Skip to content

Open source CapCut alternative.

TypeScript

chore: run reviews and fixes on the session model

David Le780edb03e026
diff --git a/.claude/agents/code-reviewer.md b/.claude/agents/code-reviewer.md
index af943df8..4147aba5 100644
--- a/.claude/agents/code-reviewer.md
+++ b/.claude/agents/code-reviewer.md
@@ -1,7 +1,6 @@
---
name: code-reviewer
-description: Reviews a diff for correctness bugs and real cleanup opportunities on Opus. Use for any review request — /review, "review this", "check my changes" — whatever model the session is on.
-model: opus
+description: Reviews a diff for correctness bugs and real cleanup opportunities. Use for any review request — /review, "review this", "check my changes".
tools: Read, Grep, Glob, Bash
---
diff --git a/.claude/skills/review/SKILL.md b/.claude/skills/review/SKILL.md
index afa8eb5a..bfcbaf31 100644
--- a/.claude/skills/review/SKILL.md
+++ b/.claude/skills/review/SKILL.md
@@ -1,10 +1,8 @@
---
name: review
-description: Review the current diff (or a branch, PR number, commit range, or path) on Opus and relay the verified findings. Use for /review, "review this", or "check my changes".
+description: Review the current diff (or a branch, PR number, commit range, or path) and relay the verified findings. Use for /review, "review this", or "check my changes".
---
-Reviews run on Opus and fixes run on Fable, whatever model the session is on. This skill puts the review in the `code-reviewer` agent, whose definition pins the model.
-
-1. Launch the review with the Agent tool: `subagent_type: "code-reviewer"`. Pass the target from `$ARGUMENTS` (default: the uncommitted working tree). Name any files another session owns so the reviewer skips them. Do not pass `subagent_type: "fork"`: a fork runs on the session model.
+1. Launch the review with the Agent tool: `subagent_type: "code-reviewer"`. Pass the target from `$ARGUMENTS` (default: the uncommitted working tree). Name any files another session owns so the reviewer skips them.
2. Wait for the agent's report. Relay every finding it confirmed, ranked most severe first, with the file and line, the defect in one sentence, and the failure scenario.
-3. If the user asks for fixes and the session is on Fable, apply them here. If the session is on another model, apply them in an agent launched with `model: "fable"`.
+3. If the user asks for fixes, apply the verified changes.
diff --git a/CLAUDE.md b/CLAUDE.md
index 4972426f..256f51ab 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -47,7 +47,6 @@ Every Cut change has to hold on every surface, and the plan for it says how:
- Do not touch repository-root `prototype/` unless the user explicitly asks for that filesystem path. By default, assume requested product changes are for the Mac app or the site/landing page.
- Ask before creating any new plan document.
-- Code reviews run on the latest Opus model; fixes and implementation run on the latest Fable model. Make the switch yourself: a review request (`/review`, "review this") runs in the `code-reviewer` agent (`.claude/agents/code-reviewer.md`, which pins `model: opus`) whatever the session model is, and a fix runs on Fable, spawned in an agent with `model: "fable"` when the session is on something else. A forked skill runs on the session model, so a review never goes through a fork. Never ask the user to `/model`.
- All writing follows `docs/guides/writing-style.md` exactly, for documentation, marketing, and every other writing surface. Read it before writing. Engineering docs under `docs/` also follow the structure in `docs/guides/eng-doc-style.md`.
- Write straight up — in prompts, docs, commits, code comments, summaries, and UI copy. State what a thing is, once, and stop. Never frame it against what it is not: no "X, not Y", no "X rather than Y", no "instead of Z", no "…, which is exactly what not to do". Cut filler.
- Keep replies short and action-oriented. For implementation questions, give the recommendation first, then one to three short bullets on why; when the answer is obvious, just say what to do. Skip long explanations, caveats, and "one last thing" sections; flag a real blocker or risk with "One issue:" and explain it briefly.