This commit updates the EE repository reference after PR #783 was merged in windmill-ee-private.
Previous ee-repo-ref: 5c853e2c20eca6b748415fc0d6862a6ebfb5fec4
New ee-repo-ref: a1873dbb67f2302b85ff5362f8387b48eccdb607
Automated by sync-ee-ref workflow.
The matcher it covers is enterprise code now, so without the enterprise
edition the test hit the stub and failed the default windmill-common run. It
runs with private and enterprise, and a counterpart without them asserts that
no tenant list covers anyone, the wildcard and a workspace admin included.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
Each test that exercises roles runs with private and enterprise. Two tests run
without them: every roles route answers the Enterprise refusal, and a data
table saved under roles, or a named role, is refused a connection while one
not under roles resolves as before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
The superadmin list, create, update and delete handlers move to
windmill-ee-private. The routes stay registered and, without the enterprise
edition, refuse after authentication.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
The permissions read, save and usable-roles handlers move to
windmill-ee-private; the routes stay registered and, without the enterprise
edition, answer that data table roles are an Enterprise Edition feature.
ensure_governs_datatable and ensure_reaches_datatable keep their paths: the
first refuses, the second passes a data table not under roles.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
Roles are an Enterprise Edition feature. The catalog, the Postgres logins,
CONNECT convergence, tenant evaluation and the role half of connection
resolution move to windmill-ee-private. Every public function keeps its path
and signature and forwards through datatable_roles_oss, which re-exports the
enterprise implementation or, without it, refuses.
Without the enterprise edition a data table under roles, or a caller naming a
role, is refused a connection rather than resolved as admin, and the reach and
admin-access checks refuse one under roles. A data table not under roles
resolves as before in every edition, and an instance database keeps the
CONNECT grants it was created with. The catalog lock, the stream lock, the
tenant cascades and the permissions stripping stay in OSS: they only restrict.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
The previous guard only caught a new name replacing an entry under roles.
A whole-map save could also repoint an existing entry without roles at
that database, or another workspace could point one there, and every
caller of that entry would connect as admin. The rule is now stated on
the saved entries: one that carries no roles and newly points at an
instance database any entry under roles uses, in this workspace or
another, is refused. A declared rename carries its roles and passes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
A data table's roles follow its entry only through a declared rename. A
settings sync sends the whole map and never declares one, so renaming a
data table under roles there read as a delete and a new entry on the same
database: the new entry carried no roles, and every caller connected as
admin. Such a save is now refused, naming both entries.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
A settings save reported the fork pointers left resolving to nothing only
for the names in `deleted_datatables`, which `wmill sync push` never sends.
The save now works out what it removed from the locked entries, and the
CLI prints the stranded pointers it returns.
Also correct the replication helper's contract: no role or admin check
makes a replication connection safe, so a data table under roles is
refused outright rather than gated as an admin operation.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
Turning roles on counted a trigger as gone once disabled, and a capture
once its client stopped pinging, but the listener keeps its replication
connection until its next heartbeat notices. A trigger or capture whose
listener pinged in the last 15 seconds, the window a server holds a
listener for, now still counts as streaming.
Data table names could contain `?` before they were restricted, and such
entries are still stored. Splitting `?role=` off a reference misread them:
`a?b` became `a` with an unknown parameter, and the clone checks looked at
a different entry than the one copied. An entry stored under the whole
reference is now looked up first, in the Postgres executor, DuckDB ATTACH
and the clone checks. Agent workers cannot read the workspace and keep
the strict parse, which refuses such a name rather than misreading it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Turning roles on looked for enabled triggers and live captures once,
without a lock anything starting a stream also took. A trigger enabled in
that window could have its listener connect before roles committed, and a
healthy listener never checks again. Both transitions now serialize on one
advisory lock: roles going on hold it exclusive while they look, and
trigger create, edit and enable, and capture setup and ping hold it shared
while they commit. Either the look sees the stream, or the listener
connects after roles are committed and refuses.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Deleting a role reassigns and drops what it owns in each registered
database on its own connection, and each of those passes commits as it
goes. A database failing part-way left the role enabled in the catalog and
able to log in, but already stripped in the databases reached before it.
The role is now disabled in its own commit first, so a failed delete
leaves a disabled role to retry.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Creating or editing a trigger that points at a data table under roles was
accepted, and its listener then retried the refused connection every 30
seconds forever. The save is now refused, and a trigger that reaches such a
data table anyway (re-enabled, or cloned into a fork) is disabled by its
listener with the reason, as a missing replication slot is.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A replication stream reads every row of every table whatever the data table's
roles grant, and its listener checks access only when it connects. Rather than
chase every way access can change and bounce the streams each one affects, a
data table now carries one or the other:
- a Postgres trigger or capture cannot be created on, or connect to, a data
table under roles;
- roles cannot be turned on while an enabled trigger or a live capture reads
the data table, its own or a fork's through its pointer. The refusal names
each one to disable.
This removes the stream bounces on roles edits and on data table and workspace
deletion, and the trigger gate that admitted admins. The fork schema baseline
fix from the same review round is kept.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three fixes from review.
`edit_datatable_config` took `forked_from` wholesale from the stored entry, so
the fork schema diff's save of an advanced baseline was silently discarded and
an applied change was offered again. Whether an entry carries a clone stamp is
still carried from the store, since that is what marks its database droppable,
but the baseline inside it is now taken from the request.
The stranded-pointer warning and the stream bounce ran over the optional
`deleted_datatables` hint, which the settings-sync CLI never sends, so removing
a governing data table through `wmill` bounced nothing. Removals are now derived
from the stored configuration against the saved one.
`delete_workspace` read the pointers to bounce before its transaction, so a fork
committing a pointer during the deletion was missed. The read now happens inside
the transaction, after the workspace row is deleted: a fork's insert key-share
locks that row through its parent foreign key, so it is either seen or fails on
the missing parent.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Deleting a governing data table, or the workspace that holds it, only collected
the fork pointers it stranded, for the warning. A Postgres trigger or capture
already streaming through one of those pointers kept the replication connection
it opened while the pointer still resolved, so it went on dispatching the
governing database's rows after the fork lost access — until its connection
happened to restart. The governing workspace's own streams on a deleted entry
did the same.
Both deletion paths now bounce the affected listeners inside their own
transaction, through the helper a permission change already uses, so a
listener that reconnects re-resolves the entry and finds it gone. The helper is
split so a caller can pass the (workspace, local name) pairs it already holds.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A clone is three requests and `CREATE DATABASE` is not transactional, so a
failure after the first leaves a registered `wm_fork_*` behind, as it did
before data table roles. Accepted for this PR: it is harmless to data and goes
away once the clone is a single server-side operation.
The comment also records why the obvious fix is wrong: reclaiming the leftover
on retry, without durable clone ownership, can drop another workspace's fully
copied database between its import and its final fork request.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This reverts commit 7dd3275a10.
The reclaim tied the caller to the source they administer, but not to the
database it dropped. Between another workspace's import and its final fork
request, that workspace's target is full, registered, unnamed and has no open
connection, so an admin of any instance data table could name it and have it
dropped and recreated empty. The victim's fork would then commit pointing at
the empty copy. Safe reclaim needs durable clone ownership and serialization
with the request that names the database; until then the leftover stays, as it
did before this PR.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A clone creates its target database one request before it copies into it, and
the fork that would name it is written a request after that. Any failure in
between — a pg_dump error, a bad restore, a dropped connection, the source's
roles changing mid-flow — left a registered `wm_fork_*` that no entry names,
and every retry then failed on its name. This predates data table roles.
`create_pg_database` now reclaims such a leftover before creating: only a
`wm_fork_*` database Windmill registered as a data table database and that no
data table or ducklake entry names, in any workspace, archived ones included.
The drop never terminates connections, so a clone still copying into it makes
the reclaim fail instead of being cut off. It is limited to callers who
administer the source — reaching it is not enough, since on a data table
without roles every member reaches it — and anyone else gets the refusal an
existing database always got.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A clone is three stages a workspace apart — `create_pg_database`, then
`import_pg_database`, then `apply_forked_datatable` inside the fork transaction.
Only the third can roll back, and `CREATE DATABASE` is not transactional, so any
refusal that lives there strands a registered `wm_fork_*` that no entry names
and whose name blocks the retry.
That orphan has now been fixed three times, most recently reintroduced by a
guard added one commit ago. Patching each new refusal into the first endpoint is
not the fix; having two places that can refuse is. `ensure_datatable_is_clonable`
now answers every reason a copy can be refused and returns what it resolved, and
the stage that writes the entry only does the work.
Also takes an ACCESS EXCLUSIVE lock before the rollback guard counts, so a role
created concurrently cannot slip between the check and the drop.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
The down migration dropped the table and left every role behind: live Postgres
logins whose passwords only that table carried, so after a revert Windmill could
neither use, disable nor delete them, and re-applying could not recreate them
because the names were taken. Cleaning up here is not possible either — dropping
a role means reassigning what it owns in every instance database, and a
migration runs in one — so it now refuses while the catalog is non-empty and
says to delete the roles through instance settings, which does the cluster work.
Also enforces the instance-only invariant the resolved-pointer clone relies on
rather than only asserting it in a comment.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
Forking a fork with cloning left an orphan database. The preflight resolves the
pointer and sees the governing entry, so both endpoints ran and filled the new
database; `apply_forked_datatable` then refused the inherited pointer and rolled
the fork back, stranding a registered `wm_fork_*` that no entry names and whose
name blocks the retry.
Refusing earlier would have been the smaller change, but forking a fork and
cloning worked before pointers existed, so it would trade an orphan for a
regression. Resolve what the pointer names and write the terminal entry the
clone needs: the whole `database` object rather than a patch of its
`resource_path`, since a pointer has none, and `reference` removed with it.
Also accepts `-- role=x` and `-- Role = x`, two more spellings that fell through
to the default role.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
Two review findings, both real.
`attach_datatable_parses_name_and_role` never compiled: `parse_attach_datatable`
returns `Result<Option<_>>` now and one call site kept a single `unwrap`. Its
`?Role=analytics` case also asserted a refusal, contradicting the parser in the
same commit, which matches the key case-insensitively. Replaced with the cases
that are genuinely malformed, and a positive one for the cased key.
`-- role: analytics` fell through to the default role — the silent fallback the
strict parser exists to remove, for the spelling most likely to be typed. The
keyword now accepts an optional colon, attached or spaced, while a word that
merely starts with it (`rolebased`) is still not an attempt.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
Enforcement ships first. The permissions drawer is what turns roles on, and the
catalog section is what creates them — both are only useful once there is a way
to grant a role the privileges it needs, which arrives with the ACL editor. Left
mounted they would offer a feature whose other half does not exist.
The two components are complete and reviewed; only their call sites here are
commented out, with a note pointing the follow-up PRs at them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
The retry rebuilds its SQL from `pruneComments(code)`, so the leading comment
block never reached the second attempt — and with it the `-- role <name>` line
that decides which login the query runs as. The retry connected as the data
table's default role instead, so a query the first attempt was denied could
succeed on the second, reported as "recovered with the row_to_json fix".
Carry the leading comment block over. The retry itself is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BjfMkJyKzodxkobqGZ6Lqb
`?Role=analytics`, `?role=` and `?x=1&role=…` all fell through the reference
parser's exact-match rule, so the connection resolved to the data table's default
role and ran under a login the caller never asked for — the URI half of the same
trap as a malformed `-- role` annotation.
The key now matches case-insensitively, and anything else in the query string is
an error naming it; `role` is the only parameter a reference takes. Callers that
only need the entry keep a lenient `datatable_ref_name`, since they never act on
the role. The DuckDB `ATTACH` parser propagates it rather than attaching under
the default.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
`-- Role operator`, `-- role operator;` and `-- role operator -- why` all failed
the annotation parser's exact-match rule, so the query fell through to the data
table's default role and ran, silently, under a login the author did not choose.
Naming a role exists precisely to not do that.
A leading comment whose first word is `role` is now an annotation attempt: the
keyword matches case-insensitively, one trailing `;` is tolerated, and anything
else is an error naming the line. Only callers that already know the target is a
`datatable://` reference ever run this, so ordinary SQL keeps its comments.
Also bumps the dev shell's postgres client to 18 — it trailed the server the dev
database runs, which takes out every data table export, clone and fork-with-data.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
A clone is two endpoints: `create_pg_database` then `import_pg_database`. Only
the second refused a data table under roles, so a fork asking to clone one
created and registered an empty `wm_fork_…` instance database and then failed —
and nothing collects it, since `drop_forked_datatable_databases` only drops
entries carrying `forked_from` and no entry names this one.
Refuse in both, so the clone stops before a database exists.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
pg_dump carries no roles and the import runs with --no-privileges, so a copied
data table arrives owned by the admin connection with no GRANT for any role.
The settings clone brings `permissions` across, so the fork's tenants pass
Windmill's check, connect as the role they were given, and are denied by
Postgres on everything: an entry that reads as configured and answers nothing.
Refuse the copy — in the import endpoint before any data moves, and in the fork
path the CLI takes. Replaying the source's owners and ACLs into the clone is
what lifts this, and is a change of its own. Dropping `permissions` from the
copy instead would be the unsafe half, since the copy holds the parent's rows.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
The tenant cascade on leaving went onto `/users/leave`. The UI and the generated client call
`/workspaces/leave` — a different handler in a different crate with the same name — which
deleted the membership and left `u/<username>` in the tenant lists. Leaving and rejoining
therefore restored the access the leave was supposed to end, and a later account taking the
username would have inherited it. The regression test drives the route the client actually
calls; without the fix it fails with "leaving kept the tenant".
The migration endpoints authorized too late. `run_datatable_migrations` opened the data table's
admin connection, created `_wm_migrations` and read it before reaching the per-migration role
check — so with nothing pending, nothing was checked at all. Rollback returned before its check
when nothing was applied, and the status endpoint had none. All three now ask, before any
connection is opened, whether the caller can reach the data table as any role at all; which role
a given migration runs as is still decided per migration, and by the executor after that.
Deleting a role committed the cluster drop and the catalog row, then swept the tenant lists in
separate transactions. A sweep failing part-way left workspaces naming a role nothing can connect
as, while the retry answered `NotFound` because the catalog entry was already gone. The sweep now
runs in the same transaction, so the drop, the row and every tenant list commit together.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
Three from the round, all about deciding on state that could already have moved.
A permission save resolved the data table and checked it was instance-backed before taking any
lock, then wrote under one. A config save committing in between could move the table onto a
PostgreSQL resource — recreating exactly what the transition guard refuses — or rename it, in
which case the write targeted a key that no longer existed and reported success having changed
nothing. It now re-resolves and re-checks on the locked state.
Rename validation checked that the source existed before and the target existed after, which
still accepts `main -> decoy` against a save that keeps both: every fork of `main` then follows
onto a different data table, silently, because it keeps resolving. The rule is now the actual
old-to-new key transition — a source may only survive if another rename took its name, and a
target may only pre-exist if another rename freed it. That also stops two sources sharing one
target, and it admits a swap, which the previous guard refused: `datatables` is keyed by name, so
a swap cannot be done one save at a time, and refusing it was a regression against main. The
pointer cascade now runs in two passes through a temporary name, the way the migration cascade
one layer down already handles the same shape, so `A -> B` with `B -> C` moves each pointer once
from what it named before the save.
The tenant mutators say what they are for: they write an access decision for any workspace named,
with an arbitrary mutation, and exist for the transaction that frees or renames a principal.
Editing a decision on purpose belongs in the permissions endpoint.
Carried in the same change: the stranded-fork list is a field rather than a phrase to grep out of
a success string; the pointer cascade matches with `EXISTS` instead of a `LIKE` over the whole
document, so a workspace whose pointers name something else is not rewritten to a byte-identical
value under an exclusive lock; and `InstanceDatatableRole` drops the serde derives left over from
the JSON document, one of which would emit `pwd`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
Five findings across three rounds were all the same choice. A set of live Postgres credentials
was living in `global_settings`, which has generic read, list, write, config-export and CLI
round-trip paths that know nothing about what they carry: the passwords reached the instance
config and its YAML editor, a full-row upsert of a neighbouring key erased the catalog,
`GET /settings/global/{key}` and the settings listing returned them raw, and this round the
redaction that fixed the last two turned `wmill instance push` into something that wipes every
password — a fix breaking the assumption the previous fix made. `POST /settings/global/datatable_roles`
could also empty it outside the lock.
The approved plan offered a table or `global_settings`, so this is the other option it already
allowed rather than a new design. `datatable_role` is a table: no generic settings path can read
it, list it, export it, write it or round-trip it, so none of the five needs a guard. The
redaction, the hidden/protected/agent-denylist entries and the JSON document all go with it.
One row per role also removes the read-modify-write the concurrency work was about: two
concurrent creates are two inserts, and the unique index on `name` is what settles a collision.
The advisory lock stays for the one window rows do not cover — `CREATE ROLE` is invisible to
another transaction until commit, so without it both creates pass their `pg_roles` check.
Also from this round: rename mappings are checked against the configuration they claim to
describe, since fork pointers are rewritten from them — a caller could otherwise submit
`main -> missing` against an unchanged config and repoint every fork of `main` at a name nothing
has, and `A -> B` plus `B -> C` moved what pointed at `A` all the way to `C`. And the warning
naming forks a delete stranded reached the response but not the screen: both the data table
settings save and the workspace delete now show it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
The raw settings readers hand back whatever is in the row, so moving the catalog into its own
`global_settings` key protected the config machinery and left `GET /settings/global/datatable_roles`
and the settings listing returning every live password. Both now filter that one key. The
neighbouring `custom_instance_replication_pwd` has the same shape and is not touched here: it
predates this and widening the fix to it is a decision about an operator workflow, not a
consequence of this change.
Three ways a save could leave something resolving to nothing:
A permissioned data table could be moved to a PostgreSQL resource. The block was carried across
as a server-owned field, the runtime refuses roles on a resource-backed table, so the save
succeeded and every job afterwards failed. Refused instead — turning roles off first is one step,
and it keeps discarding an access decision something somebody chose.
Renaming a governing data table left every fork pointing at the old name: the data table
disappears from their pickers and their jobs stop, with nothing in the renaming workspace to
suggest why. The rename now follows into the pointers in the same transaction.
Deleting one cannot be followed the same way, so it is reported instead — the response names what
it stranded, the way deleting a workspace does, and the fork's own error already says which
workspace is gone.
Also: `ensure_instance_db_grant_options_unchecked` claimed superadmin while the permissions
handler reaches it as a workspace admin (the same class fixed last commit, one instance missed);
the role entry kept an `instance_config_schema` derive it no longer needs; `write_role_catalog`
was the one writer of that table not stamping `updated_at`; and the concurrency test dropped its
roles only on success — a failing run is exactly the one that creates them without recording them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
The concurrency test reimplemented the read-modify-write inline, so deleting the lock from all
three handlers left it green — it pinned Postgres, not the code it was written for. It now
drives `create_datatable_role` twice concurrently and asserts the catalog kept both names.
Checked the way the last one should have been: removing the lock from the handler makes it
fail with "wmtest_a_… is a live cluster login the catalog forgot".
The contracts added last commit were stricter than this PR's own callers, which is worse than
none — the next reader sees a rule already broken and learns to ignore it.
`read_role_catalog` said superadmin-only while two of its four callers are open to any
workspace member, and `converge_connect_grants` said superadmin while
`set_datatable_permissions` reaches it as a workspace admin. Both were fine on substance: the
rule that actually holds is about the credential never reaching a response, log, audit record
or export, not about who may call. They now say that. `read_datatable_entry` gets the same
treatment rather than the one the earlier message claimed for it: it is the primitive every
resolution goes through, so it is deliberately open, and what must not escape is `permissions`
— it names the governing workspace's users, groups and folders.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
Putting it inside `custom_instance_pg_databases` was the wrong call, and it cost two ways.
The catalog serializes a generated Postgres password per role, and that row is the
operator-facing instance config, so the passwords reached `get_instance_config` and its YAML
editor — a live cluster credential in a response body, a UI field and any log of either.
Worse in the other direction: `to_settings_map` strips the catalog, so a full-row upsert of
that key writes the row back without it and the catalog is gone, while the cluster keeps every
login it described.
`custom_instance_replication_pwd` is the precedent and says exactly why — a generated secret,
written only by the server, never operator-authored, hidden so the config machinery cannot
read, rewrite or drop it. The catalog is the same thing, so it now has the same shape:
`datatable_roles`, in `HIDDEN_SETTINGS`, `PROTECTED_SETTINGS` and the agent-worker denylist.
No redaction to keep in step with three code paths, and no way for a neighbouring write to
take it out.
Two races on the same shared documents. `edit_datatable_config` read the stored data tables
outside its transaction and then wrote the whole `datatable` document, so a permissions save
committing in between was silently rolled back; it now reads under `FOR UPDATE`. And
`set_datatable_permissions` validated role ids against the catalog before opening its
transaction, so a deletion in between let it write a deleted role back — including as the
default, which every later job then fails on; it now holds the catalog lock and the settings
row across validation and write.
Completes the authorization contracts the previous commit claimed but did not finish:
`read_datatable_entry` (which it named and missed), `resolve_governing_datatable`, whose whole
job is to answer for a workspace the caller may not belong to, and
`converge_connect_grants_with`, which had not inherited its wrapper's.
Also the generic Python SDK reference: `_format_py_params` learned the bare `*` last time, but
`extract_py_functions` is a second formatter and still rendered `datatable(name, role)`, so
code written from that page passed a keyword-only argument positionally.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
The catalog is one JSON document, so create, rename, enable and delete are all
read-modify-write. Two concurrent creates read the same snapshot, both succeed in the
cluster, and the second write drops the first — leaving a live Postgres login with a password
nobody recorded, which is the exact state the delete path exists to prevent. Every mutation
now runs in one transaction holding an advisory lock across the read, the cluster DDL and the
write, so a lost update cannot happen and a failure rolls the whole thing back. The DDL
helpers take that transaction rather than the pool, which is what makes the lock cover them.
Their statements moved off `sqlx::raw_sql`: the simple protocol is only needed for genuinely
multi-statement SQL, and its future is not `Send`, which an axum handler holding the
transaction requires. Each of these is one statement anyway.
The new cross-crate surface now says what callers must do. `read_role_catalog` returns
plaintext credentials; `create`/`rename`/`set_login`/`drop_instance_role` and
`converge_connect_grants` mutate cluster-wide state; `read_datatable_entry` reads a workspace's
raw config. All of them are superadmin-gated by their current handlers, but nothing said so at
the definition, which is where the next caller looks.
Also: the roles table reloads after a failed login toggle instead of leaving it claiming a flip
that did not land; the rename affordance is the design-system `Button`, not a raw one; and
`resolve_datatable_pg_as_caller` drops a `role` parameter no caller ever filled — browsing
resolves as the data table's default until the database manager grows a picker.
Why role passwords stay a plain `String` while the instance user's password beside them is a
`StringOrSecretRef`, asked three times across reviews: that one is a secret ref because an
operator supplies it and may want it from their own backend, while these are minted here and
never entered by anyone, so there is nothing for a ref to point at. Encrypting generated
secrets at rest is a separate change that would take the replication password with it. Now
said at the field.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ti5HyeTikPMYyW8YSdiHR
