codex SDK. This allows you to run Codex programmatically in your own containers and so you can get structured output of code out of Codex and we'll cover this kind of in the the back half of the presentation here. So OpenAI at its core is a research company and Codex is backed by our state-of-the-art models, most recently GPT-5.1 Codex Max, which is our best model for agentic coding. So it's specifically trained in [...] give it this output schema param and just have it generate JSON. I'll fire this off here and you can see rather than going into an interactive terminal mode, Codex is just streaming kind of the ra

from being defined anywhere else except our canonical async utils package. And this is just a vibecoded eslint rule which we can assert is 100% covered because we can just do that which makes codeex exhaustively write tests for the positive and negative cases here. And this is kind of what it means to dive deeper into the systems thinking within your codebase. observe the mistakes or classes of misbehavior that the agents are making and do what it takes in order to just statically disallow them. Um, this has been really really cool because this allows us to actually encode what it means to have trusted engineering in the code. It's just not possible for AI slop to enter the codebase. And one other neat thing is that these things have stacked incredibly well, right? We have kind of [...] repository uh on my file system at the same time without having to clone the repo multiple times and like these commands are pretty fiddly. Uh I hate using them. So it's great that Codex app is just goi

### The preview iteration system One of the most interesting new features is the preview system. When you submit a task, Codex now generates 2-4 different implementation approaches and lets you pick which one to execute. This has been genuinely useful. For a recent API endpoint, the variations included: A minimal implementation focused on speed A more robust version with comprehensive error handling An approach that prioritized backwards compatibility A version optimized for future extensibility It feels like having multiple senior developers propose solutions, then picking the best approach for your specific context. ## What still frustrates me ### Model selection opacity [...] ## What still frustrates me ### Model selection opacity You still can't choose which model handles your task. Codex picks internally based on task complexity, repository size, and probably other factors I'm not privy to. As someone who understands the trade-offs between different model sizes and ca