[Bug] Desktop host dies with CordisError INACTIVE_EFFECT during plugin hot-reload (Fiber._reload)

[uahoi]

Describe the bug

The desktop host can die during a plugin hot-reload with CordisError: cannot create effect on inactive context (code INACTIVE_EFFECT). The error is raised inside Fiber._reload while dsh-desktop-host re-applies the bundle; the host process then exits and the desktop app quits.

Stack trace (from crash-<ts>-main.log)

DesktopHostFatalError: cannot create effect on inactive context
    at ChildProcess.<anonymous> (file:///.../resources/app.asar/lib/main.js:3711:49)
    at ChildProcess.emit (node:events:509:28)
    at emit (node:internal/child_process:972:14)
    at process.processTicksAndRejections (node:internal/process/task_queues:91:21)

--- host diagnostic (as reported by the Host process) ---
CordisError: cannot create effect on inactive context
    at Fiber.assertActive (file:///.../app.asar/dsh/node_modules/@deepseek-ai/cordis/lib/index.js:1133:9)
    at Proxy.plugin (file:///.../@deepseek-ai/cordis/lib/index.js:1622:18)
    at Object.apply [as callback] (file:///.../@deepseek-ai/dsh-desktop-host/lib/index.js:80:12)
    at async Fiber._reload (file:///.../@deepseek-ai/cordis/lib/index.js:1356:5) {
  code: 'INACTIVE_EFFECT'
}

The companion crash-<ts>-web-boot.log adds:

Error: Error invoking remote method 'dsh-desktop:boot': Error: cannot create effect on inactive context
    at file:///.../resources/app.asar/lib/main.js:11573:15
    at Session.<anonymous> (node:electron/js2c/browser_init:2:122853)

Environment

• @deepseek-ai/dsh-desktop 0.2.1-rc.2
• Electron 44.0.0 / Node 24.18.1
• Windows 11 x64

To reproduce

Not deterministic. Observed once, during a window in which the loader was reloading plugin entries (a plugin config patch had just been written). It looks like a race between Fiber._reload and context disposal: the reloaded entry's apply callback (dsh-desktop-host/lib/index.js:80) calls ctx.plugin(...) after its fiber context has already been deactivated.

Additional context

• Both crash logs are written to the desktop logs directory (%APPDATA%/@deepseek-ai/dsh-desktop/logs).
• After this fatal, the renderer also reports N entries did not activate.
• Happy to provide the full crash logs if useful.

[PerryLink]

The second mount happens after an await. Nothing re-checks that the fiber is still active.

apps/desktop-host/src/office.ts:28-37: apply is async, line 29 does await ctx.plugin(workspaceDependencies, config), lines 30-32 do synchronous path work, and line 33 does await ctx.plugin(officeSkills, {...}). If a hot reload disposes or replaces that fiber while line 29 is pending, the context is inactive by the time line 33 runs, and Cordis refuses to create an effect on it — INACTIVE_EFFECT, which is exactly your stack (Fiber._reload → plugin() → assertActive).

The fix belongs on this side, not in Cordis: either mount both children in one synchronous turn — they do not depend on each other — or re-check activity after the await before the second ctx.plugin.

What I could not check from here: what triggered your reload (a cordis.patch.yml write or the plugin manager's restart), and whether the 25 bundles matter. The code path is the same either way, but I did not reproduce it.

Source read at f3803c96 (0.2.2-alpha.2); this file is unchanged since 22ddee3.