Subagents started in the same turn as an approved exit_plan_mode get none of the built-in tools (regression after 1.0.95) #5243

Describe the bug

When exit_plan_mode is approved with the interactive action, the agent keeps working in the same turn. Every subagent it starts in that turn runs without any built-in tools: no powershell or bash, view, create, edit, grep, glob, sql, task, skill, or web_fetch. The subagent only gets the external tools the SDK host registers, such as openBrowserPage, readPage, runPlaywrightCode, navigatePage, semantic_search, issue_fetch, and activePullRequest. Its system prompt has no <tools> section.

The main agent keeps all its built-in tools in that same turn, and subagents started from any other turn get the full set. So any plan-then-implement workflow that delegates to subagents breaks without warning. In my case, a reviewer subagent could not write its report or run the test suite, and fell back to reading files through the browser tools.

I hit this through an SDK host (the VS Code Agents window, reported there as microsoft/vscode#343716). The host does not seem to be involved. It answers the exitPlanMode.request with approved: true and selectedAction: "interactive", and as far as I can tell from its code, it sends nothing else to the runtime before the subagent starts. Subagent tool sets are built inside the runtime.

Affected version

1.0.100-1, as bundled with VS Code 1.143.0 (@github/copilot-sdk-win32-x64 1.0.23-preview.5). 1.0.92 and 1.0.95 are not affected.

Steps to reproduce the behavior

1. Create ~/.copilot/agents/Probe.agent.md:

   ---
   name: Probe
   description: Reports the tools it has
   ---
   
   Reply with the exact names of every tool you have, as a comma-separated list. Do not call any tools.

2. Start a session in plan mode through an SDK host. I used the VS Code Agents window with the Copilot harness.

3. Send: "Plan a single step: run the Probe subagent and report its answer verbatim."

4. Approve the plan with the interactive action (VS Code labels it Implement plan).

5. The agent runs Probe through the task tool in the same turn.

Probe lists only the host's external tools. The session log shows exit_plan_mode completing with "Plan approved, exited plan mode (interactive)" and session.mode_changed from plan to interactive, both before subagent.started.

Expected behavior

Probe gets the same built-in tools it gets when started from a normal turn.

Additional context

Counted from my local session logs over the last 20 days, by whether the subagent's system prompt had a <tools> section:

• All 6 failing calls were custom agents with no tools key in their frontmatter, started through task with agent_type, model, and reasoning_effort. subagent.selected reports tools: null in both failing and working calls.
• The host creates the session with availableTools set to explicit builtin:<name> entries plus custom:* and mcp:*. The failing subagents kept the external tools and lost every built-in one. My guess is that the built-in set given to subagents is computed differently right after the mode switch, but I can't see the runtime code to confirm.
• I have two preToolUse hooks. They returned success in both failing and working calls.
• I only tested the interactive action, not autopilot or autopilot_fleet.
• OS: Windows 11 (10.0.26300), x64
• Host: VS Code 1.143.0 Agents window, Copilot harness
• Model: a BYOK Claude Opus model from a VS Code language model provider extension, for the main agent and the subagents