Daily harness signal

Make “done” call its own witnesses

Claude Code 2.1.215 stops choosing to run its bundled verification and code-review skills. Any completion loop that relied on that discretion now has a silent gap.

July 20, 2026 · JST Fresh · source date July 19 UTC One implementation finding
Verification is a workflow edge, not a personality trait. Trigger it explicitly after the last change, preserve its evidence, and reopen the gate whenever a fix changes the artifact.
01 · Fresh · source date 2026-07-19 02:56 UTC

Remove acceptance from model discretion

Use when. You run Claude Code 2.1.215 or later and your interactive, CI, or unattended workflow assumed Claude would opportunistically invoke the bundled /verify or /code-review skill before declaring work complete. This includes prompts such as “verify before done” that never inspect the resulting tool trace.

Action. Change the supervisor’s completion transition, not merely its prose. After the final edit, submit /verify as a standalone skill call, then submit /code-review high with an explicit target such as main...feature. Capture both invocations and their outputs in the run record. If either step causes another edit, invalidate both receipts and rerun them in that order. Keep deterministic tests, builds, and policy checks outside these model-operated skills; for unattended sessions, enforce those checks with CI or a blocking Stop hook.

final edit → explicit /verify → explicit /code-review high <target> → inspect receipts → release or repair and restart the gate

Acceptance check. Canary the upgraded harness on a disposable change. Pass only when the transcript contains explicit post-edit invocations of both named skills; /verify records an observed running-app result; review reports no unresolved correctness finding; independent checks are green; and no write occurs afterward. A prose assurance, a pre-edit run, or a missing trace fails the gate.

Evidence. Anthropic’s signed-asset 2.1.215 release states the behavior change directly. Its current skills and commands documentation identifies both workflows as bundled, prompt-based skills, documents direct slash invocation, and defines their distinct oracles: live application behavior for verification and targeted diff analysis for review.

Caveat. The release note publishes neither rationale nor regression output. /verify can infer the wrong launch recipe for projects with databases, environment files, GUIs, or multi-step startup; record a project run skill or use an external oracle. A project skill can also shadow a bundled name, and code-review findings do not themselves approve or block a pull request.

Compact source notes

  1. Claude Code v2.1.215 (published 2026-07-19 02:56 UTC / 11:56 JST). Official release and signed checksum assets; states that Claude no longer runs /verify or /code-review on its own.
  2. Claude Code skills (retrieved 2026-07-20). Official invocation, bundled-skill, shadowing, and run-recipe behavior.
  3. Claude Code commands (retrieved 2026-07-20). Official syntax and oracle definitions for /verify and targeted /code-review.
  4. Claude Code best practices (retrieved 2026-07-20). Official guidance for executable feedback, independent review, and blocking Stop-hook gates.