server: AgentNotFoundError for custom agents when ServerProcess.start drains sessions before agent registry loads #59585

Summary

When a ServerProcess starts, OpenCode drains pending/interrupted sessions before the
custom-agent registry (agents/*.md from config directories) is ready. Sessions whose
agent is a custom .md agent fail to drain with Session.AgentNotFoundError, and
subagent dispatches issued in this window fail with Agent not found: "<custom-agent>".
Built-in agents (e.g. general) resolve fine in the same window.

Environment

• opencode version: 2.1.4 (update check reports current == latest)
• OS: Windows_NT 10.0.19045 (win32 x64)
• Terminal: VS Code integrated terminal (TERM_PROGRAM=vscode, TERM=xterm-256color)
• Shell: cmd.exe host; agents run PowerShell via the shell tool
• Install/channel: release (npm @opencode/cli)
• Active plugins: one local plugin at ~/.config/opencode/plugin/memory.js (loads OK; unrelated)

Reproduction

1. Define custom agents in ~/.config/opencode/agent/ (e.g. backend-dev.md, mode: subagent)
   plus a custom primary agent set as default_agent.
2. Trigger a new ServerProcess start while sessions hold pending steps. Observed triggers:
   service restart; a subagent session killed by a provider error (free-quota exhaustion)
   leaving pending state; dispatching a subagent right after such an interruption.
3. Dispatch the custom subagent immediately.
4. Dispatch fails: Subagent failed (sessionID: ses_***): Agent not found: "backend-dev".
5. Server log at the same moment (UTC):
   • Failed to drain Session cause=Session.AgentNotFoundError: Agent not found: "backend-dev"
     at SessionContext.select <- SessionRunner.prepareContext <- SessionRunner.drain <- ServerProcess.start
   • a second identical error ~40-60 ms later for the parent session's custom agent
     (Agent not found: "nemesis"; earlier days: "fullstack-lead")
   • only AFTER those errors: watcher stopped/watcher started lines for the config/agent
     directories - i.e. the agent registry loads after the drain has already run.
6. Retry the same dispatch after the process is warm -> succeeds.

Expected Behavior

Drain at ServerProcess.start should wait for (or defer until) the custom agent registry
is loaded; custom agents should resolve like built-ins during drain.

Actual Behavior

Paired Session.AgentNotFoundError for every mid-flight session whose agent is a custom
.md agent; the subagent dispatch fails and the child session never runs a step. The parent
session recovers on new turns, but in-flight subagent work is lost.

Additional Context

• Intermittent: 5 failure bursts over 5 days (Oct 22 x2, Oct 24, Oct 27 x3), stack identical
  every time (always ServerProcess.start -> drain).
• Custom-agent dispatches succeed whenever the ServerProcess is warm; failures cluster right
  after service interruptions/idle-unload windows and after quota-killed subagent sessions.
• Built-in general subagent succeeded in the same window in which custom agents failed -
  points at registry load ordering, not session corruption.
• All agent definition files were present and parseable throughout (frontmatter verified).
• Workarounds: retry after warm-up; route work to built-in general; set
  experimental.location_idle_timeout: false to reduce cold starts (just applied, unverified).