server: failed plugin reload on config change silently drops custom agents and commands until restart

Summary

On v2.0.9 running as the background service (opencode serve --service, desktop app), config edits repeatedly fail with ERROR failed to reload plugins — TypeError: pe is not a function. After these failed reloads, the live service silently loses all file-discovered custom agents and commands: /api/agent and /api/command (and the TUI pickers) degrade to builtins only until the service is restarted. A freshly started server on the identical config loads everything, so the definitions and config are valid.

Environment

• opencode version: 2.0.9
• OS: Darwin 27.0.0 (darwin arm64), macOS
• Terminal: OpenCode desktop app (session host); TERM_PROGRAM=Apple_Terminal, TERM=xterm-256color
• Shell: /bin/zsh
• Install/channel: latest; binary at ~/.opencode/bin/opencode, background service mode
• Active plugins: none in final state (transient local plugins were registered and removed during the session that triggered the bug; config has no plugin key afterwards)

Reproduction

1. Set up custom agents (e.g. ~/.config/opencode/agent/review.md) and commands (project .opencode/commands/*.md ×23, global ~/.config/opencode/commands/*.md ×1). Start the background service and confirm via /api/agent / /api/command that they are loaded.
2. With the service running (active sessions, MCP servers connected, two locations booted: home + project), edit ~/.config/opencode/opencode.json repeatedly — in my case registering then removing local plugin entries, including a moment where a registered plugin's files were deleted before its config key was removed.
3. Each reload logs: ERROR failed to reload plugins cause="Cause([Die(TypeError: pe is not a function. (In 'pe()', 'pe' is an instance of Object))])"
4. Afterwards /api/agent returns only the 7 builtins (custom review agent gone) and /api/command returns only the 2 builtins (all 24 custom commands gone). No error is logged at the moment the catalogs degrade — the loss is silent.
5. opencode service restart → everything comes back.

Honest trigger caveat: on an isolated opencode serve (no --service, no connected clients) I could not trigger the crash with an mtime touch, a byte-identical rewrite, or adding a valid no-op plugin entry. It reproduced only on the shared --service instance with active sessions and multiple booted locations, which points at the concurrent/multi-location reload path also seen in #48121.

Expected Behavior

A failed plugin reload should not take down the agent/command catalogs, and file-discovered agents/commands should remain registered. Note the agent/command directories are not subscribed by any watcher (only skills and config files are), so once the catalogs degrade there is no self-healing path short of a restart.

Actual Behavior

Log excerpts (paths trimmed; timestamps UTC):

13:07:26 ERROR failed to reload plugins cause="Cause([Die(TypeError: pe is not a function. (In 'pe()', 'pe' is an instance of Object))])"
13:25:02 ERROR failed to reload plugins cause=… (same defect, recurs ≥10× over ~30 min)
13:35:19 WARN  failed to load plugin target="<project>/.opencode/plugin/aios-canary" cause="ENOENT …"   (plugin files removed before config key)
13:35:19 ERROR failed to reload plugins cause=… (same TypeError, ×3 in <200 ms)

After this window, on the still-running service:

/api/agent    → build, compaction, explore, general, plan, summary, title   (custom "review" agent gone)
/api/command  → init, review                                                 (23 project + 1 global commands gone)

Same binary + same config on a fresh opencode serve: 8 agents (incl. review) and 26 commands load correctly — verified twice on isolated instances.

Additional Context

• Workaround: opencode service restart.
• During the same window the background service also restarted twice without user action (may be unrelated, included for completeness).
• Recurring WARN failed to load OpenCode provider config … 401 GET https://opencode.ai/console/api/v2/config also appears on both healthy and unhealthy instances — likely unrelated auth noise.
• Related issues in the same subsystem: #48121 (concurrent location plugin reloads crash with r.base.get), #44920, #44975. The new part reported here is the silent loss of the agent/command catalogs until restart.

Start by reproducing the failure in opencode serve --service with active sessions and multiple booted locations, then inspect the config/plugin reload path and the /api/agent and /api/command entry points. Compare the behavior with a fresh isolated server and review related issue #48121. Done means a failed reload does not silently remove file-discovered agents or commands, and the catalogs remain available without a restart.