- nginx and lighttpd allow single quoted strings, so use python's
built-in methods to escape quotes
- only apache requires parentheses on the outside (and only for obscure
reasons)
- caddy expects double quoted strings. at least escape inner quotes
directly (re.escape does not do this since python 3.7), should any
appear
Re-submission of #248, which was merged then reverted in 80c19fc. The revert was
caused by main.yml's commit-message handling, not by this data — that is fixed
separately in #254.
Six entries move from "Unclear at this time." to the operator's own documented
value, each citing the primary source inline:
Applebot Yes support.apple.com/en-us/119829#retrieval
Meta-ExternalAgent Yes developers.facebook.com/docs/sharing/webmasters/web-crawlers/
Meta-ExternalFetcher No (same Meta page — user-initiated fetch, documented
meta-externalfetcher No as not checking robots.txt)
DuckAssistBot Yes duckduckgo.com/duckduckgo-help-pages/results/duckassistbot/
Google-Agent Yes developers.google.com/search/docs/crawling-indexing/google-common-crawlers#google-agent
table-of-bot-metrics.md is regenerated and included, so `robots.py --convert` has
nothing left to write and the workflow's commit step is skipped entirely.
Verified: code/tests.py 13/13 · robots.py --convert exits 0 and is idempotent ·
robots.txt, .htaccess, the nginx/lighttpd/haproxy configs and the Caddyfile are
byte-identical · the JSON diff touches only the six `respect` fields, nothing else.
`git commit -m "${{ github.event.head_commit.message }}"` splices arbitrary text
straight into a double-quoted bash string. A message containing a double quote
closes the string early and the remainder is re-parsed as shell words.
This is what broke main after #248 merged. That PR's title contained
"Unclear at this time." (with quotes), so the merge commit message did too, and
the line bash actually ran was:
git commit -m "Fill "Unclear at this time." from primary operator docs ..."
which git received as:
-m "Fill Unclear" at this "time. from primary operator docs ..."
Hence `error: pathspec 'at' did not match any file(s) known to git`, a failed
run, and the revert in 80c19fc. The data in that PR was fine — robots.py
--convert exits 0 against it and code/tests.py passes 13/13.
The same interpolation is also a script-injection vector, which is the more
important reason to change it: a PR title is attacker-controlled, and
chore: tidy"; <any command>; echo "
executes that command on the runner with the workflow's token. `inputs.message`
has the same shape in the `if [ -n ... ]` test and its own `git commit -m`, so
all three are moved.
Passing through `env:` and quoting the shell variable is GitHub's documented
recommendation for untrusted values. The variable is expanded by bash after
parsing, so quotes, newlines and `$(...)` stay literal text.
Verified: code/tests.py 13/13 · python code/robots.py --convert exits 0 and
leaves robots.txt, table-of-bot-metrics.md and every server-config output
byte-identical · reproduced both the parse failure and the injection locally
against the old form, and confirmed the env form commits the same message
verbatim, quotes included.
Applebot, Meta-ExternalAgent, Meta-ExternalFetcher, meta-externalfetcher,
DuckAssistBot and Google-Agent carried 'Unclear at this time.' in both
operator and respect, but each operator documents the behaviour on its
own site. Values are linked to the primary source inline, matching the
style already used by Applebot-Extended and meta-externalagent.
robots.txt output is unchanged — no keys added or renamed.
