Daily harness signal

Background review still needs a foreground quality gate

July 23, 2026 · JST 2 fresh findings Review lifecycle
Claude Code now backgrounds its built-in reviewer and forked skills by default. The implementation win is context isolation; the release risk is allowing a downstream ship step to outrun its evidence.
01 · Fresh · source date 2026-07-22

Insert an explicit join after background review

Use when: A Claude Code workflow on 2.1.218 or later invokes /code-review, or a custom context: fork skill must finish before commit, merge, deployment, or another mutating stage.

Action: Classify every forked skill as side work or a gate. Add background: false to gate skills. For the bundled reviewer, record:

git diff --binary HEAD | git hash-object --stdin

Invoke /code-review high, then add this orchestration rule: “Do not enter the next mutating stage until /tasks marks the review done, its completion notification has been consumed, and the diff hash still matches.” Rerun review after any repair.

Acceptance check: In a disposable branch, start review on a known diff and immediately queue a marker-producing “ship” step. Pass only if the marker is absent while review is running, /tasks later shows done, the completion result is present, and equal before/after hashes prove the reviewed bytes did not move. A changed hash must force a second review.

Evidence: Anthropic’s signed-asset 2.1.218 release explicitly moves /code-review and context: fork skills to background execution; current official docs define /tasks and later-turn completion notifications as the observable join surfaces.

Caveat: The release publishes no regression output, and the skills reference has not yet added background to its frontmatter table. Named background subagents also receive a different built-in tool pool, so canary the exact skill and runner before rollout.

02 · Fresh · source date 2026-07-21

Promote reviewers with incident-derived evals

Use when: You want automated review to replace human review on selected, lower-risk paths.

Action: Build a path-scoped corpus from PRs that caused incidents, preserving the causal diff and expected defect category; add representative clean PRs. Run the candidate reviewer on that frozen set before changing CODEOWNERS. After each new incident, add its causal PR and rerun the entire set.

Acceptance check: Every historical incident fixture is flagged at the expected file and defect category, clean controls remain within a predeclared false-positive budget, and critical paths still require a human code owner. Any miss blocks promotion.

Evidence: In Simon Willison’s dated transcript, Claude Code leaders describe this progression: retain manual review for core areas, automate outer layers only after measured coverage, and turn incident-causing PRs into permanent code-review evals.

Caveat: Anthropic released neither the eval set nor its results; the reported 65% automated product-PR share is internal and does not transfer.

Compact source notes

  1. Claude Code v2.1.218 (published 2026-07-22 21:24 UTC / 2026-07-23 06:24 JST). Official release with signed checksum asset; directly states both background-execution changes.
  2. Claude Code subagent reference and commands reference (retrieved 2026-07-23). Official definitions of background completion notifications, reduced built-in tool pools, and completed work in /tasks.
  3. Claude Code skills reference (retrieved 2026-07-23). Official context: fork semantics; its current frontmatter table does not yet expose the new skill-level background override.
  4. A Fireside Chat with Cat and Thariq (Simon Willison, 2026-07-21). Dated edited transcript linked to the source video; operational testimony, not published evaluation data.