[Bug]: Pi sessions declare every t3-code tool on every turn #16651

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Start a Pi thread, or delegate a task to Pi.
2. Send a short prompt.
3. In the Pi session file, count the first system message's toolsAdded entries named mcp__t3-code__* and check the first assistant usage.

Expected behavior

T3 tools a session rarely uses aren't declared on every turn. Pi 0.99+ supports this for extension tools: registerTool takes exposure (codemode, deferred), and inactive tools also leave out their promptSnippet and promptGuidelines (extensions.md, Tool exposure).

Actual behavior

piT3McpExtensionSource.ts registers every tools/list result with pi.registerTool() and no exposure, so all of them default to direct. Each one sends its full schema plus a snippet line and a guideline line.

From one Pi session on Opus:

That's about 116k of the 287k-char prefix, roughly 46k tokens at the 2.5 chars per token this session measured (first turn: 114.5k input tokens). Every delegated Pi child pays it again. Children get the same list as the orchestrator, because McpHttpServer.ts registers toolkits statically and session capabilities only gate calls.

main (517188b) registers tools the same way and defines 79. #16466 and #16528 add more through the same path.

Impact

Minor bug or occasional failure

Version or commit

0.0.46-nightly.20261006.2735, also checked main @ 517188b

Environment

macOS 26.7.1, T3 Code Nightly desktop, pi 1.0.4, anthropic/claude-opus-5-5 and claude-sonnet-5-5

Workaround

Untested: a second Pi provider instance for children, with --exclude-tools 'mcp__t3-code__*' in its launch arguments.

Possible fix, for Pi 0.99 or newer (MINIMUM_PI_VERSION is still 0.80.5):

pi.registerTool({
  name: registeredName,
  exposure: DIRECT_TOOLS.has(name) ? "direct" : "codemode",
  namespace: { name: "mcp__t3-code", description: "T3 Code orchestration" },
  // ...
});

DIRECT_TOOLS would be a short list such as delegate_task and task_status. Alternatively, Pi 1.0 ships a built-in MCP client, so the extension could call pi.registerMcpServer("t3-code", { url, headers, exposure: "codemode" }) instead of running its own client. The comment saying "Pi core has no MCP client" is out of date.