perf(agent-manager): overlap MCP startup with worktree creation

[marius-kilocode]

Cold Agent Manager worktree sessions currently initialize MCP servers only after worktree setup, session creation, and snapshot tracking. That makes MCP connection time a separate serial stage immediately before the first model request, even though MCP initialization does not depend on snapshot completion.

This change starts a directory-scoped MCP status request as soon as worktree setup finishes, before the CLI session is created. Session creation and the initial prompt continue without awaiting that request. The existing mcp.tools() call remains the synchronization point before model dispatch, so tool resolution still waits if initialization is genuinely unfinished.

The warmup lives in the shared Agent Manager worktree-session creation path. This covers normal New Worktree creation, configured and multi-version creation, imported worktrees, and Agent Manager tool-created worktrees without duplicating orchestration. It does not initialize MCP for unrelated directories or during extension startup. Warmup failures are caught and logged, so they cannot reject worktree/session creation or become unhandled promise rejections.

Observed cold-flow difference in the same repository and modal flow:

Stage	Before	After	Difference
MCP/tool resolution after snapshot	1.356s	0.106s	1.250s less serial work (92%)
Provider dispatch to first backend delta	1.304s	1.323s	Effectively unchanged
Modal submit to first rendered token	11.642s	12.144s	0.502s slower in this sample

The end-to-end sample varied because worktree/session setup and snapshot tracking were slower in the after run. Backend ordering confirms the intended optimization independently of that variance: MCP discovery and client creation moved into snapshot initialization, completed about 4.5 seconds before snapshot tracking finished, and no longer occupied the serial resolveTools() stage.

This deliberately does not change snapshot behavior or shared CLI code. Snapshot startup remains the dominant cold-worktree cost and can be optimized independently, while this change ensures MCP initialization is already overlapped rather than becoming the next serial bottleneck.

[kilo-code-bot]

Code Review Summary

Status: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	1
SUGGESTION	0

Issue Details (click to expand)

WARNING

File	Line	Issue
packages/kilo-vscode/src/agent-manager/mcp-warmup.ts	13	Potential memory/retention leak: if client.mcp.status() never settles, the warm async function suspends indefinitely, holding the log closure (which captures this/the provider instance) alive even after dispose(). No timeout or AbortSignal bounds the warmup lifetime.

Other Observations (not in diff)

Import grouping nit (AgentManagerProvider.ts line 38): The startSession import is separated from the contiguous block of local ./ imports by a blank line after line 36. The existing style in this file groups all local imports together without internal blank lines. This is cosmetic and won't affect behaviour.

Files Reviewed (5 files)

• .changeset/warm-agent-manager-mcp.md — no issues
• packages/kilo-vscode/src/agent-manager/mcp-warmup.ts — 1 issue
• packages/kilo-vscode/src/agent-manager/AgentManagerProvider.ts — no issues (call site looks correct)
• packages/kilo-vscode/tests/unit/agent-manager-arch.test.ts — no issues
• packages/kilo-vscode/tests/unit/agent-manager-mcp-warmup.test.ts — no issues (tests exercise real implementation, ordering is reliable)

Fix these issues in Kilo Cloud

---

_{Reviewed by claude-sonnet-4.6 · 446,226 tokens}

_{Review guidance: REVIEW.md from base branch main}