fix(plugin): drop scope-close deregistration of shutdown hooks

[Alezander9]

What

Drop the Effect.addFinalizer that deregisters plugin shutdown hooks on InstanceState scope close. Fix 3 from the cubic review on PR #75 was a lifecycle regression.

Why

Verified on staging V4 after PR #75 shipped as v0.1.8-rc1: trace 0ee0b31c-4970-c840-be54-d7bf5da53b37 from a real cloud bcode run still has the exact same problem PR #75 was supposed to fix — top_span_landed=False, total_cost=0, all session.llm children point to one orphan parent (the bcode-laminar "turn" span) that never landed.

The synchronous shutdown hook from PR #75 was never being invoked. Root cause is the lifecycle, not the hook itself:

yargs.handler {                              // src/index.ts try
  effectCmd.handler {
    try {
      AppRuntime.runPromise(run command)     // bus events, plugin events, etc.
    } finally {
      AppRuntime.runPromise(store.dispose(ctx))   //   ← plugin layer scope closes HERE
                                                  //     my Effect.addFinalizer runs HERE
                                                  //     pluginShutdownHooks.delete(hook) HERE
    }
  }
}                                            // src/index.ts top-level finally
finally {
  for (const hook of pluginShutdownHooks) hook()   //   ← Set is EMPTY → no-op
  forceFlush()                                      //   ← nothing queued, exports nothing
  process.exit()
}

So the turn span never gets .end()'d, never gets queued for export, and the top-level forceFlush(3000ms) drains nothing useful.

pluginShutdownHooks is intentionally module-level — that's the entire point: the Set must outlive the Effect scopes so the host can reach it from outside the runtime. A scope-close finalizer defeats that.

Fix

Just drop the Effect.addFinalizer. Multi-instance TUI bloat (the cubic reviewer's concern) is the lesser evil: rare in practice, closures are tiny, and the cost is bounded by the number of instances opened over a TUI session lifetime. Comment block updated so the next reviewer doesn't re-add the dereregistration.

Diff

-        const registered: Array<() => void> = []
         for (const hook of hooks) {
-          if (!hook.shutdown) continue
-          pluginShutdownHooks.add(hook.shutdown)
-          registered.push(hook.shutdown)
+          if (hook.shutdown) pluginShutdownHooks.add(hook.shutdown)
         }
-        yield* Effect.addFinalizer(() =>
-          Effect.sync(() => {
-            for (const fn of registered) pluginShutdownHooks.delete(fn)
-          }),
-        )

Verification plan

1. Merge → tag v0.1.8-rc2 → bump cloud staging Dockerfile to rc2.
2. Submit a real cloud V4 task.
3. Pull trace, expect top_span_landed=True for the bcode-laminar turn parent and total_cost > 0 on the trace aggregate.
4. If green → v0.1.8 proper, then bump prod.

Refs

• v0.1.8-rc1 verification trace: 0ee0b31c-4970-c840-be54-d7bf5da53b37
• PR fix(plugin): synchronous shutdown hook for OTel span drain (revert PR #74) #75 (the original sync-shutdown-hook PR)
• EXCEPTIONS.md updated with the lifecycle analysis under the existing PR fix(plugin): synchronous shutdown hook for OTel span drain (revert PR #74) #75 section.

---

Summary by cubic

Stop deregistering plugin shutdown hooks on scope close so they survive InstanceState teardown and run at process exit. This fixes orphan spans and ensures telemetry flushes correctly.

• Bug Fixes
  • Removed Effect.addFinalizer that deleted hooks on scope close.
  • Hooks stay in module-level pluginShutdownHooks until the top-level finally executes them.
  • Prevents unlanded parent spans and zero-cost traces seen in v0.1.8-rc1.
  • Added a comment explaining the lifecycle to avoid reintroducing deregistration.

^{Written for commit adfa914. Summary will update on new commits. Review in cubic}

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}