Fewer interruptions: the language server’s checks never ask for approval, and Apex test runs ask only in production. The preflight check counts triggers per object, and each request to Claude carries about half the context.
Approval and safety
- Apex and LWC diagnostics and SOQL checks from the language server run without an approval card.
- Apex test runs, through the Salesforce MCP server or sf apex run test, run without asking in developer, sandbox and scratch orgs. In production they get a review card, without typing the org’s alias.
- A test run no longer shows in the Changes panel as a change that can’t be undone: its data rolls back on its own.
- An org without an alias is named on approval cards the way the org list names it, not by its 18-character ID.
The agent
- When existing automation already does what you asked for, the preflight check’s instruction to stop now reaches the agent, instead of the raw report.
- The preflight check counts active and inactive triggers for each object, so a request that touches Account and Contact no longer reads Contact’s triggers as Account’s.
- Recent setup changes reach the agent in full, not just how many there were.
- Each request to Claude carries about half the context it did.
- After your first message, how hard the agent thinks only goes up within a session, which keeps the conversation in Claude’s cache.
- “Yes, deploy it” and similar replies get the thought a deploy needs, instead of being taken for a question.
The app
- Tool calls and approval cards show paths within the org’s source folder, such as main/default/classes/AccountHandler.cls, with the full path on hover.
- Each tool call is named for what it does, such as “Check Apex for errors” or “Find files”, and a shell command by what it is for.