From 142b680141cf4fd956adc3b63c7a4c07cce83e3a Mon Sep 17 00:00:00 2001 From: Maximilian Roos Date: Sun, 9 Aug 2026 05:24:55 -0700 Subject: [PATCH] docs: drop the pinned version from the port's preamble MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit `claude/action.yaml` moved to 2.1.226 while #904 was in flight, so the sentence naming 2.1.220 as the version it pins shipped already false. The skill no longer names a version at all — it had no reason to, and a version sitting in a file nothing re-checks is a claim waiting to rot. `TODO.md` keeps the provenance, where the drift is the point rather than an aside: the port came from 2.1.220, CI now runs 2.1.226, and the restructuring between them is exactly what decides whether the port should survive. Co-Authored-By: Claude Opus 5 (1M context) --- TODO.md | 15 ++++++++++----- .../tend-ci-runner/skills/code-review/SKILL.md | 2 +- 2 files changed, 11 insertions(+), 6 deletions(-) diff --git a/TODO.md b/TODO.md index e38b3d3e..729c6499 100644 --- a/TODO.md +++ b/TODO.md @@ -280,11 +280,16 @@ The same gap covers `tend-outage`; #816 tracks it from that side, and the ## Decide whether the built-in `/code-review` replaces the vendored port `plugins/tend-ci-runner/skills/code-review/` is a copy of Claude Code's -built-in `/code-review`, read out of the 2.1.220 binary `claude/action.yaml` -pins. The two share their method almost verbatim — the same ten angles, the -same three-verdict verify with "PLAUSIBLE by default", the same sweep — so the -copy buys nothing on method and costs the usual: it drifts silently, and -nothing re-checks it against the binary. +built-in `/code-review`, taken from a 2.1.220 binary. The two share their +method almost verbatim — the same ten angles, the same three-verdict verify +with "PLAUSIBLE by default", the same sweep — so the copy buys nothing on +method and costs the usual: it drifts silently, and nothing re-checks it +against the binary. + +That drift is no longer hypothetical. `claude/action.yaml` now pins 2.1.226, +which restructured the built-in without touching those texts, so the copy is +already a version behind the binary CI runs — and the thing that changed is +exactly what decides this question. Reaching the built-in is possible. It carries `disable-model-invocation`, which the `Skill` tool waives for a turn whose own user message names the diff --git a/plugins/tend-ci-runner/skills/code-review/SKILL.md b/plugins/tend-ci-runner/skills/code-review/SKILL.md index 5e544629..1becb9ab 100644 --- a/plugins/tend-ci-runner/skills/code-review/SKILL.md +++ b/plugins/tend-ci-runner/skills/code-review/SKILL.md @@ -8,7 +8,7 @@ metadata: # Code review -A structured pass over a diff that returns ranked findings. Ported from Claude Code's built-in `/code-review`, read out of the 2.1.220 binary `claude/action.yaml` pins. The Codex harness has no built-in equivalent, and on Claude the built-in carries `disable-model-invocation`, which no tend prompt lifts. This copy is tend-owned: reachable from the model on either harness. +A structured pass over a diff that returns ranked findings. Ported from Claude Code's built-in `/code-review`. The Codex harness has no built-in equivalent, and on Claude the built-in carries `disable-model-invocation`, which no tend prompt lifts. This copy is tend-owned: reachable from the model on either harness. **Return findings; don't act on them.** No review, comment, commit, or artifact from this skill — the caller folds the findings into its own single review.