v2026.5

Perforce Agentic Gateway v2026.5

Token-savings reporting for code mode, a readable answer when a server is
running but serving no tools, and richer typed results for sandbox code.

claude mcp add pag -- uvx perforce-agentic-gateway@latest

New features

• Token efficiency reporting. A new Efficiency page in the dashboard
  reports how many tokens code mode kept out of the client's context, split
  into Execute (arguments and results that stayed in the sandbox) and
  Explore (tool schemas the model never loaded), with a per-server and
  per-tool breakdown of Execute and a selectable date range. A badge on the
  navigation item carries the combined total, and each server's detail panel
  carries its own figure. The same report is available to agents through the
  new pag__metrics tool and over HTTP at GET /api/metrics. Figures are
  gateway-side estimates and are labeled as such.
• Discovery readiness on running servers. A server whose MCP handshake
  succeeded no longer looks identical whether its capability discovery
  finished, is still in flight, or failed. Each running server now reports
  discovering, discovered, or discovery_failed on pag__status, on
  GET /api/health, in the server list, and in the Tools tab empty state.
  When discovery did not complete, the dashboard and pag__status say so and
  name restarting the server as the remedy, pointing at the Activity tab and
  pag__server_logs for the reason.
• Output schemas and structured results in code mode. pag__explore
  catalog entries now carry each tool's declared outputSchema, and the array
  callTool() returns now carries the downstream result's structuredContent
  when the server sends one. Sandbox code can read the typed object directly —
  r.structuredContent ?? JSON.parse(r[0].text) — instead of guessing the
  return shape and re-parsing a text blob. Existing content indexing is
  unchanged.

Improvements

• Downstream tool changes take effect mid-session. A server that adds,
  removes, or changes a tool and sends notifications/tools/list_changed (or
  the prompts, resources, or resource-templates equivalent) is now re-queried,
  so the change reaches the model without restarting that server.
• Startup failures explain themselves. A server that reaches "Running" but
  never publishes a session used to produce a single unexplained "server
  unavailable during discovery" line. Init-handshake failures and post-init
  failures now log the concrete error.

Bug fixes

• A server that came up with no tools now recovers on its own. When a
  server's capabilities were not yet published as startup discovery ran, the
  one corrective re-query was discarded, so the catalog stayed empty for the
  rest of the session unless the user restarted the server.
• Slow-starting servers are no longer abandoned at 15 seconds. The wait for
  a server's session and the capability list calls that follow it now get
  separate budgets, so a server still resolving its package (for example a
  uvx cold start) contributes its tools instead of coming up empty.
• Servers that do not implement logging/setLevel now expose their tools
  when the client has set a log level. Previously the server showed as Running
  and reported "server unavailable" with no tools routed — reproducible under
  VS Code, which always sets a level after initialize.
• Template variables in catalog environment values are now substituted. An
  environment variable combining a publisher-set value with a user-supplied
  variable, such as https://{mcp_node}/mcp/, reached the server process with
  the placeholder intact.
• The built-in ui_builder server can no longer be deleted. The dashboard
  offered a Delete button that always failed, and the underlying API had no
  guard on the reserved alias.
• First-run config seeding is now atomic, fixing a startup failure that
  reported version field is required for a config file that was valid by the
  time anyone looked at it.
• Shutdown no longer takes ten seconds. Signaling the gateway waited out a
  supervisor timeout before exiting, and a closed stdin left the stdio surface
  restarting itself in a loop against closed pipes.
• Code mode no longer races on tool arguments. callTool() read its
  arguments off the JavaScript event loop and could hand the dispatcher a map
  the sandbox program still held, crashing the process on concurrent access.
• A callTool() issued from a promise continuation is now waited for
  rather than escaping the sandbox's dispatch barrier, which could also panic
  the process.