[frahlg]

Problem and change

A CLI full-backup request sends {}. When its HTTP/1 connection closes, Core can keep copying and publish the archive because the handler has not read the body. Separate HTTP backup verification has the same unread-body issue, and Inspect(ctx, ...) also dropped the caller context during full archive verification.

Read the ignored create/verify body with the existing 1 MiB API limit before acquiring operation locks. An unreadable body returns 400; an oversized body returns 413 before work starts. Pass the caller context through the full Inspect verification, as well as the final archive hash, so cancellation releases the verification workspace and backup lock. Create also releases its update lock. SQL, durability, deadlines, retry and backup pacing stay unchanged.

Validation

• Create disconnect and oversized-body regressions fail on exact published .142.1 before the fix. The two added verify regressions also fail on that source: oversized JSON reaches archive lookup, and a canceled inspection parses the archive.
• All relevant create/verify regressions pass in three race runs; full make verify passes.
• Published .142.1 Core/CLI on an own 250000-row copy keeps JSON/CLI creation running after SIGINT; empty POST cancels. A separate valid archive with 512 MiB of synthetic padding shows JSON verification writing metadata after SIGINT.
• A local modified ARM64 Core with the published CLI cancels all three create cases without an archive. Empty and JSON verification release their active workspace within 70/39 ms of the signal in this probe and write no verification metadata. Normal JSON creation and verification still pass. All own probe containers are removed.

These are isolated copy/protocol checks on a virtual SSD. They do not prove Raspberry Pi backup throughput, the old household request’s exact network behavior, or the first household write-delay cause. No box runtime, release or version changed.

[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 2026-10-13T15:41:30.337810Z	f1421ec	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.

[miravoss26]

Reviewed. The PR does what it says: a full backup create/verify request that disconnects mid-flight used to keep running and could still publish the archive, because Go only notices a closed HTTP/1 connection once something actually reads the request body. readBackupRequestBody consumes it (capped at the existing maxJSONBody limit) before work starts, so a disconnect now cancels the context in time; Inspect threading ctx through to verifyInWorkspaceContext instead of the context-less Verify closes the matching gap on the verification side.

Security screen: no secrets, no new deps, no new network destinations, no authz surface touched. The new oversized-body path returns 413 before any lock is taken — looks right.

Correctness: the regression tests are doing real work here, not padding — TestBackupClientDisconnectCancelsBeforePublishing actually opens a raw TCP conn, half-writes a request, closes it, and asserts no archive lands; TestInspectCanceledContextDoesNotParseArchive catches the exact Inspect gap the PR description calls out. CI green across the board (a few skipped jobs look unrelated to this diff — web/drivers/module-boundaries).

Safe to merge from my read.