MangoCode

Had an Oracle APEX Interactive Grid that forgot its checkboxes after every refresh. Worked through it with GitHub Copilot. Three fixes, all plausible, none of them worked. The bug was somewhere else entirely.

The root cause is in APEX's own source. It defines its own pageChange in the grid options and passes those after yours, so a pageChange in defaultGridViewOptions never runs. Listening to the gridpagechange event on the grid view works.

Had an Oracle APEX Interactive Grid that forgot its checkboxes after every refresh. Worked through it with GitHub Copilot. Three fixes, all plausible, none of them worked. The bug was somewhere else entirely.

Every Copilot fix changed when the restore ran. None of them made it run after every refresh, so the guard stayed up and real clicks were thrown away too.

Had an Oracle APEX Interactive Grid that forgot its checkboxes after every refresh. Worked through it with GitHub Copilot. Three fixes, all plausible, none of them worked. The bug was somewhere else entirely.

Showing a row and ticking its checkbox are separate in an Interactive Grid. The query can return the row again after a refresh, but the grid does not know it should be ticked. So you keep the IDs yourself and put the ticks back.

Had an Oracle APEX Interactive Grid that forgot its checkboxes after every refresh. Worked through it with GitHub Copilot. Three fixes, all plausible, none of them worked. The bug was somewhere else entirely.

I spent my Friday night typing Dutch into the new APEX 26.2 Interactive Report search bar. The assistant told me it had filtered my report. The report said "No data found." Here is the whole story, in English and Dutch on the same data.

What I could not fix: grouping. "what is the total amount per category" errors in English and Dutch. The user sees "An unexpected error occurred." The real ORA-20960 is in debug only. If you made it work, I would love to hear how.

I spent my Friday night typing Dutch into the new APEX 26.2 Interactive Report search bar. The assistant told me it had filtered my report. The report said "No data found." Here is the whole story, in English and Dutch on the same data.

Then I set the application language to Dutch. "filter status op Afgerond" ran locally and "opnieuw instellen" reset the report, the same words as the reset button. The English phrase stopped working. Finding the other Dutch phrases is trial and error for now.

I spent my Friday night typing Dutch into the new APEX 26.2 Interactive Report search bar. The assistant told me it had filtered my report. The report said "No data found." Here is the whole story, in English and Dutch on the same data.

In a default English application, English key phrases work on Dutch column names and values. "filter status by Afgerond" and "break on categorie" both did exactly what they should. Values with a space need quotes.

I spent my Friday night typing Dutch into the new APEX 26.2 Interactive Report search bar. The assistant told me it had filtered my report. The report said "No data found." Here is the whole story, in English and Dutch on the same data.

Spent some time trying to make my own APEX app leak data with the new AI assistant in 26.2. Logged in as Bob, I asked for Alice's order 1001, and the careless version I wrote gave me a working link. Thread on what I found.

I would also like to know how far an agent that reads data and renders markdown links is exposed to indirect prompt injection. My test agent reads nothing, so I could not test it. If you have tried it on 26.2, I would love to hear what you saw.

Spent some time trying to make my own APEX app leak data with the new AI assistant in 26.2. Logged in as Bob, I asked for Alice's order 1001, and the careless version I wrote gave me a working link. Thread on what I found.

The careful callback checks the ID is a number and that the order belongs to the current user. Same question, five refusals out of five.

Spent some time trying to make my own APEX app leak data with the new AI assistant in 26.2. Logged in as Bob, I asked for Alice's order 1001, and the careless version I wrote gave me a working link. Thread on what I found.

The careless callback passed the model's ID into APEX_PAGE.GET_URL, which signs it with a valid checksum. Bob clicked and read "Confidential merger fee". Five runs out of five.

Spent some time trying to make my own APEX app leak data with the new AI assistant in 26.2. Logged in as Bob, I asked for Alice's order 1001, and the careless version I wrote gave me a working link. Thread on what I found.

The model writes a placeholder, and APEX calls a PL/SQL callback you write. Oracle's docs say it directly: "the navigation callback is the application security boundary."

Spent some time trying to make my own APEX app leak data with the new AI assistant in 26.2. Logged in as Bob, I asked for Alice's order 1001, and the careless version I wrote gave me a working link. Thread on what I found.

A UAT tester reported that some of their selected rows never got processed. No App Builder in UAT, so I could not look. A thread on what I built, what I learned afterwards, and what the bug actually was.

For a session only someone else can reproduce, APEX_SESSION.SET_DEBUG switches debug on for them, and APEX_DEBUG_MESSAGES gives you the trace as a plain query.

A UAT tester reported that some of their selected rows never got processed. No App Builder in UAT, so I could not look. A thread on what I built, what I learned afterwards, and what the bug actually was.

So I put the query in a report region on the page, where it does run inside the session, and made it visible only in debug mode. Condition: PL/SQL Expression, apex_application.g_debug.

A UAT tester reported that some of their selected rows never got processed. No App Builder in UAT, so I could not look. A thread on what I built, what I learned afterwards, and what the bug actually was.

The selected rows go into an APEX collection, and a cursor joins it to two tables. In DEV the collection had every ID, so the rows had to be dropping out in one of the joins.

Six ways to fire many REST calls at once from inside an Oracle database, with working code for each. Four of them turn out to be the same mechanism wearing different names, which changes what the choice is actually about.

And APEX background execution chains when a user is watching, purely for the free progress reporting.

Six ways to fire many REST calls at once from inside an Oracle database, with working code for each. Four of them turn out to be the same mechanism wearing different names, which changes what the choice is actually about.

195 real embedding calls, 8 workers requested, timed first row start to last row end so no mechanism is charged for its own submission overhead: DBMS_AQ 7.74x APEX background chains 7.29x DBMS_PARALLEL_EXECUTE 7.09x DBMS_SCHEDULER 6.87x

Six ways to fire many REST calls at once from inside an Oracle database, with working code for each. Four of them turn out to be the same mechanism wearing different names, which changes what the choice is actually about.

Everything follows from one constraint: PL/SQL has no asynchronous HTTP. UTL_HTTP and APEX_WEB_SERVICE block the calling session for the entire round trip. One session holds exactly one REST call in flight, so twenty at once needs twenty sessions, and every way of getting them is job-based.

Someone built a small Go tool that fires twenty AI image calls simultaneously instead of one at a time. Half an hour of generating becomes about a minute. I wanted to know whether Oracle could do the same thing from inside the database. Thread on what I found.

Then the parallel hint approach, using PX query slaves rather than jobs. With DBMS_VECTOR_CHAIN.UTL_TO_EMBEDDING inside it the session never came back. No error, no timeout, no ORA number because no row got far enough to write one. Killed it by hand after 65 minutes.

Someone built a small Go tool that fires twenty AI image calls simultaneously instead of one at a time. Half an hour of generating becomes about a minute. I wanted to know whether Oracle could do the same thing from inside the database. Thread on what I found.

So twenty calls at once needs twenty sessions, and every way of getting more sessions in Oracle is job-based.

Someone built a small Go tool that fires twenty AI image calls simultaneously instead of one at a time. Half an hour of generating becomes about a minute. I wanted to know whether Oracle could do the same thing from inside the database. Thread on what I found.