Comments and test titles added on this branch cited review labels and
planning documents that do not exist in the repository, and some used
dashes as separators. They now say what the code does and why in plain
sentences, and the Postgres lane's suites share one header format. Test
ids stay as they were.
The boot refusal beside a SQLite install told the operator to run the move
with plain `docker compose run`, which uses the SQLite compose file. It now
gives the command the wiki gives, and PGBOOT-037 holds both to the page.
The move read unique indexes by joining indkey to pg_attribute. An expression key such as lower(email) has attnum 0 there, so idx_users_email_lower was missing from the plan. The catalog query now reads each index's key as Postgres prints it, whether any part of it is an expression, and the columns it reads: the plain key columns and the Var nodes of indexprs, without INCLUDE columns or the predicate.
A broken unique key no longer names --drop-nul as its cause. The refusal names the key and the shared value from Postgres's detail. It explains that the SQLite file held the two rows apart, and the hint about SQLite's lower() appears only for an expression key. Values --drop-nul changed in the key's columns are listed as rows to check first, and only when the duplicate value holds their text. An unrelated value that lost a NUL is no longer named. A plain key over a quoted column is named by its column, not as an expression.
A row that breaks a constraint (SQLSTATE class 23) is now a MoveRefusal that says what to correct, in the SQLite file or in a plugin's plugin.db. Before, it was a bare error with a stack. A value Postgres refuses by its type (class 22) stays a failure: it names no row, and values.ts should have refused it first, so the message asks for a bug report. MoveRefusal and the remedy text move to refusal.ts so copy.ts can use them without an import cycle. The usage text of --drop-nul and the wiki's move section describe NUL in JSON, the key listing and these refusals.
A raw U+0000 in TREK's JSON text, which an old or hand-edited travel.db can hold, was refused as "not valid JSON" without naming the NUL or the remedy. It is now refused with the NUL reason, so the refusal names --drop-nul. With --drop-nul, raw and escaped NULs are removed and nothing else, the way TREK's own writes drop them, and each change is listed. Text that is not valid JSON even without its NULs is refused as that, and the checksum's source side follows the same rule.
A \u escape of half a surrogate pair parses in JavaScript but jsonb refuses it. Before, that failed the insert without naming a row. It is now refused by table, row and column.
The plugin SQL guard now refuses what still reached outside a plugin's
schema or depended on the server:
- reads of the system columns (xmin, xmax, cmin, cmax, ctid, tableoid),
which show transaction ids and storage layout and do not exist on
SQLite; a value of the plugin's own may still be named xmin with AS
- the xid and xid8 types wherever the word stands for the type, since
age() of an xid is the cluster's transaction counter
- the text search functions that coerce a text argument to regconfig, a
catalog lookup by name, which leave the allowlist
- collations and access methods outside the set every server has, which
could tell whether a locale or an extension exists
Every plugin transaction pins, with SET LOCAL, the settings that decide
how values print and parse (TimeZone UTC, DateStyle, the C locales, float,
bytea and xml output, the text search configuration), so a plugin's output
neither depends on nor reveals the server's configuration.
Both stores now keep one contract. Its texts and caps live once, in
plugin-data-contract.ts, and both stores throw only those: NUL in SQL,
an argument or a migration id is refused on SQLite as on Postgres,
transaction control is refused in every call on both engines (tx() is
how a batch becomes atomic), a closed store refuses every call, and the
Postgres store answers the cases better-sqlite3 names with its words.
The SQLite store moves beside the Postgres one as SqlitePluginDataStore
in db/engine/sqlite, takes its file from the backend so src/db never
imports src/nest, and no longer needs the driver exemption. A contract
test runs every violation on both stores and fails when they answer
differently. The Postgres lane holds the allowlist, the pinned settings
and the SQL plugins commonly write against a real database.
ROUTEID walks every route the app registers.
- ROUTEID-001 sends 'abc' and the non-canonical spellings of a seeded row's id to each path param named like an id, and requires no 5xx and the answer 'abc' gets. It then sends the spellings of an id no row has and records what the database drivers are handed (tests/helpers/sql-capture.ts, below MikroORM and Kysely). It fails when a spelling, or the number Number() makes of it, reaches SQL, so a route over a table the suite seeds nothing in is held too. Stale NOT_ROW_IDS entries fail, and the walk runs on one listening keep-alive server.
- ROUTEID-002 walks every @trek/shared request, body and query schema and every route's body and query DTO, and fails on a field named like a row id that accepts such a spelling. String ids are listed with their reason; open fields name the handler test that holds them.
Three more suites cover the other entry points:
- MCPROWID does the same for every MCP resource template.
- RPCROWID refuses a plugin RPC id param read past rowIdParam.
- BODYID, JRNID and QUERYID cover the open accommodation update body, the collab reply, the journey fields a handler parses itself and the notifications limit and offset.
All suites that boot the app run in the SQLite suite and in the Postgres lane on 14 and 18. server/CLAUDE.md states the rule and the four gates.
Eleven MCP tool files carried their own parseId, which used Number() and so took '0x1' and '1e20'; they now share uriRowId. The websocket join and book:join read the id with toRowId instead of Number().
Plugin RPC ids go through rowIdParam instead of num(), and a param that may be left out goes through the new optionalRowIdParam. num() let '0x5', 1.5 and 1e20 through, which Postgres refuses or reads its own way while SQLite matches nothing.
accommodations.create read its refs with Math.trunc(Number()), and accommodations.update bound its open input's refs as sent. collab.createMessage took replyTo only when typeof was 'number', which let 1.5 and 1e20 through. All three now read their ids as row ids. A canonical numeric string still passes, as the published contract says, and the refusal keeps num()'s message.
The services and controllers that took an id from a request with Number(), parseInt() or not at all now use toRowId. That covers admin users, notifications (and their limit and offset), trip members and guests, day notes, assignments, memories (photo owner, album link, body photo_id), photos, mail-ingest, the assistant id parsers, dawarich, registration invites, the plugin contribution routes and the collab reply.
The accommodation update body is open, so AccommodationsService reads the place and day refs of every stay write with toRowId and answers a refused one with the create schema's 400.
AdminService.deleteUser checked the caller with parseInt and deleted with Number(), so an admin could delete their own account by spelling its id in hex; both now use one parse.
The OAuth session revoke called Number(id) first and so revoked session 5 for '0x5'; the service now reads the id as sent.
Request fields that hold a row id took z.union([z.number(), z.string()]), a bare z.string() or z.array(z.number()), so '0x5', '5e0', 1.5 and 1e20 reached the handler and from there SQL, where SQLite matches nothing, Postgres 14 and 15 refuse the cast and 16 and later read hex and underscores as a number. Two primitives join idParamSchema: rowIdInputSchema takes a safe integer or its digits and keeps the value as sent, so field types do not change; rowIdStringSchema takes digits only, for multipart fields. Both refuse with 'Invalid id'. The accommodation, assignment, reservation position, admin invite, collab reply, vacay, file and memories photo ids now use them, and every list of ids is z.array(z.number().int()). idParamSchema takes only the decimal spelling or a whole number, and the trip-journal-link and tour library ids use it. A body id spelled '0x5' is now a 400 on SQLite too.
RowIdPipe (src/nest/common/row-id.pipe.ts) parses a path param with toRowId. It answers a refused spelling with the route's own not-found answer, so '0x5', '5e0' and 'abc' get what an id that names no row gets. ParseRowIdPipe replaces Nest's ParseIntPipe with the same 400 and message: ParseIntPipe let '-1' and numbers past 2^53 through, and Postgres refuses the latter as out of range. Journey, journey recordings, day import, book design, map, public photo, collections, doc-sync, vacay, atlas and the plugin user-data export declare their ids through it. The upload routes parse in the handler with withRowId, so a refused id takes the existing cleanup path, or carry SpoolCleanupInterceptor (collections).
toRowId moves to src/db/row-id.ts so repositories can use it (src/db sits below src/nest), together with a new withRowId helper; src/nest/common/row-id.ts re-exports both. Its docstring is rewritten to state the rule plainly. TripsRepository read its number | string trip ids raw: SQLite matched '0x1' against nothing, Postgres 14 and 15 refused the cast with a 500, and 16 and later read it as trip 1. Its access reads, DaysRepository.existsInTrip and the packing template lookups now parse the id first, so an id toRowId refuses is no row on every engine. TripAccessGuard, TripOwnerGuard and JourneyAssistantGuard used Number(), which read '0x1' and '1e0' as row 1; they now use toRowId.
An attack on the move to Postgres with real SQLite files found five values
that changed without a refusal or a listing, or a valid file that failed:
- Negative zero in a plugin column became 0: pg sends the number -0 as "0",
and the checksum printed both as "0". It is sent as text and kept apart in
the checksum on both sides.
- Text whose bytes are not valid UTF-8 was read with each bad sequence
replaced by U+FFFD, and the checksum was built from the replaced text.
Text is now read as its bytes and decoded strictly in the file's encoding
(UTF-8, or UTF-16 for a file created so); such a value is refused in
TREK's and in plugin data, with --drop-nul too, and listed with its bytes.
- --drop-nul serialized a JSON value again, so a large number lost digits
and 1.0 became 1. Only the \u0000 escapes are removed now.
- A plugin default Postgres refuses for the column (DEFAULT 0 on a BLOB
column) failed the move, and DEFAULT 1.5 on an integer column would have
been rounded. Such a default is left off and listed as not carried over.
- A plugin index or primary key SQLite compared with COLLATE NOCASE or RTRIM
became one that compares exactly. It is still created, and the summary
lists the collation it lost.
Each case has a regression test that fails without its fix, on SQLite and
on the Postgres lane.
Postgres reads x.f as the call f(x) when x has no column f. The guard counted
that spelling only after ")." or "].", so over a function in FROM that
returns a single value, or a parameter, the allowlist was bypassed:
"select g.obj_description from generate_series(...) g" returned the comments
on every object of the database, and g.current_schemas, g.txid_status and
$1.ts_token_type passed as well. A refused catalog function after a dot is
now refused too, except the names in COLUMN_NAMES (money, name, oid, version,
xid, xml), which take no argument or only convert it to their type, so t.name
keeps working. PGPLUG-021 holds that list to pg_proc on the lane's server.
The type aclitem names roles: '=r'::aclitem printed the cluster's first
superuser and 'x=r/x'::aclitem told whether a role exists. It is refused in
every spelling like the reg* types.
The allowlist itself moves to assets/postgres/plugin-functions.txt, beside
the catalog list it is checked against. As TypeScript it spelled 'like' and
'json_each', names of Postgres functions that the SQLite dialect gate reads
as SQLite SQL.
The booking body is not schema-checked, so day_id, end_day_id, place_id,
assignment_id, accommodation_id and the create_accommodation ids can arrive
as any string. The reference guards passed them to SQL as they came. SQLite
compares '0x5', '5e0' or 'abc' against an integer id and matches nothing
(or, for '5e0', the row), so the guard answered "Unknown reference". Postgres
refuses the cast on 14 and 15 and reads hex as a number from 16 on, so the
same body answered a 500 on one major and reached a real row on another;
the Postgres 14 lane caught it in L2.
Both guards now read each id through toRowId, the rule every other untrusted
row id follows: an id it refuses resolves to nothing, so referencesOutsideTrip
does not name it and unresolvedReferences does, and the answer is the same
400 on every engine. create() reads the payload's accommodation_id through
toRowId as update() already did, instead of storing '0x5.0', which the
reads' join could not cast back on Postgres.
ReservationRow and ReservationTraveler move to reservation-rows.ts, the way
reservation-inputs.ts holds the input shapes, so the service stays within
its size baseline; the service exports them on and it is lowered to 1166.
The fakes that answer safeFetch by API took their init as any; safeFetch
declares it as RequestInit, which is what they read (the body as
URLSearchParams or text). The SSRF album case had gained one more of them,
which put the tests' no-explicit-any count one over its baseline. All five
are typed now, and the baseline is lowered to the 1545 the tests hold.
The plugin in the move suite gains a BLOB column holding text and an
untyped column holding numbers, and the plugin store must read both back
as they were stored. A new suite moves a file whose plugin and own rows
hold NUL: the plugin's refuses before anything is written, TREK's
refuses until --drop-nul, a username that dropping the NUL makes
duplicate is named with users_username_unique, and a move without a
collision lists the value it changed and stores it without the NUL.
The move stripped NUL from a plugin's own text, merging keys that were
distinct, and turned a plugin's values into another type (text in a BLOB
column into bytea, numbers in an untyped column into text). Its checksum
was built from the converted values, so both passed as verified.
A plugin's data now moves exactly as the Postgres plugin store would
hand it back, or not at all: a NUL, a value of another storage class
than its column, or an integer past 2^53 refuses the move before
anything is written, listing plugin, table, row and column. A column
left open by SQLite's affinity takes the kind its values need, so text
in a BLOB column stays text.
TREK's own text holding a NUL from an older install is refused by
default too. --drop-nul drops it, lists every changed value in the
summary, and names a unique key that dropping a NUL broke.
Every table's checksum is now built from the values read from SQLite,
so a conversion that changes a value fails the verification.
nulFreeInput left a string as sent under any key containing "password".
That covered the admin settings password_login and password_registration,
which are flags: 'true\u0000' failed the lockout check as a disabled
password login while the database stored 'true'. It also tested the raw
key, so 'pass\u0000word' lost the NUL of the password it carried.
The exemption is now HASHED_PASSWORD_KEYS, the fields TREK only ever hashes
with bcrypt (password, current_password, new_password), matched exactly on
the key without its NUL. A unit test reads every field of the shared
schemas and the instance settings named like a password, passphrase or
secret and fails until each is either in that list or decided against with
its reason.
The plugin SQL guard refused functions by name, so every function that reads
the catalogs or the server and was not listed passed: obj_description,
col_description and shobj_description returned the comments on TREK's
tables, columns, database and roles by OID, format_type named a type by its
OID, and version(), current_database(), inet_server_addr() and the session
keywords described the server.
A call to a pg_catalog function is now refused unless the function is on
PLUGIN_FUNCTIONS (string, number, date and time, JSON, array, aggregate,
window, UUID, range, geometry, network and text search functions), in
either spelling of a call, f(...) and the column notation (x).f. The
catalog's function names of Postgres 14 to 18 are kept in
assets/postgres/catalog-functions.txt, and PGPLUG-021 fails on the lane
when the server has a function that list does not name, so a new major
cannot widen what a plugin may call. The session keywords current_user,
session_user, current_role, user, system_user, current_catalog and
current_schema are refused unquoted.
The token guard kept leaking: on a superuser connection an unqualified
`alter type int4 rename to int4x` renamed pg_catalog's type for the whole
server, sequences and catalog-only objects grew the database past the quota,
regtype/regconfig/regdictionary looked TREK's objects up by name, and a
column alias named insert or merge in front of INTO made a table. Instead
of a longer word list, the store now asks the database:
- No plugin SQL while TREK's role is privileged in the sense of the restore
check (PRIVILEGED_ROLE_SQL, imported from server-checks.ts): every db:own
call fails with PRIVILEGED_ROLE_PLUGIN_REFUSAL, read once per pool
connection, and the boot log says why.
- The objects an ALTER, DROP, REFRESH, TRUNCATE or CREATE INDEX ... ON
changes, and the tables a CREATE or ALTER TABLE ties to its own, are
resolved with to_regclass/to_regtype in the transaction right before the
statement, and it is refused unless each one found lies in the plugin's
schema. An unqualified int4, text or box is pg_catalog's and refused even
with the role check off.
- The quota counts every relation of the schema with its TOAST table, and a
schema holds at most PLUGIN_SCHEMA_MAX_OBJECTS catalog objects (relations,
types, constraints, enum labels).
- The guard refuses every reg* type and its array in every spelling, the
has_*_privilege functions, currtid2 and the BRIN/GIN maintenance
functions; INTO only after the INSERT or MERGE that is the statement's
verb or a WITH body's; TEMP only where it makes a temporary object
(CREATE TEMP/TEMPORARY, SELECT ... INTO TEMP), so a column named temp
works again; a base type; and a NUL anywhere in the SQL text.
- A bound value or a migration id holding NUL is refused with the reason
instead of being stripped, so the plugin's keys never change meaning.
Every bypass from the re-review is a test on the real database, on the
role the plugin data is meant to run as and on a superuser connection with
the role check off.
5ff5e7660 stripped U+0000 from every argument a plugin bound to its own
SQLite file, but left the rows it had stored alone. A key written with a
NUL (through a bound arg or char(0) in the plugin's SQL) could no longer be
found by the same key, upserts added duplicates, and keys that differ only
in where a NUL sits ('a\0bc', 'ab\0c') collapsed into one and collided.
A plugin's data is the plugin's, so its file binds the arguments exactly as
sent again, as before that commit. TREK's own data still holds no NUL; the
Postgres store, whose text cannot hold the character, refuses it instead.
The privileged-role check missed three ways a role can reach the database
host from SQL. Before PostgreSQL 16 a member of a CREATEROLE role sets its
role to it and grants itself pg_execute_server_program, since role
attributes are not inherited but the grant checks the role in effect. A role
granted EXECUTE on lo_export, lo_import, pg_read_file, pg_ls_dir and their
siblings reads or writes server files, and lo_export keeps the file after the
transaction rolls back. From PostgreSQL 15 on a role can be granted ALTER
SYSTEM on a setting or SET on a superuser-only one.
PRIVILEGED_ROLE_SQL now counts all three. The parameter grants are read
through query_to_xml inside a CASE taken only on 15 or later, because a
statement naming pg_parameter_acl does not parse on 14. A SET grant on a
setting every role may change anyway, and pg_monitor, stay ordinary. The
lane test creates each kind of role and shows the escalation the check is
for on the server it runs against.
The database stores no NUL on either engine, but the application still held
the value a client sent. A key or a check made from it was then a different
value from the one the database compared: `victim@x.test` followed by NUL
was a fresh key for the per-email password-reset throttle while the lookup
found the victim's row, so every spelling mailed a new link and voided the
last one.
Every input now loses its NUL where it enters TREK: the nulFreeRequest
middleware after the body parsers (path, query string, body, cookies), the
first global pipe for what a route parses itself (a multipart body), the
websocket adapter for every frame, and the MCP registry through a new
host-supplied input mapper, since /mcp keeps its raw body. Headers need
nothing: Node refuses a NUL in one with a 400. A string under a key naming a
password stays as sent: it is only ever hashed, and stripping it would lock
out an account whose password was set with one. The stripping in the
database layer stays as the backstop for what does not come from a request.
createFromBaseline created the migration history's table through MikroORM's
storage before, and outside, the transaction that runs the baseline. Without a
master transaction MikroORM commits that table on its own, so a first boot that
failed in the baseline (a server without ICU, a timeout, a dropped connection)
left an empty mikro_orm_migrations behind. The next start with
RESTORE_FROM_BACKUP took that for a database in use, skipped the restore with a
warning and brought up an empty instance, after which the restore never ran.
The table is now created on the baseline's own transaction, so a failed first
boot leaves the database empty. The boot restore also treats a lone migration
history that records nothing as no database, which a restore clears anyway; any
other answer keeps it, so the check fails closed. Database-Postgres says both.
The two checks over every file under src/ awaited ESLint's config file by
file, which lint:warnings counts as no-await-in-loop. They read the configs
with Promise.all instead.
Plugin statements on Postgres went over the extended query protocol through a
marker parameter and a hook that replaced query() on every client of TREK's
pool. That kept a plugin to one statement per round trip, but it changed
TREK's own hot path for a concern only plugin data has.
The plugin data store now checks a pg client out of the ORM's pool for each
transaction, through MikroORM's getNativeClient, and runs the transaction on
it directly. Every plugin write goes as a query with queryMode 'extended' and
every read through a pg-cursor on that client, so Postgres parses exactly one
statement either way. The settings the transaction makes are local, and
before the client goes back to the pool the store drops temporary tables,
cursors, listeners, session advisory locks and sequence state, or closes a
client it cannot bring back to that state. The row cap, quota, time budget,
untimed migrate, per-plugin queue, pool share and script commits are as
before, and plugin args are still sent without NUL.
The marker, the hook and its install in the Postgres driver are gone.
The ESLint walls against raw(), the sql tag and SQL text handed to
execute/run/all/get applied to src/db/repositories and src/db/seeders only. A
service, job, guard or MCP tool could send its own SQLite-only SQL through
em.getConnection().execute(...), which the Postgres probe never runs, since it
only calls the repositories. The walls now cover all of src/. The dialect layer,
the engines and the migrations stay the homes of raw SQL, and the four boot-time
lifecycle files that hold SQL text today are named and pinned to their count.
The wrapper took the chunk size of a stream as optional, which lint:strict refuses, and dropped the abort options Kysely hands executeQuery and streamQuery, so a cancelled query kept running.
resetDatabase empties tables and puts back the default categories one after another through the repo's oneByOne helper, where the order can matter, and the default addons and photo providers together; the MCP suites switch their addons on together and create ordered stops through oneByOne. The warnings this brings below the baseline, and the no-explicit-any and no-unused-vars counts earlier commits lowered, are lowered in scripts/eslint-baseline.json.
gen:pg-baseline rewrote the baseline with its cut at the newest migration on
disk, so after a migration landed after the cut, running it moved the cut over
that migration while Postgres installs created from the old baseline would
still run it through the migrator. It now refuses while a migration follows the
committed cut, names those migrations and the routine, and moves the cut only
with --move-cut (lint:dialect keeps such a migration under the post-cut rules
either way).
check:pg-baseline stopped comparing SQLite triggers with POSTGRES_TRIGGERS once
the baseline was frozen, so a trigger a later migration created on SQLite went
unnoticed. The comparison now runs in both states, with advice that points at
the migration when the baseline is frozen.
lint:dialect missed several classes of SQL that SQLite reads and Postgres does
not, or reads differently:
- group_concat, ifnull, instr, iif, printf, typeof, total(), changes() and
json_object/json_array/json_valid/json_type/json_quote, as calls in SQL text
and as the bare name handed to eb.fn.
- LIKE in SQL text, Kysely's 'like' operator and MikroORM's $like, which ignore
case on SQLite only. A LIKE that is exact on purpose now says so in a
like-exact comment; the six existing ones (setting keys, storage paths, the
lowercased reference search, sqlite_master) carry it.
- The sqlite member of engineDdl was not read at all, so a trigger or a data
backfill put there never reached Postgres. It may now hold only the schema
change and a rebuild's copy and sqlite_sequence bookkeeping, and SQLite
statements with an empty postgres member need a sqlite-only reason.
- execBestEffort, bestEffort and execUnlessTableMissing in a post-cut
migration: on Postgres a swallowed failure aborts the migration transaction.
- The post-cut rules follow POSTGRES_FIRST_CUT as well as the manifest, so
regenerating the baseline over a new migration no longer releases it.
src/db/engine/sqlite/ is SQLite's own code and is no longer scanned.
The gate matched only isSqlite(/isPostgres(, instanceof on a platform class
and a comparison on a property named engine or id. A destructured engine name,
the literal on the left, a switch over the engine, instanceof on a driver class
and a comparison with a driver class all passed with an empty baseline, and
standalone-orm.ts already held one such check. The regex now matches every one
of them, the guard in standalone-orm.ts moves into the engine port as
assertSqliteOrmConfig, and the advice names where a check belongs.
Postgres text cannot hold U+0000, so any write or filter carrying one (a
request body, a query string, an import, a feed, a forwarded mail, a
plugin's own SQL) answered a 500 on Postgres while SQLite stored it. Both
engines now drop the character where every value leaves for the database
(src/db/nul-free.ts): the MikroORM platforms when they inline a string or
serialize a JSON value, the Kysely dialects for every parameter, a
plugin's own SQLite file for its arguments, and the move from SQLite,
which used to refuse such text. The same request now answers the same
status and leaves the same row on both engines, so the places NUL suite
runs on Postgres as well and leaves SQLITE_ONLY.
The release queue narrows each owner's rows with LIKE '%run%' before it
parses them, and relied on SQLite's LIKE ignoring ASCII case. On Postgres
LIKE is exact, so a row naming an object in uppercase was not found and
the release deleted an object that row still names. likeAny now matches
LOWER(column) against the lowercase run, which reads alike on both
engines. RELI-010 fails on Postgres without it.
The nearest-day fallback of a booking's time cast the bound date to a
timestamp on Postgres, which refuses 2024-02-30 with a 500, while SQLite's
JULIANDAY rolled it over into March and placed the booking there.
dayDistance now refuses anything but a calendar date on both engines
(isCalendarDate), so a caller that forgets the check fails on SQLite too,
and findNearestDayId answers no day for such a date. The size baseline of
Reservations.repository.ts drops by one line.
A vacation day can be taken as a half day, so what a year carries into
the next can end in a half. vacay_user_years.carried_over was an INTEGER
column: SQLite stored 29.5 as a REAL, Postgres (bigint) rounded it to 30,
and the move refused such a SQLite value. The column becomes REAL on
SQLite (a table rebuild that keeps every id and the AUTOINCREMENT counter)
and double precision on Postgres, through the first migration after the
Postgres baseline's cut. VACAY-046 books a half day and reads the
carry-over on both engines.
The album list asks three sources at once; which call reaches the fake first
depends on how database reads interleave, which differs between the engines.
The fake now answers by API and category and the cases assert the retry with
a fresh session itself.
Every integer column is bigint on Postgres. SQLite stored a fraction sent
to one (a packing quantity of 1.5, a budget item for 2.5 days, an id of
3.5) as a REAL, while Postgres rounded it in a statement MikroORM renders
and refused it with a 500 in a Kysely statement, so the engines held
different data for the same request.
The request schemas of those fields now take whole numbers only, and
INTCOL (tests/unit/contracts/integer-columns.test.ts) holds the class: it
reads the integer columns from the entities and fails on any DTO field or
inline MCP tool field of that name that takes any number.
The privileged-role check looked only at the role's own rolsuper and at
pg_execute_server_program and pg_write_server_files. A role that is no
superuser itself but a member of one passed it and could SET ROLE to the
superuser from a crafted archive; on PostgreSQL 14 and 15 a role with
CREATEROLE could grant itself pg_execute_server_program; and
pg_read_server_files let COPY read the server's files into TREK's tables.
The check now counts membership in any superuser role, the three
predefined server roles, and CREATEROLE before 16. The boot warning and
the restore refusal describe the same list from one constant. A Postgres
lane suite creates each kind of role and evaluates the check as it, and
shows on the server it runs against whether CREATEROLE can grant itself
pg_execute_server_program.
The guard ended a -- comment at LF only, while Postgres ends it at CR as
well, so "select 1 --\rfrom public.users" hid a qualified name from the
guard that Postgres then read, and with no args a second statement after
the CR ran as TREK's role. It also only looked for INTO at the top level
of a SELECT, so "(select ... into temp users)" made a temporary table that
shadowed TREK's own users table on the pooled connection.
The tokenizer now skips exactly what Postgres skips as whitespace, ends a
line comment at CR or LF, and refuses any character the two read
differently (other control characters, non-ASCII whitespace such as
U+00A0, U+FEFF and U+2028, and the non-decimal and underscored numbers
Postgres before 16 reads another way) instead of guessing. INTO is refused
anywhere but after INSERT or MERGE, at any depth, and the words TEMP and
TEMPORARY anywhere outside a string or quoted name.
The guard tests carry both payloads and their variants; the Postgres lane
runs them against the real database, proves a second statement in one
text is refused by Postgres over the extended protocol, and that a
temporary table never shadows TREK's or a plugin's tables.
The plugin SQL guard was the only thing between a plugin and TREK's tables
on Postgres, and two gaps in it were enough to leave the plugin schema.
Two layers now hold even when the guard is fooled.
A plugin write with no bound args went over the simple query protocol,
which runs every statement in its text. Plugin statements now go through
extendedQuery, which pg sends over the extended protocol whatever the
args, so a second statement is a syntax error instead of a statement that
runs. A marker parameter carries the request to the pool's clients, and a
client without the hook binds one value too many, so it fails closed.
pg_temp is searched before the search_path unless the path names it, and
a temporary table lives as long as its pooled connection, so one left
behind would stand in for users, sessions and every other table for the
next caller. Every connection of TREK's pool now puts pg_temp last on the
search_path the role configured, and a plugin transaction sets its schema
with pg_temp last.
They seed through the booted container's ORM instead of the SQLite snapshot
mock and reset through resetDatabase, so they run in the Postgres lane as
well. The tag notification case waits for the fire-and-forget row with
vi.waitFor instead of one turn of the event loop: on SQLite the insert
lands within that turn, on Postgres it needs a few round trips more.
It takes the ORM from the app buildApp boots, seeds through the factories
and resets through resetDatabase instead of the SQLite snapshot mock, so
it runs in the Postgres lane as well.