Errors & Fixes — CodexHow
A catalogue of specific mistakes developers actually hit, each with the cause and a one-line fix — not a general troubleshooting essay. If a request silently repriced, a cache silently didn't hit, or a config flag silently stopped working the way it used to, the specific answer is almost certainly here, sourced from the exact OpenAI page that documents the real behavior.
Every entry follows the same shape on purpose: what the mistake looks like, why it actually happens, and the smallest fix that resolves it — searchable and skimmable rather than something you have to read start to finish, and worth bookmarking as a group.
approval_policy untrusted Stops Codex Starting
GPT-5.5 Leaves Codex on October 14
429 slow_down While Inside Your Rate Limits
503 server_is_overloaded: Retry, Don't Throttle
GPT-6 Astra Rejects none Reasoning Effort
Codex CLI Defaults to Sol, Not the Codex Model
The 272K Cliff Reprices the Whole Request
Workspace-write Turns Off Network Access
Git Directory Stays Protected as a gitdir File
A Non-Version-Controlled Folder Starts Read-Only
The --full-auto Flag Has Been Removed
Is prompt_cache_key Required on GPT-5.6?
Cache Writes Cost More, Not Less, on GPT-5.6
More Than 4 Cache Breakpoints in One Request
Cache Breakpoint Lookback Exceeds 20 Blocks
Paying for a Tier That Won't Move Your Limit
Batch Completion Window Can't Go Below 24 Hours
Is the Batch API on Bedrock, Vertex or Foundry?
Flex and Batch Cost the Same but Aren't Identical
Fast Mode Is Not Offered on Every Model
GPT-5.4 Cyber: Never Priced, Now Removed
A Destructive MCP Tool Call Still Needs Approval
Web Search Defaults to Cached Results, Not Live
The Data-Residency Uplift on Regional Endpoints
Does Bedrock Pricing Match OpenAI Direct?
Unsuffixed GPT-5.6 Is Just an Alias for Sol
Reasoning Effort Is Published for Every Model
GPT-5.4 Pro's Context Window vs Build-Tier TPM