[Bug] Claude Code presents unverified hypotheses as diagnoses without web search verification

Bug Description
Subject: Claude Code repeatedly presented unverified guesses as diagnoses, costing real build and deploy cycles I'm a student building an iOS fasting app with CloudKit for a project that's due soon. I used Claude Code for a long debugging session on CloudKit push notification subscriptions. What happened The app hit a CloudKit error: "attempting to create a subscription in a production container". Over roughly six rounds, Claude offered a series of confident explanations, each presented as the diagnosis rather than as a hypothesis, and each requiring me to spend a real cycle to test: 1. Said the cause was an OR predicate in the subscription. I changed the code, rebuilt, uploaded to TestFlight. Wrong. 2. Told me to run a Debug build so subscriptions would be created in Development, then deploy the schema. I did. The deploy diff came back empty. 3. When the diff was empty, Claude concluded its own theory was disproven and said "subscriptions aren't schema, that path was a dead end." 4. Only after I told it that we were going in circles and that it was wasting my time and money did it search the web for the exact error string. The first search result gave the documented cause and fix immediately: subscriptions cannot be created at runtime in a production container, they must be created in Development and promoted by deploying the schema, and this works even though the diff shows no subscription changes. So the answer was one search away the entire time, and the "empty diff" that Claude treated as disproof was in fact the documented expected behaviour. The cost Each wrong theory meant a new build, a TestFlight upload, and waiting for processing. I eventually hit App Store Connect's daily upload limit (error 90382, "Please wait 1 day and try again"), so I now cannot submit my project for testing at all today. That is a hard external limit I cannot work around, on a deadline. I had also raised this same pattern with Claude in a previous session. It happened again anyway, because nothing about it persisted between sessions. What I think should change - When Claude has a tool that can verify a factual claim, like web search for an exact error string, it should use it before proposing a fix, not after the user complains. - Claude should distinguish clearly between "this is the cause" and "this is my best guess, here is what would confirm it." It consistently used the first framing for the second kind of claim. - Claude should account for the real cost of the actions it asks for. Telling a user to rebuild and re-upload is not free. There are rate limits, deadlines, and money involved, and it should weigh that before proposing a test. - Feedback about working style should persist across sessions by default, rather than depending on the user repeating themselves.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.233
• Feedback ID: b85ab6b4-e2cf-4914-a441-97f443555c6f

Errors

[]