safeImageURL and ValidURLRef refused only //, so a stored og_image of
/\evil.example/x.png loaded a third-party image in the admin. Both now
refuse /\ like models.validTarget already does.
refreshEntry set entry.url to the descriptor's address unconditionally, so
a slug change the job saw before any RecordSaved listener lost its 301 for
good. An entry whose URL differs now goes through AddressChanged (via
auto), which creates the 301 and moves the URL; an empty URL is filled.
A locale given with a URL write overrode the matched one unchecked, so
"xx" created an entry nothing ever reads while the answer reported
success. SetURLSEO (API, MCP, batch set_seo, records import) now requires
one of the record's locales and refuses others with api.locale_unknown.
The API, MCP, batches and the CSV import copied any value and Normalize
replaced it with the default, answering 201 for a 307 saved as a 301.
RedirectFields.Problems refuses values outside {301, 302, 410}, {exact,
pattern, regex} and {ignore, pass} with value_invalid; the import marks
such rows refused.
Only the admin seoimage field refused values that are not a path or an
http(s) URL; SetURLSEO, SaveRule and SaveSettings stored javascript:,
data: or protocol-relative values, which /resolve then served as
og:image. All three now refuse them with image.url_invalid.
(page-1)*perPage overflowed for a large page and sliced out of range.
Export compares the page with the page count before multiplying, and the
list endpoints cap the page number.
redirectHopsLogged was a never-evicted map keyed by the anonymous
request's URL with its query, so requests to a looping address with fresh
query strings grew it without limit and logged each one. The log is now
keyed by the cycle's rule ids, repeats at most once an hour and starts
over at 1024 keys.
A decoded path such as /%00 or /%FF reached Postgres as text and answered
500 with an error log line per request. parseTarget refuses a path or
query that is not valid UTF-8 or holds NUL (422), NotFoundPath does the
same for the logged address, and a referer is cleaned of both.
Two concurrent undos of one entry both passed the already-undone check and
both restored. Undo and UndoBatch now read the entries FOR UPDATE, and the
final mark is conditional on undone_at IS NULL, refusing with
already_undone when another undo got there first.
The save-time simulation read the rules at READ COMMITTED without a lock,
so A -> B and B -> A written concurrently both passed. CheckRedirect,
AddressChanged and every batch now take a transaction-level advisory lock
before they read the rules.
AddressChanged wrote its 301 and re-pointed rules without Problems or the
save-time simulation, and removed only an exact rule whose source equals
the new address literally. It now removes every exact rule that matches
the new address with its options, simulates the 301 and the re-pointing
with the stored rules and writes nothing when that loops, stores an
absolute old address on the site as its path, and skips an address that
is not a path on the site.
The resolver fetched exact rules with lower(source) IN keys, while the keys
drop one trailing slash and keep the query as written, so a source stored
as /old/ or /list?Page=2 was never read. A new migration adds the stored
generated column source_key (the SQL form of exactKey) with an index for
exact rules, and the fetch compares it.
Pure tests of the classes package (no database): the resolver's value
chains with every layer present and absent, robots per directive, the
staging lock and drafts; the redirect graph (chains, flattened and
back-and-forth graphs, loops by exact, pattern and regex probes, the
10-hop cap); URL normalisation across every address rule; the robots
grammar over every directive combination; Google's robots.txt matcher
examples; templates and unknown placeholders; CSV export and import;
owner resolution; and every kind of the issue report.
go.mod: gorm.io/driver/postgres becomes a direct requirement (the tests'
dry-run database).
Issues loads the entries, rules, redirects, robots row and noindex
addresses, then issuesOf builds the report and blockedNoindex checks the
addresses against robots.txt, so the report's rules can be unit-tested
without a database. No behaviour change.
The loop and chain path of a redirect notice appended the final target
when it was an absolute URL of the site itself, so a loop read
"/a → /b → /a → https://example.test/a". The site URL now counts as
internal, as it does when the chain is followed.
CheckRedirect loads the stored rules and hands them to
checkRedirectRules, which needs no database, so the loop and chain rules
can be unit-tested. No behaviour change.
- POST /_seo/api/v1/batch runs changes in one transaction: dry_run
answers the diff and the conflicts (loops, chains, noindex on a
published record, canonicals to redirected or noindex addresses) and
rolls back; a refused change refuses the batch (D-12)
- undo of a change or a whole batch from History and through MCP, with
the undo_stale, undo_loop and undo_gone refusals (D-13)
- CSV export and import of redirects, rules and records by URL: BOM,
comma or semicolon, formula cells escaped, a dry-run preview, all or
nothing with via import; the screen's apply token is HMAC-SHA256 over
the file and a fingerprint of the data (transfer.stale)
- Import and export singleton with the seo_transfer widget action,
History record actions undo and undo_batch, navigation entry
- MCP tools undo_change, undo_batch, apply_batch, export_csv, import_csv
- TestBatch, TestUndo, TestCSVTransfer, TestLoopRefused mcp and import
- /_seo/api/v1 routes for the effective SEO of a URL with sources and
owner, URL rules, redirects (loops refused with redirect_loop), the
404 log actions, robots.txt and its tester, settings and history
- the CMS-data issue report: missing and duplicate titles and
descriptions, rules over records, redirect chains, canonicals to
noindex or redirected addresses, noindex addresses robots.txt blocks
- 24 MCP tools on those routes (backend guard, golem15.seo.manage)
- AdminOperations maps every admin operation to its tool;
TestEveryAdminOperationHasTool and TestMCPReadToolsAreGET guard drift
- robots.txt writes go through classes.SaveRobots from the admin too
- PUT /_seo/api/v1/seo writes to the record behind a URL in its locale,
else to an exact URL rule (force_rule writes a rule); the answer says
where (D-11)
- management routes resolve the backend administrator from a token or
the MCP dispatch and require golem15.seo.manage (D-10, D-15)
- MCP tool set_url_seo; history records via mcp
- classes.WithTx, FindOwner, SetURLSEO, SaveRule and DeleteRule
- GET /_seo/api/v1/verification/{name} serves a verification file stored
in the SEO settings with its content type, nosniff and a sandboxing CSP
(D-05, D-103); the name pattern is enforced in the route and the
lookup, and no route is registered at the site root
- golem15.seo.robots singleton (UI-SPEC C10): staging callout and the
served file (P4); the plugin's own extension adds the editor, the
tester and the AI bots tab, saved to the robots row with history
- field type robotstxt (E4, live lint, size meter; Decode refuses over
500 KB) and widget golem15-seo-robotstester (E5) on robots_test, which
tests the unsaved text and switches through grobotstxt
- saving rules that block a noindex address answers robots.blocks_noindex
- github.com/jimsmart/grobotstxt v1.0.3 (D-99, decision note in the
framework repo)
- golem15_seo_robots singleton: rules and AI bot switches; eleven bots
incl. Claude-User, all allowed by default
- GET /robots.txt: rules, a Disallow group per switched-off bot and the
golem15.seo.robots_sitemaps lines; outside production the lock with
X-Robots-Tag; never a 5xx (last good body, else allow-all)
- TestRobots on Google's matcher with the deciding line, and the
noindex-blocked check (D-24)
- Normalize: https, www or bare host, double slashes and index files,
trailing slash add/remove, optional ASCII lowercasing (defaults change
nothing but scheme, slashes and index files)
- strict URLs through signpost MatchPath: a path that names a record
without being its address 301s to the record's address
- resolve runs normalisation, redirects, strict URLs, then SEO; the
location is the final address, so a visitor makes one hop
- P1 warns when a record's canonical URL redirects
- POST /_seo/api/v1/notfound with the golem15.seo.notfound_token bearer
(constant-time, off when unset), URL and referer caps, a row cap with
an hourly warning, throttle bucket seo-notfound
- golem15.seo.notfound list by hits with Resolved rows, a preview with the
resolved hint, a prompted 301 and a 410 record action, a bulk 410 that
refuses a selection with a covered address
- every /_seo/api/v1 answer carries X-Robots-Tag: noindex
- golem15.seo.redirects list and form (UI-SPEC C8): chain rows marked,
status, match, origin and chain filters, the navigation item lands here
- a field-less extension on the controller refuses a loop with
redirects.loop on target, answers the chain notice and records history
- the status dropdown posts text, read through RedirectStatus
- rule set: exact sources by key, patterns by literal prefix, regex by id
(stdlib RE2, compiled once per source); ignore case, trailing slash,
query ignore/pass, a source with ? matches its query exactly
- targets: captures escaped, a relative target never leaves the site, an
absolute one keeps its host
- SaveRedirect validates, simulates the rule in place (exact, pattern and
regex probes, 10 hops) and refuses a loop; chains answer the notice
- resolve follows the chain to its final target, 410 answers gone, a
runtime cycle answers not_found and is logged once
- golem15_seo_redirects table and Redirect model
- AddressChanged: old -> new 301 in the save transaction, earlier
redirects pointed at the new address, self-loops dropped
- the SEO tab save compares descriptor URLs with the entry URLs and
answers the redirects.auto_created notice
- resolve answers redirect with the absolute target for an exact source
