Debugging via MCP 0.5.0

0.5.0 — Exception settings (preview)

Added

• Exception settings: exception_list, exception_set, exception_reset. The agent can see which exceptions Visual Studio stops
  on when they are thrown (Debug → Windows → Exception Settings) and turn Break When Thrown on or off for named exceptions,
  so a debugging session stops where the exception is thrown instead of where it is caught. A call checks every name first and
  changes nothing when one is unknown (it offers close names). exception_reset puts back only what the agent changed, and Visual Studio puts it back by itself when the agent lets go or disconnects. These are
  Control tools: they ask like the other Control tools when you set that in Options. Needs the extension and the debugviamcp
  server of the same release (protocol 2.8); an older extension answers that it must be updated.

• Mask sensitive values (Tools → Options → All Settings → Debugging → Debugging via MCP, on by default): values that look
  like secrets (strings under names such as password or token, JWTs, Bearer tokens, PEM private keys, GitHub and AWS keys,
  Password= in connection strings, passwords in URLs) are replaced by [masked] before they reach the agent, in vars, eval,
  the arguments of a break, exception texts and output_read. The result carries masked: true. Turning it off asks you to confirm.
  A tracepoint that would print such a value is refused while it is on. A heuristic against accidental leaks, not a security boundary (see docs/security.md).

• vars opens several levels in one call. depth (1-4, default 1) puts the members of each object under children;
  expand names more paths to open in the same call (listed under expanded). Each object shows at most 20 members
  (childrenTruncated says there are more) and the answer stays small: what does not fit stays closed, and hasChildren
  still says it can be opened with path.

• debug_status with detail. Stopped at a break, detail: true also returns the top 10 frames of the Call Stack, the
  first 30 Locals and the exception (when the break is about one): where you are and what is going on in one call. A part
  Visual Studio cannot read is named under notes and the rest still comes.

• process_list lists your processes that debug_attach can attach to (pid and name), with an optional filter on part of
  the name, so the agent can pick a pid when several processes share a name.

• symbol_status says why a breakpoint may not bind: the debugger options that decide whether symbols load (Require source
  files to exactly match, Just My Code, Source Server, Suppress JIT optimization), the breakpoints that are not bound, and what
  can be behind it. It does not say which modules have symbols loaded; Visual Studio's automation does not offer the symbol path.

• bp_set with temporary. A one-shot breakpoint: it removes itself after the break it causes, and bp_list shows it
  (temporary: true) until then. A breakpoint that already existed is left as it is, and a tracepoint cannot be temporary.

• A dialog Visual Studio waits on is named, not waited for. When a modal dialog is open (for example the Hot Reload
  "changes not supported" question), debug_status returns its title under dialog, and every call that would move the
  program, build or change a value fails at once with busy and says which dialog to answer. Reads keep working. The check
  asks Windows about the main window only, so it costs no pause on Visual Studio's UI thread, and each held-back call is
  logged with its call id.

Changed

• debug_start builds first. A failed build comes back to the agent (read the diagnostics with errors_list) instead of
  leaving Visual Studio's "run the last successful build?" dialog open on screen.
• The Output panes are named by what they hold: Debugging via MCP Log (the technical log, formerly Debugging via MCP)
  and Debugging via MCP Audit (what the agent did, formerly Debugging via MCP Activity). The menu entry Show Activity
  is now Show Audit. The agent still cannot read either pane.