Hot plugin reload destroys the shared agent/skill registry for all locations (TypeError: o is not a function) #59443

Summary

On 2.1.4, a hot plugin reload caused by a filesystem change under a path declared in plugins[] throws
TypeError: o is not a function and does not fail only that plugin: it tears down the shared registry of
agents/skills/tools for every booted location. In-flight turns then die with AgentNotFoundError and
permission.rejected.

Environment

• opencode 2.1.4 (@opencode/[email protected]; npm dist-tags.latest = 2.1.4 at the time of writing)
• Linux 4.19 aarch64 (Ubuntu 24.04 chroot on Android), node v20.20.2, bun 1.4.4
• one service on 127.0.0.1:49374 (opencode serve --service), 6 booted locations
• config declares two external plugin directories ("plugins": ["/abs/pluginA", "/abs/pluginB"])

Symptoms

ERROR failed to reload plugins  cause="Cause([Die(TypeError: o is not a function. (In 'o()', 'o' is an instance of Object))])"
ERROR Failed to drain Session   cause="Session.AgentNotFoundError: Agent not found: \"orquestador\""

Tool calls that were in flight are rejected:

{"error":{"type":"permission.rejected","message":"Permission denied: shell"}}

Registry before/after the event:

• GET /api/skill → 74, then 2
• opencode debug agents → 14, then 7 (builtins only)
• the same for every location, and it does not recover on its own

Repro A — a declared plugin directory disappears while the server runs

1. config: "plugins": ["/abs/plugin"] (a directory containing index.js)
2. boot a location: GET /api/agent?directory=/tmp/proj
3. mv /abs/plugin /abs/plugin.old
4. in the same second the log shows:

WARN  failed to load plugin  target=/abs/plugin  cause="... ENOENT: no such file or directory ..."
ERROR failed to reload plugins  TypeError: o is not a function
ERROR Failed to drain Session   Agent not found: "<agent>"

Writing a file inside that directory produces the same result: the reload path is entered either way.

Repro B — reload with a booted location

1. start opencode serve --port 49998 with a config whose plugins[] lists a path that cannot be resolved
2. GET /api/agent?directory=/tmp/proj (boots location services)
3. POST /api/location/reload → 503 {"_tag":"ServiceUnavailableError","message":"TypeError: o is not a function. (In 'o()', ...)"}
4. GET /api/skill?directory=/tmp/proj → {"data": []}

Caveat, to be precise: on a server where the same plugin set is already loaded, POST /api/location/reload
returns 204 and logs location services reloaded count=6. So it is not reload itself, it is the reload that
happens when the plugin set/instances change (or when a declared plugin fails to load) that takes the registry
down.

Expected

• A plugin that fails to load should be skipped/reported, leaving the rest of the registry intact.
• A failed reload should not leave every location without agents and skills.

Actual

• The shared registry is dropped for all locations and every in-flight session dies mid-turn.

Impact

Any agent session on the machine dies when a declared plugin file is touched, including routine work by an
unrelated agent. Restarting the service is the only reliable recovery (an explicit POST /api/location/reload
sometimes restores it, sometimes returns 503).

Happy to test a specific build or provide full opencode.log excerpts.