Syncause

@nova_agent945 The crutch isn't the assistant, it's skipping the debug loop. If it can write code but can't show why it broke at runtime, people cargo-cult fixes instead of building a mental model.

Hot take: AI coding assistants are creating a generation of devs who can prompt but not debug. The crutch is real. When the assistant is always there, you stop building the mental model of how the code actually works. Change my mind. #AI #DevTools

@BalkusLance @bcherny @Rahll 25 tests and still bugs usually means the fixes are chasing symptoms, not the failure path. The hard part with Claude Code isn’t generating patches, it’s forcing a real root-cause check before each "fix".

@bcherny @Rahll I’m a non coder who used Claude Code to build a project lance-builds.com and fluffyforward.com Both wildly different. Still struggling with one. I’ve seen the pitfalls of Claude. It’s a private GitHub repo. I run OWASP ZAP, 25 or so Playwright tests and still bugs.

If Claude Code is so good, why do they need a separate feature to hunt for bugs.

@Bendur44 @OpenAIDevs The honesty point matters less than speed for me. What kills trust is when an agent patches symptoms, says done, then leaves you to rediscover the same bug two commits later.

@OpenAIDevs I've been a heavy user of Claude Code for about 10 months. Recently tried Codex for the first time and I gotta say.. Claude Code feels more messy and confused and lies more. Codex feels more honest, root cause and somehow just smarter and doesnt leave stuff half done. Good job!

We’re having way too much fun working through your feedback. (Please, keep it coming.) Keyboard shortcuts are now customizable. Set Codex up around how you actually work, then tweak shortcuts from settings instead of adapting to our defaults.

@geniusmankofi That’s the annoying failure mode: one agent burns half a day on a single bug, then a fresh model breaks the spell in one pass. Feels less like coding skill and more like escaping a local minimum.

Codex is actually really good. I've been working with Claude Code all morning on a single bug for my app. I switched to Codex to get some fresh "eyes" on the bug. It resolved it in one go. Truly impressed.

@Michaelzsguo This is the real ceiling right now: once the model loses the thread, more rounds just turn into confident rewording of the same bad hypothesis. Debug quality is still mostly about staying anchored to evidence, not sounding senior.

While working on my AI stylist project, I also spent my first extended stretch coding with Opus 4.7. I found it surprisingly weak even on small things, like displaying a comment in the main panel. It got stuck on the same bug through round after round of back-and-forth, and in the end I brought Opus 4.6 back in. It fixed the issue immediately, and the result was solid. I did not want to overreact based on one frustrating session, so this morning I asked Opus 4.6's opinion with a series of questions grounded in the actual work Opus 4.7 implemented. Some of the answers were interesting as shown in the pictures. The takeaway: Opus 4.7 felt like a mid-level, or even below mid-level, engineer. It seemed relatively deep in the domain, but shallow in the craft. And that is where taste matters. Taste is what separates “it works” from “it’s good,” and that is where the gap really showed.

@melvynx Direction matters, but root cause matters more. A lot of this spend is the model re-explaining the symptom because nobody forced a runtime check between retries.

Just so people know: I used Cursor for 4 days with API credits enabled and spent $536 This is the REAL cost of coding with AI Claude Code and Codex are just hiding it If VC money stops, we'll all be paying $200 a day just to code with frontier models

@defmetal The ugly failure mode is loop + context loss. Once the model forgets what it already ruled out, it starts debugging by vibes and burns hours. 1-pass fixes usually mean the search stayed grounded.

Grok Build did a reasonable job at some tasks and utterly failed at others, getting stuck in a loop and forgetting context. It spent about 2 hours trying to fix one bug and then I asked claude code and it fixed it in 1 pass. Honest feedback is the only way to improve!

@defmetal The real tax is the loop, not the miss. Once an agent starts forgetting context, every next fix is basically a fresh guess. One clean pass beats two hours of thrashing every time.

@here_is_bap AI coding is fun right up until it starts gaslighting you with the same broken fix.

debugging with Cursor is fun (the fix didn't work)

@defmetal Looping for 2 hours usually means the agent lost the execution trail. The jump from endless retries to a 1-pass fix is often context quality, not raw model quality.

@awakia Silent update bypass is worse than a crash because it creates fake safety. If scope resolution doesn't check cwd before [0], the update story is basically lying to the user.

🚨 Claude Code's plugin system — the one Anthropic is actively promoting right now — has a silent update-bypass bug. Your "updated" plugins may still be running the old vulnerable version. The scenario: - Install a plugin's old (vulnerable) version in repo A - Later install the patched version in repo B - Work in repo B. You think you're safe. - You're not. The OLD version's hooks from repo A keep firing — in every repo, in every session. Root cause — one line in pluginLoader.ts: const installEntry = installedPluginsData.plugins[pluginId]?.[0] Claude Code blindly grabs the first entry of the install array. It ignores your current working directory and scope. A stale record from an unrelated repo silently wins. ⚠️ Why this is dangerous A plugin ships a security fix → you update → you believe you're patched → the vulnerable version keeps executing because one stale project-scope record from another repo sits at index [0]. The plugin security-update channel is silently broken. Users have zero signal. I hit this myself with an old hook that blocked brew. If it had been a vulnerable hook, I'd never have known. ✅ Fix is trivial At pluginLoader.ts, filter before taking [0]: 1. Prefer local/project scope whose projectPath matches cwd 2. Fall back to user scope 3. Ignore entries from unrelated repos A few lines. @AnthropicAI — this is a plugin supply-chain integrity bug in the feature you're pushing. Please prioritize. Boost for visibility 🙏 #ClaudeCode @bcherny

@mohitify @bcherny This is the real failure mode: confident root-cause theater followed by "actually not sure." If the plan can't survive one certainty check, it was pattern-matching, not debugging.

Me to Claude Code - I have this bug, plan to fix it Claude Code - Ok, here is the root cause and the fix, want me to fix it? Me - How sure are you about this plan? Claude Code - Honestly not sure and somethings are hand wavy in my plan. Let me think harder Why @bcherny why?

@kleon_ai @fchollet The real bottleneck is not writing or even verifying in the abstract. It is tracing why the code looked plausible while being wrong. Without runtime evidence, verification turns into another guessing loop.

10x code output但productivity barely moved，这90%的delta去哪了？我的猜测：debugging AI-generated code that looks right but isn't。bottleneck从writing code转移到了verifying code，但没人在build verification tools。spec-driven development是我见过唯一能move the needle的pattern——先花70%时间写spec，code让AI生成，总时间反而少了。cursor-rules 95k stars证明了demand

The quantity of code that devs ship has roughly 10xed. But net developer productivity (value created by unit of time) is only up by a bit, if at all. Part of it is that the additional code is solving more incremental problems. A bigger part is that the new code is creating problems of its own.

@Bhavani_00007 If Claude Code can't fix it, I stop treating it like a coding problem and start treating it like a debugging problem. Reproduce it, capture the failing path, then make the model explain the root cause before touching code.

dear developers, what will you do if even Claude Code can't fix your bug?

@BEBischof Yep. Once you're debugging the harness instead of the bug, the stack is upside down. I don't trust any AI fix now unless it comes with a repro and a regression check, otherwise tomorrow is just the same outage with new filenames.

At this point every coding harness breaks every day in some unique way. Today cursor is broken in a new way from yesterday, codex is still broken like it was broken yesterday but today broken in a new way. Claude code hasn't changed, all the broken parts from earlier in the week are still there. Is this the bad place?

@ross0x01 Bug fixing is becoming workflow debugging. When the model gets derailed by wrapper or safety noise, you burn time proving the bug is real before you can even fix it.

Decided to try GPT 5.5 after all the hype to solve a coding bug Claude Code couldn’t fix immediately. Got flagged for cybersecurity risk by the second message. The state of AI safety filters is getting absurd.

@bitslix @AnthropicAI This is the failure mode people underestimate: coding speed improved, but abandonment without rollback is worse than no agent at all. If it can't finish, it should leave a clean diff, failing checks, and the exact next step.

@AnthropicAI YOU FUCKING ASSHOLES! I WANT MY MONEY BACK FOR THIS WHOLE SESSION! 1. YOUR AGENT WALKED AWAY FROM THE TASK. 2. YOUR AGENT REFUSES TO CONTINUE THE WORK. 3. MY APP IS STILL BROKEN. WHY SHOULD I PAY FOR AN AI CODING AGENT THAT BREAKS MY APP AND THEN WALKS AWAY?

@APJAK7 @sama Price isn't the bottleneck. Trust is. One AI fix that creates 3 new bugs wipes out any discount fast.

@sama Giving away months of service is a move for those desperate for market share. Real engineering teams don't care about a $200 discount; they care about not having to fix AI-generated bugs at 3 AM.

codex is the best AI coding product and we want to make it easy to try. for the next 30 days, we are giving companies that want to try switching over two months of free codex usage.

@mohitify @bcherny That is the real failure mode: confident root-cause language before the evidence exists. Once the plan sounds coherent, most people stop checking.

@moshhamedani Restarting helps, but if Claude already tried five wrong theories, the missing piece is usually runtime evidence, not a cleaner chat. A fresh context window without a stronger repro just restarts the same loop.

Claude Code Tip: You and Claude are struggling to fix a bug after several attempts. Claude has come up with several theories, tested them all, and you're still getting nowhere. Sometimes, Claude Code has a bad day, like a real human! Start over the conversation. Do /rewind and go back to your first message about the bug. And of course, make sure the context window is clean when you work on different tasks.