[pawaca]

Closes #373.

Why

computerd 0.3.1 finished a command on the child's exit event. Output read after that (a background job started with & that keeps the pipes, or a Durable Object reader slower than the command) reached EventLog.append, threw append after exit from a stream listener, and the daemon's uncaughtException handler exited it, stopping the container. On staging a burst of a few hundred KB through the container failed almost every time. cloudflare/computer#260, released in 0.7.0, finishes a command on close, or exitDrainMs (250 ms) after exit when a background job holds the pipes, and drops later output.

What changes

• @cloudflare/computer 0.3.1 → 0.7.0 (app and standalone manifests, both locks) and computerd 0.7.0 pinned by digest in container/Dockerfile; the wire protocol has no negotiation, so they move together. Notices regenerated.
• Container backend: CloudflareContainerBackend is now LegacyContainerBackend from backends/container-legacy, still platform-scheduled with the same containers block in wrangler.jsonc. Computer's new ContainerBackend needs scheduling_policy: "durable_object". Cloudflare cannot switch an existing application to that policy, and the replacement must use a new Durable Object class and namespace, which would leave every instance's storage behind. The policy is also in public beta.
• output: false on the Workspace: since 0.5.0, runtime.exec saves the full output of long commands to /.computer/output by default. The bash tool already keeps each stream's tail and spills to /workspace/.spill, so Computer's copy would double the writes.
• RecordingWorkspaceStub: follows 0.6.0's stub. getWorkspace() now reads hasAssets across RPC, and useThink is gone. This was found by integration: The RPC receiver does not implement the method "hasAssets".
• VFS resolve-plan patch: re-ported to 0.7.0, because resolveViaCte still probes vfs_dirents by parent_inode only. audit.json and vfs-resolve-plan.spec.ts follow it.
• minimumReleaseAgeExclude: @cloudflare/computer@0.7.0 was published on 2026-10-23, inside the minimum release age, so it is excluded explicitly, as the 0.2.1-rc.2 Harness packages are.
• README: the container output limitation is replaced by the 0.7.0 behavior.

Compatibility

The workspace VFS schema migrates from 8 to 10 on first open (vfs_changes.prior, _vfs_mounts.version). 0.30 refuses schema 10, so rolling an instance back after this lands cannot read its workspace. This goes in the release notes.

Evidence

• pnpm run typecheck, pnpm run test (1,113 passed), pnpm run lint, doc-sync, dependency-sync, upstream-parity, test:snapshot (13 passed).
• Session integration passes against both release artifacts (Direct and Isolated).
• test:container passes with computerd 0.7.0. The binary in the built image is byte-identical to the published 0.7.0 image and contains exitDrainMs. A new check runs (sleep 1; seq 1 200000) & echo started in the container, then reads a file kept on the container's own disk: it passes here and fails on main ("the container restarted after late output").
• row-cost.probe.mjs against 0.30.0: every step identical (443 SQL rows and 36 KV and alarm writes over plain turns).
• upgrade-from-release.probe.mjs from published 0.30.0, with the version temporarily bumped: 11 checks pass, including workspace files read back after the VFS migration.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-24T02:39:36.003177Z	8047106	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.