Claude Code Refuses to Edit GBK Files

Because Edit and NotebookEdit save the whole file back as UTF-8. The tools reference states the reason for the refusal: The tools refuse because they save the whole file back as UTF-8, which would replace every byte they can't decode with the replacement character U+FFFD. Before 2.1.296 that replacement is exactly what happened, so one edit filled a GBK or Shift-JIS file with replacement characters. Since 2.1.296 those edits are refused: the changelog says Fixed Edit and NotebookEdit replacing every non-ASCII character in files that are not valid UTF-8 (Windows-1252, Shift-JIS, GBK); such edits are now refused, and the error reference adds Nothing was written, so the file is as it was. How the characters look on screen is up to your editor; the character the docs name is U+FFFD.

Use the generic file command. The file(1) manual says -i causes the file command to output mime type strings rather than the more traditional human readable ones, its example being text/plain; charset=us-ascii, and that --mime-type and --mime-encoding print only the specified element. Treat the answer as a hint: file classifies text by byte ranges and sequences, and the manual notes some encoding logic is hard-coded in encoding.c, so confirm by converting with iconv and reading the text back. The six Claude Code documentation pages this article consulted on 2026-10-13 (settings, troubleshooting, common-workflows, checkpointing, terminal-config and overview) mention no command that prints or sets a file's encoding.

Back up first, convert second, compare third. The iconv(1) synopsis is iconv [ options ] [ -f from-encoding ] [ -t to-encoding ]: -f names the encoding of the input characters and -t the encoding of the output, and the manual's own example is iconv -f ISO-8859-15 -t UTF-8 < input.txt > output.txt, so a GBK file becomes -f GBK. Remember what the manual says about failures: unconvertible characters terminate the run unless you pass -c, //TRANSLIT approximates them and prints a question mark when it cannot, and //IGNORE discards them with an error after conversion. All of them lose characters, so keep the copy and check the Chinese or Japanese passages. These are standard Unix tool usages, not Claude Code parameters.

Same rule, same set. The 2.1.296 sentence names Windows-1252, Shift-JIS and GBK together; the error reference names the same three, and the tools reference gives Windows-1252 and Shift-JIS as examples. The check is whether the file's bytes decode as UTF-8 as a whole, so a file that should be UTF-8 triggers the same message when it holds even one invalid byte sequence. For conversion, use the names from the GNU libiconv encoding list, which includes SHIFT_JIS, CP932, GBK, CP936, GB18030, BIG5 and the CP1250 to CP1257 family, with UTF-8 as the target.

Two variables. QCode's Claude Code tutorial gives export ANTHROPIC_BASE_URL = https://api.qcode.cc/api and export ANTHROPIC_AUTH_TOKEN = cr_ followed by your key, and the endpoint table in QCode's documentation gives Claude Code the same value https://api.qcode.cc/api; the ANTHROPIC_BASE_URL setup guide on qcode.cc adds that the base URL takes no trailing slash and that Claude models go only to the Anthropic endpoint. As of 2026-10-13 claude-sonnet-5-5, claude-haiku-4-5-20251001 and claude-opus-5-5 are available on QCode. Billed per token; per-model rates are listed on /models.

Alipay and WeChat Pay both work. QCode's payment-methods page lists Alipay, WeChat Pay, Credit / Debit Card (Visa / Mastercard / AMEX), Apple Pay / Google Pay and Cryptocurrency (many coins and networks, with manual reconciliation of the on-chain transfer) on the top-up page; billing is prepaid — you top up, then buy a plan that draws down the balance. The same page states the limits it prints: CNY top-ups start at ¥100 and USD top-ups run from $1 to $2,000. After registering, create a cr_ key in the dashboard and point Claude Code at it.