Building thoughtful software with AI.Notes from the systems behind the work.
Making settings saves queue on one workspace lock opened a gap: the permission check ran before the wait, so a signed-out, demoted or expired admin could still write. Now authority is rechecked under the locks, with second-factor proof read outside them and validated again at the end.
Read moreJob titles were swapped for permission-based roles in the scheduling app, and the diff shows that the seeding step run on every staff listing rewrites any edit to a built-in role, alongside non-atomic writes, no last-admin guard, and a few access changes I can't swear I meant.
Custom plans resolved to nothing and nothing meant unlimited, so unknown codes now fall to the free tier. Invites, reactivations, transfers and first sign-in all ask the same question, and a copied existence check would have blocked every workspace edit.
Hiding a product from one store was a single database write. Keeping it off real orders meant checks in the cart, checkout, reissued orders and the suggestion report, which as I read it would otherwise have flagged a hidden item and maybe suggested a full case of it.
A paid AI job moved from plan-based access to a per-organization flag and got checked in six places, then reading the diffs back showed the workers skip the check without their key, an outage reads as a plain no, and my job table needed three corrections in one day.
The dashboard's export route gathered the whole result into memory before replying. The rewrite streams but never waits for a slow reader, and the plan check guarding it had its key respelled the same day.
The licensing code I shipped checks how many devices a key already has, then inserts the new one in a separate step, so two activations racing for the last slot can both get in. The test activates devices one at a time and can't see it, a device racing itself gets a server error, and a key-only check passes for keys never bound anywhere. I found this by reading, not by running anything concurrent.
Swapping a local copy of billing state for plan entitlements in the session was the right call, but it left the free-tier seat cap answered by three pieces of code that disagree, a random retry key on customer creation, signup that now depends on the payment provider, and billing endpoints that trust whatever organization the request names.
The apply endpoint on my self-hosted coding assistant had no login, took any folder, used a prefix check for paths and handed a request-supplied command list to the shell. Now the server owns all four decisions, and reading it again turned up a few doors that are still open.
Restoring a backup puts the old accounts back, including the one running the restore. The agent's first attempt saved only the admin's password; the second saved the whole account and its grants, and I walk through the gaps I can still see in the code, starting with the login rename and id reuse.
The way my admin app links a code-host install to a tenant got a rewrite. It checks that an installation exists, not that it belongs to you, and when the webhook arrives first, nothing ever finishes the setup.
Workspace creation waited for a later webhook so invited people wouldn't get a spare one, and self-serve people could log in before it fired. Now the creation handler asks the identity provider for memberships on the spot, a later event repairs misses, and the session lookup retries, but the comment that justifies it contradicts the one it replaced and nothing tests which is right.
A commit collapsed the app's org roles to admin, manager and member. The script header said executives become managers, the SQL made them admins, and the new ranking code treated them as members.
Every member could add, remove and re-role team members behind a cap on which roles they could hand out. Reading the diff back, the cap stops promotion and nothing else: demotion, removal, unknown role names and missing roles all walk past it.
Moving a teams screen from route handlers to server actions deleted two files and kept every permission check, but left the role list in three places, turned refusals into ordinary successful calls, and tested some actions' gates far better than others.
Promoting tenant ids out of a telemetry map into indexed columns went from computed-from-the-row, to a collector copy step that kept the collector from starting and likely left history empty, and back. What I can and can't verify about which version is live.
A cache, most likely agent-written, stored a copy of each good key while the revoke route only knew a row ID, so a cancelled or rotated key would have kept working until its cached copy aged out, and the usage counter could run backward.
A checkpoint added an MCP endpoint to my observability tool that validates the request envelope, then runs any registered tool with whatever repository ID the caller typed. Canned data is the only reason nothing real has leaked.
An auth step I put in front of my agent-event collector decoded the bearer token instead of verifying it, so the caller's own claims picked the user, the organization, a default role and the author of every stored event.
The first commit of my coding-agent dashboard let any caller send SQL to the database, with only a check on the first word and no login. I take apart the statements that slip past that check, why a perfect read would still have exposed every prompt I typed, and what I'd put there instead.
The schedule parser built its name lookup from a staff table shared by every company, admins included. It now has to be told the company, both import endpoints check it against the token, and anyone outside that company comes back unmatched.
A Delete button was added to the backup page in the scheduling app. It sends a file path, so the server resolves it, fences it to the selected company's folder with a separator-aware check, refuses anything but a regular file, and only then deletes. What still bugs me: name-derived folders, hard deletes, no password, no tests and a vague audit entry.