fix(cloud-agent-next) skip container wake on idle stop

[St0rmz1]

Summary

stopWrappers inspects the sandbox by calling listProcesses, which is a
container fetch and therefore boots a sleeping container. On the idle timeout
path that boot is wasted work: cleanupIdleKiloServer only requests a stop
after hasWrapperRuntimeOrPendingWork() has already confirmed from Durable
Object state that no wrapper runtime and no pending work remain. The inspection
then wakes a stopped container to re-derive that same fact, and the woken
container sits idle until the 900s sleepAfter expires it.

In 90 minutes of production logs, reason=idle-timeout produced 233 stop
inspections. Every one returned absent. None returned present.

This confirms absence from container state instead, when the container is
stopped. A wrapper is a process, and a process cannot outlive its container
(activity expiry SIGTERMs the container, exit code 143), so a stopped container
cannot be hiding a leaked wrapper.

Changes

• container-usage.ts: adds MeteredSandbox.isContainerRunning(), returning
  this.ctx.container?.running === true. It reads Durable Object state only, so
  calling it over RPC does not boot the container.
• container-usage-context.ts: adds the method to MeteredSandboxInstance and
  adds isSandboxContainerRunning(), using the same runtime guarded cast as
  configureSandboxBillingInput. It returns boolean | undefined, yielding
  undefined when the method is missing or throws, so an unknown state is never
  read as stopped.
• cloudflare-agent-sandbox.ts: in stopWrappers, when the reason is
  idle-timeout and the container is confirmed not running, return
  { status: 'absent' } without inspecting. Every other stop reason is
  unchanged.
• cloudflare-agent-sandbox.ts: the wrapper_stop_inspection log is now emitted
  once for all outcomes rather than once per branch, with a local
  StopInspection type adding an absent-no-container status for the new case.

Verification

• No manual testing of the fix. Reproducing it needs a sandbox container that
  has already gone to sleep plus a Durable Object alarm firing the idle sweep,
  which is not reproducible locally. The problem itself was measured from
  production logs (233 idle timeout inspections, all absent).

Visual Changes

N/A

Reviewer Notes

• The narrow scope is deliberate. The stopWrappers inspection was added by
  fix(cloud-agent-next): recover leaked sandbox wrappers safely #3555 ("recover leaked sandbox wrappers safely") to catch wrapper processes the
  Durable Object lease does not track, including duplicates in shared sandboxes.
  That recovery is preserved: it can only apply when a container is running, and
  in that case this change falls through to the existing inspection untouched.
  Only idle-timeout short circuits.
• isSandboxContainerRunning sits behind a && on the reason check, so the probe
  does not run at all for other stop reasons. There is a test asserting this.
• Returning bare { status: 'absent' } drops stoppedInstanceIds. That field is
  optional on StopWrappersResult and has no readers anywhere in the codebase
  (produced in stopWrappers, never consumed).
• Residual risk worth a look: this trades a container fetch for trusting
  ctx.container.running. If that flag were ever stale and reported false while a
  container was up, a legitimate leak check would be skipped. The same flag
  already gates billing adoption in configureBilling.
• Four tests cover the branch matrix: idle timeout with a stopped container
  (asserts listProcesses is never called), idle timeout with a running
  container, a non idle timeout reason with a stopped container, and a sandbox
  that does not expose isContainerRunning.
• Effect is checkable after deploy with existing logging: reason=idle-timeout
  should report observation=absent-no-container, and container_stopped with
  reason=activity_expired should fall relative to reason=exit.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

The idle-timeout stop path now short-circuits on DO-only container state (ctx.container.running) before any container fetch, and the behavior degrades safely: an unknown or failing probe returns undefined and falls back to inspection, every non-idle stop reason still inspects, and the absent result returned in the skip path is identical to what the previous inspection produced for a stopped container, so lease reconciliation in wrapper-supervisor.ts is unaffected.

Verification notes

• isSandboxContainerRunning mirrors the existing configureSandboxBillingInput capability-probe pattern (typeof guard plus try/catch), so a sandbox class without the RPC method cannot be misread as "stopped".
• All sandbox DO classes in src/sandbox-outbound.ts extend MeteredSandbox, so the new RPC method is present on every metered class; MeteredSandboxInstance has no object-literal construction sites, so the added required member does not break other callers.
• StopInspection narrows correctly: absent-no-container is returned before the initial.status !== 'present' branch, and observedWrapperCount still only reads observed on present.
• Added tests cover stopped/running/non-idle-reason/unknown-capability paths; default metadata() uses a ses- sandbox id, so discoverSessionWrappers only needs the mocked listProcesses.
• Not run in this read-only review: pnpm --filter cloud-agent-next test / typecheck.

Files Reviewed (4 files)

• services/cloud-agent-next/src/agent-sandbox/cloudflare/cloudflare-agent-sandbox.ts
• services/cloud-agent-next/src/agent-sandbox/cloudflare/cloudflare-agent-sandbox.test.ts
• services/cloud-agent-next/src/container-usage-context.ts
• services/cloud-agent-next/src/container-usage.ts

---

_{Reviewed by claude-opus-5 · Input: 64 · Output: 15.6K · Cached: 2.3M}

_{Review guidance: REVIEW.md from base branch main}