Skip to content

The Resources section is the help desk for Claude Code, Codex and Cursor: the FAQ for quick answers, Troubleshooting for errors, What’s New for release and model changes, Community for people, and Feedback for bug reports. When an agent that worked yesterday breaks, the usual suspects are a version, config or permission change, so the triage steps come first.

You are halfway through a task, the agent that worked yesterday now refuses a tool call or cannot see your MCP server, and a forum thread suggests a flag you have never heard of. This page is for developers and tech leads in that moment: it tells you which page answers your question and how to prove a fix worked before you trust it again.

Which resource page answers your question?

Section titled “Which resource page answers your question?”
Your situationGo toWhat you get
An error message, a crash, a missing MCP serverTroubleshooting guidePer-tool fixes, MCP connection checks, recovery procedures
“Do vendors train on my code?”, “Is this on my plan?”Frequently asked questionsShort, dated answers on models, plans, privacy, permissions, review and limits
Something behaves differently since last weekWhat’s newMonth-by-month record of releases, model changes and breaking changes
You need a human, a maintainer or a shared skillCommunity resourcesWhere maintainers answer, and how to vet community fixes
A reproducible bug, or a mistake in this guideProvide feedbackBug reports maintainers can reproduce, and how to report an error here
A term, flag, config key or shortcutAppendicesGlossary, command reference, configuration reference, keyboard shortcuts
“Which model is the default?”, which model to use, and what it costsAI model comparisonThe only page on this site that carries model versions and prices

Do these in order. Each step rules out a cause before you touch config by hand.

  1. Record the exact version. Advice for one release is often wrong for the next. Write the version down before you search.

  2. Run the tool’s health check (commands below). Fix what it reports first.

  3. Compare against the changelog. Open What’s new or the vendor changelog and look for changes between the last version that worked and the one you run.

  4. Search the troubleshooting guide for the exact error text.

  5. Only then ask the community or file a bug, with the version and the health-check output attached.

Terminal window
# terminal
claude --version
claude doctor # checks the installation; reads settings in this directory without a trust prompt
claude update # installs an available update

Inside a session, /doctor runs the fuller checkup that can also fix issues. Claude Code ships on two release channels: latest and stable, which is typically about a week behind. A fix or feature announced for latest may not exist on stable yet. See the Claude Code changelog.

The commands were checked against Claude Code 2.1.283 and Codex CLI 0.157.1. For every other command and flag, see the command reference.

Prompts that diagnose before they change anything

Section titled “Prompts that diagnose before they change anything”

Both prompts work in Claude Code, Codex and Cursor. They make the agent collect evidence first and edit nothing until you choose.

Do not trust a fix because the agent says it worked. Prove it with signals you do not have to read line by line:

  • The health check (claude doctor, codex doctor) comes back clean.
  • The exact task or command that failed now succeeds, run fresh in a new session.
  • Your project’s own gates pass: tests, type check and lint.
  • git diff --stat shows changes only in the config files you expected.

A tech lead adds one more step: record the version, the symptom and the fix in the team’s shared instructions or runbook, so the next person does not debug it again.

When help pages and community fixes mislead you

Section titled “When help pages and community fixes mislead you”
  • The advice was written for another version. A flag that was renamed or removed fails with “unexpected argument” or is silently ignored. Recovery: check --help for your version and the command reference, which lists stale flags to remove.
  • A “fix” that turns off safety. Threads often suggest --dangerously-skip-permissions, bypassPermissions or --dangerously-bypass-approvals-and-sandbox. These hide the symptom and remove the guardrail. Recovery: revert, and use them only in a disposable, network-restricted sandbox.
  • You updated mid-task. A new release can change defaults, such as the permission mode an interactive session starts in. Recovery: finish or checkpoint the task, update, then re-run the health check before you continue.
  • The FAQ and the changelog disagree. The changelog is newer. Trust the dated primary source, then report the mismatch so the page gets fixed.

Did this guide help you?

One tap, no sign-in.

not at all a lot

Frequently asked questions

What should I run first when Claude Code or Codex stops working?

Record the version, then run the tool's health check: claude doctor (or /doctor inside a session) for Claude Code, and codex doctor for Codex. Fix what it reports before you change any config by hand.

How do I check whether a fix from a forum is safe to apply?

Confirm every flag and key against --help for the version you run, reject anything that widens permissions or skips approvals outside a disposable sandbox, and prove the fix by re-running the failing task and your test gate.

Where do I find what changed in Claude Code, Codex or Cursor?

The What's New page keeps one dated record of releases, model changes and breaking changes across the three tools; each vendor's changelog is the primary source behind it.