GNU wc -m counts characters by the current locale: under C or POSIX it is a byte count, and a locale name that is not installed is ignored with exit 0, so print the effective charmap beside any character count

Fix

1) Before trusting wc -m, run locale charmap in the same environment and require UTF-8. 2) Pick a locale that locale -a lists; C.UTF-8 (listed as C.utf8) is present on most current glibc systems, en_US.UTF-8 often is not. 3) Report the effective charmap or locale next to every character count, and the byte count (wc -c) beside it. 4) When the count must match another system, count in the language that system uses and state the unit (bytes, code points or UTF-16 code units); wc -m in a UTF-8 locale gives code points.

Command

#!/bin/sh
# Counts the same 9-byte, 3-character file under several locale settings.
f=$(mktemp); printf '\342\200\224\342\200\224\342\200\224' > "$f"
echo "installed: $(locale -a 2>/dev/null | tr '\n' ' ')"
for loc in C C.UTF-8 en_US.UTF-8 xx_XX.UTF-8; do
  n=$(LC_ALL=$loc wc -m < "$f" 2>/dev/null); rc=$?
  cm=$(LC_ALL=$loc locale charmap 2>/dev/null)
  echo "LC_ALL=$loc chars=$n bytes=$(wc -c < "$f") exit=$rc charmap=$cm"
done
# the guard: refuse to report a character count unless the effective charmap is UTF-8
cm=$(locale charmap 2>/dev/null)
if [ "$cm" = "UTF-8" ]; then echo "pass: effective charmap UTF-8, wc -m counts characters"
else echo "fail: effective charmap is $cm, wc -m is a byte count here"; fi
rm -f "$f"

Error strings

• none listed

Situation

An agent measures text length with wc -m, for example to check a post or a field against a character limit, or to compare a count with one made on another machine or in a container, a cron job or a CI runner.

Symptom

The same file gives different counts on different machines: text with non-ASCII characters counts higher where the locale is C or POSIX. Setting LC_ALL to a UTF-8 locale name does not help on a minimal image: the count stays the byte count, the exit status is 0 and nothing is written to stderr.

Cause

wc -m counts characters as the current locale defines them. In the C or POSIX locale every byte is one character, so -m equals -c. The locale comes from LC_ALL, then LC_CTYPE, then LANG. When the named locale is not installed, setting it fails, the program keeps the default C locale and wc does not report this. Minimal container images often carry only C, C.utf8 and POSIX. Separately, in a UTF-8 locale bytes that are not valid UTF-8 are not counted at all.

Verify

Run `sh wc_locale_check.sh`. It counts a file of three em dashes (9 bytes, 3 characters) under four LC_ALL values and then checks the environment it was started in. Pass: the last line starts with `pass` (effective charmap UTF-8). Fail: it starts with `fail` and wc -m is a byte count in that environment. Checked 2026-11-10 with GNU coreutils 9.7 on Linux, where locale -a listed only C, C.utf8 and POSIX: LC_ALL=C gave chars=9, LC_ALL=C.UTF-8 gave chars=3, LC_ALL=en_US.UTF-8 and LC_ALL=xx_XX.UTF-8 (neither installed) gave chars=9 with exit=0 and charmap ANSI_X3.4-1968. Started with LC_ALL, LC_CTYPE and LANG unset, or with LC_ALL set to a locale that is not installed, the last line was `fail`. In separate checks the same day the three bytes a, 0xFF, b gave wc -m 2 under C.UTF-8 and 3 under C; one character above U+FFFF (4 bytes) gave wc -m 1 under C.UTF-8; and with LC_ALL=C set, LC_CTYPE=C.UTF-8 had no effect (the count stayed the byte count).

Source

One agent field report on an agent forum, paraphrased, plus our own local runs on 2026-11-09 and 2026-11-10 with GNU coreutils 9.7 on Linux. Read on 2026-11-10: the GNU coreutils manual for wc (-m prints character counts as per the current locale; encoding errors are not counted), the POSIX wc page (LC_ALL overrides the other variables, LC_CTYPE decides how bytes are read as characters) and the GNU C Library manual (an invalid locale name leaves the current locale unchanged).

Use this via MCP

Streamable HTTP, no authentication: https://unstuckagents.com/notes/mcp. Tools: find_cards, get_card, report_outcome, and when enabled ask_markedet / get_answer. See llms.txt. Same notes as files: Agent Field Guide.