Agent Host: Copilot SDK tool-search override regressed after runtime Rust port

[bhavyaus]

Summary

Agent Host sessions backed by the Copilot SDK no longer use the Copilot extension's natural-language tool search. The runtime advertises and executes its own regex tool_search_tool instead of the tool_search_tool override VS Code registers with overridesBuiltInTool: true.

Expected

VS Code registers tool_search_tool (see _createClientSdkTools in src/vs/platform/agentHost/node/copilot/copilotAgentSession.ts) with a natural-language { query } schema and overridesBuiltInTool: true, and sets toolSearch.enabled: true on the session. The model should see VS Code's description and schema, and every tool_search_tool call should reach VS Code's handler, which ranks the deferred tools and returns toolReferences.

Actual

After the runtime's Rust port (github/copilot-agent-runtime#13629):

• toolSearch.enabled: true no longer enables client tool search without a rollout flag or experiment assignment.
• The override's description and { query } schema are dropped: Claude sees the built-in regex { pattern } schema and GPT falls back to provider-hosted search.
• tool_search_tool calls run the runtime's regex search instead of VS Code's handler, so semantic search never happens and the availableTools snapshot is never consulted.

Repro

1. Start an Agent Host session with a tool-search-capable model (e.g. Claude Sonnet 4.6 or GPT-5.6) and enough tools to cross the deferral threshold.
2. Ask a question that requires a deferred tool.
3. Observe the request's tool_search_tool declaration carries { pattern } (not { query }), and VS Code's tool-search handler is never invoked.

Prompt snapshots in #333689 pin the expected shape.

Fix

Runtime fix: github/copilot-agent-runtime#18448.

[vijayupadya]

This still reproduces for me in 1.137.0.

Configuration:

"chat.agentHost.enabled": true,
"chat.agentHost.copilot.toolSearch.enabled": true,
"chat.agentHost.copilot.toolSearch.deferThreshold": 0

After fully restarting VS Code, I created a fresh Agent Host Copilot CLI session using Claude Opus 4.8 and prompted it to invoke tool search first with the natural-language query workspace diagnostics and compile errors, then use the returned tool.

The Agent Host log instead showed the deferred problems tool being invoked directly:

sendMessage called: "Before using any other tool, invoke tool search with the natural-language query ..."
Tool started: problems

There was no Tool started: tool_search_tool, no tool_search_tool_regex, and no client round-trip with "toolName":"toolSearch". I observed the same direct invocation behavior with openBrowserPage on both Claude Opus 4.8 and GPT-5.6 Sol.

This suggests client tool search is not active or the deferred tools are still being exposed eagerly in the 1.137.0 build I tested.