When an AI coding task stops halfway

Open the project and review the changed files. If the task used Git, inspect the diff and latest commits. Check the last successful tool result and any side effects outside the repository: a sent email, a deployed build, a database change, or a created issue may need separate confirmation. Save the original prompt and the unfinished result. If you are unsure whether an external action completed, verify that service before asking the assistant to do it again.

What stopped	First check	Where to go next
What stoppedCodex says Selected model is at capacity	First checkConfirm the model and time; inspect local work before trying another model.	Where to go nextCapacity guide
What stoppedCodex says stream disconnected before completion	First checkCheck the last tool output and project diff; compare a short new request.	Where to go nextStream guide
What stoppedCodex returns HTTP 400 or Bad Request	First checkKeep the conversation ID, app version, and request ID; compare another client once.	Where to go nextHTTP 400 guide
What stoppedClaude Code loses connection mid-response	First checkFind the last completed tool action and inspect changed files.	Where to go nextConnection guide
What stoppedClaude API returns 529 overloaded_error	First checkCheck the response and official API component before retrying.	Where to go next529 guide
What stoppedClaude Code says OAuth session expired	First checkCheck which process needs sign-in; preserve work before reconnecting.	Where to go nextOAuth guide

These are routes for distinct symptoms, not a claim that Codex and Claude share an internal failure. If the error text is absent, write down the client, time, and last completed action rather than choosing a guide by guesswork.

A Codex desktop issue includes a detailed log of HTTP 200 and 400 requests around two affected chats. The reporter could not establish a deterministic reproduction and asked for an actionable server reason. The log is useful evidence that an old conversation and a replacement conversation can both fail; it does not prove that opening another chat repairs the task. A separate Windows report says PowerShell still worked while the app showed Bad Request. Product and client are therefore part of the report you should keep.

For Claude Code, one issue counted repeated truncated responses in that reporter's own unattended sessions. Another issue investigated network resets for its reporter. These cases do not make every incomplete Claude response a network failure. Compare your exact message and the affected component on Claude Status, and do the same on Codex status for Codex. A public status card reports known incidents; your local diff tells you what happened to your files.

Write a short handoff for the next attempt: what the original task was, which files changed, which tool actions completed, what you checked, and what is still unfinished. Ask the assistant to inspect the existing state first. For example: “The last response stopped after editing two files. Review the current diff, tell me what is complete, and continue only the remaining step. Do not repeat external actions without checking them.”

If a clean short request also fails, record the exact error and time, restart the client once, then compare the relevant official status source. If only one long conversation fails, keep that conversation and its logs for support. Do not delete it just to start over. If the task involved a public release or payment, verify the external result yourself before allowing any retry.

Will Retry run a command or send a message twice? It can if the original action already completed and the new instruction repeats it. Inspect the tool log and the external service first. When uncertain, ask for an inspection step without side effects.

Does no official incident mean the problem is my computer? No. Official status may not cover a particular account, client build, model, or conversation. Compare a short request and keep the exact error, client version, and time.

When should I open a new conversation? After you have saved the prompt and checked the work already done. A fresh conversation can isolate a broken thread, but it needs a concise, accurate handoff and does not restore hidden actions from the old one.

What if the assistant said the task finished but no file is available? Check whether a file was actually written and whether its link opens. For a ChatGPT file task, use the file-stage article; an agent's final sentence is not proof of delivery.

• First-hand Codex reports: macOS HTTP 400 with request logs; Windows app Bad Request.
• First-hand Claude Code reports: connection closed mid-response; network-reset investigation.
• Official status sources: OpenAI Status; Claude Status.

Is Codex down? · Is Claude down? · ChatGPT file problems