Oracle Fusion SQL — VS Code extension

An independent open-source project. Not affiliated with, endorsed by, or sponsored by Oracle Corporation. Oracle and Oracle Fusion are trademarks of Oracle Corporation.

Run SQL against Oracle Fusion Cloud from VS Code. No JDBC, no Java, no SQL Developer: the extension talks to BI Publisher over HTTPS by itself.

Queries travel as a parameter of a small BI Publisher proxy report — the SQL is gzipped and base64-encoded, the proxy's data model decodes it inside the database and opens a cursor, and the rows come back as delimited text.

Install

Download the packaged extension from the latest release and install it from inside VS Code: Ctrl/Cmd+Shift+P → Extensions: Install from VSIX…

Installing with code --install-extension while VS Code is running is not reliable: the running instance rewrites its extension list from memory and reverts the change, after the CLI has already reported success. Either install from inside VS Code as above, or quit it first:

curl -L -o /tmp/fusion-sql.vsix https://github.com/AleCyriaco/oracle-fusion-sql-vscode/releases/latest/download/fusion-sql.vsix
osascript -e 'quit app "Visual Studio Code"'
code --install-extension /tmp/fusion-sql.vsix --force && open -a "Visual Studio Code"

Confirm with code --list-extensions --show-versions | grep fusion-sql.

Publishing (maintainers)

Upload the .vsix at marketplace.visualstudio.com/manage — drag and drop, and the version in package.json is what identifies it.

That page needs nothing but the Microsoft account that owns the publisher. The vsce publish route additionally requires a personal access token from an Azure DevOps organization, and creating one can demand a linked Azure subscription — a detour worth skipping unless publishing is automated.

To build it yourself:

npm install && npm run compile && npm run package

Press F5 in this folder to run from source in an Extension Development Host.

Using it

1. Oracle Fusion in the activity bar → + for a new connection, or Import from connections.json to reuse an existing fusion-query / MCP setup.
2. Fill the form and press Test Connection before saving — a wrong host or password is only ever discovered by trying.
3. Click a connection to make it the window default.
4. New Query in the view toolbar opens an editor already set to SQL. Press Ctrl/Cmd+Enter, or click ▷ in the editor toolbar.

Each editor can run somewhere different. With more than one environment configured, the status bar shows which connection the focused editor uses; click it to pin that editor to another one, or to let it follow the window default again. A pinned editor is marked, so changing the default cannot silently move a query you deliberately aimed at one environment — which is the point when the two tabs open are DEV and PROD.

Results come back in one panel per connection, so the same statement run against two environments gives two tabs to compare rather than one replacing the other. Each keeps its own page and paging.

The Run button and the keybinding only appear on editors whose language is SQL. A file made with Ctrl/Cmd+N starts as plain text, so the toolbar looks empty — use New Query, or click the language name in the status bar and pick SQL.

Every query you run is kept under Query History — right-click to run it again, open it in an editor, copy it, or remove it. A query that failed is kept too: that is the one you come back to fix.

Right-click a connection for Test, Edit, Duplicate, Delete, and Sign In / Sign Out on SSO connections. Editing opens the same form, with the stored password left alone unless you type a new one; renaming carries the secrets across.

Results open beside the editor, one page at a time, with Export CSV — one panel per connection, named after it.

Signing in

Username and password — the ordinary case. Works everywhere, and can deploy the proxy report for you the first time you connect.

Single sign-on (OAuth 2.0) — browser sign-in using Authorization Code with PKCE, so no client secret ships with the extension. The identity domain is detected from the Fusion host, so all you supply is a client ID: an application registered in that domain with vscode://AleCyriaco.fusion-sql/auth as a redirect URI.

Where the domain federates onward — Microsoft Entra, Okta, anything — makes no difference. The token still comes from the Oracle domain, and the browser follows the rest of the chain with the session you already have.

SSO cannot deploy the proxy report. BI Publisher's SOAP services authenticate from <v2:userID>/<v2:password> inside the request envelope, which a bearer token cannot satisfy. An SSO connection therefore requires reportPath to point at a proxy report that already exists — deploy it once with a username/password connection, or share one.

Some pods also answer the BI Publisher REST endpoint with HTTP 500 and only work over SOAP. On those, SSO cannot run queries either, and the extension says so instead of failing obscurely.

Passwords and tokens go to VS Code SecretStorage (the OS keychain). Settings hold only the host, username and report path, so they stay safe to sync.

Generating queries with an AI model

Fusion: Generate Query with AI (or the ✨ button on a SQL editor) opens a panel: describe what you want, get a statement, then keep asking for changes — each follow-up edits the statement rather than starting over. Copy it, save it to a file, insert it into the editor, or run it straight away.

The model works inside a harness. Every reply is inspected before it can run: exactly one statement, SELECT or WITH only, and nothing that reaches outside the query — no UTL_, DBMS_ or APEX_ packages, no HTTPURITYPE, no database links, no EXECUTE IMMEDIATE. A refused reply goes back to the model with the reason. This is enforcement rather than instruction: on a live pod UTL_HTTP.request answers ORA-29273, which means the call was attempted and only the network stopped it.

It can check a name rather than guess. When the model is unsure a table or column exists, it replies with a query against the data dictionary alone. That is recognised as a lookup rather than an answer — it runs, and the rows go back so the real query is written from what is there. Up to three lookups, and the prompt carries what the dictionary actually looks like here: the connected user reads through synonyms, so ALL_TABLES is usually empty for application objects while ALL_TAB_COLUMNS, ALL_OBJECTS and ALL_SYNONYMS answer properly.

The statement is checked before you get it. A model writing SQL for a schema it cannot see will occasionally invent a table or a column, and the result looks perfectly plausible until it is run. So the extension runs it — a single-row page against your active connection. Oracle resolves every identifier and, when it objects, names the offending one:

ORA-00904: "COLUNA_QUE_NAO_EXISTE": invalid identifier

That error goes straight back to the model, which corrects that identifier rather than rewriting the query. Up to three attempts, then the statement is handed over anyway — marked as unverified, because you may know something the model does not. The panel says which happened: Runs on FUSION-DEV · 7 columns, or the error it could not get past. Turn the whole thing off with fusionSql.ai.validate if you would rather not spend the round trip.

The model is told what it needs to know to be useful here: Oracle dialect rather than generic SQL, SELECT only, no hand-written paging (the client adds it), and the Fusion conventions that trip people up — _ALL multi-org tables, _B/_TL translations, date-effective _F/_M rows — along with the usual table names across Payables, Receivables, GL, Purchasing, Inventory, Projects, Assets and HCM.

Bring your own key — Add AI Helper (API Key), offered in the Query History view until one is stored, or from the view toolbar:

Fusion: Add AI Helper (API Key)

Switch between them with Fusion: Select AI Provider; adding a key with Add AI Helper (API Key) switches to that provider on its own. Change the model with fusionSql.ai.model. Model ids drift faster than anything else here, so these are only defaults — a provider that answers 404 gives a plain "not found" naming that setting, and nothing needs reinstalling to move on. fusionSql.ai.baseUrl applies only to the compatible provider — the named ones always use their own endpoint, so a setting left over from another provider cannot send your key elsewhere.

Keys go to the OS keychain, the same place as connection passwords, and are sent only to the provider you selected. Nothing about your connections, credentials or query results is sent anywhere — only the words you type in the panel and the statement being edited.

Settings

"fusionSql.connections": [
  {
    "name": "FUSION-DEV",
    "url": "pod.fa.us2.oraclecloud.com",
    "authMode": "basic",
    "user": "FUSION_USER"
  },
  {
    "name": "PROD-SSO",
    "url": "pod.fa.us2.oraclecloud.com",
    "authMode": "sso",
    "reportPath": "/Custom/FusionQuery/v1/csv.xdo",
    "oauth": {
      "authorizeUrl": "https://idcs-xxxx.identity.oraclecloud.com/oauth2/v1/authorize",
      "tokenUrl": "https://idcs-xxxx.identity.oraclecloud.com/oauth2/v1/token",
      "clientId": "abc123",
      "scope": "openid offline_access"
    }
  }
]

Limits

• Read-only. The proxy runs SELECT; there is no DML, DDL or PL/SQL.
• Pagination is added automatically as OFFSET … FETCH NEXT, unless the statement already paginates.
• Values are returned as text, exactly as BI Publisher formatted them — so a 15-digit PO_HEADER_ID keeps every digit instead of becoming a float.

Troubleshooting

Fusion: Show Log (or Output → Oracle Fusion SQL) records every command, the editor and language it ran against, the connection it chose and the row count — so a command that appears to do nothing can be traced to the step that stopped.

Developing

npm test          # protocol, CSV, ZIP and template patching
npm run watch

The layers are deliberately separate: protocol.ts and zip.ts are pure functions with no VS Code and no I/O, which is why they can be unit-tested with plain node --test.

License

MIT.