feat(review): default agent shortcuts to skip-permissions mode (#2847)

Co-authored-by: Cyrus Gray <cyrus@tinyhumans.ai>

scripts/shortcuts/review/cli.sh

@@ -37,6 +37,10 @@ Env:
   REVIEW_REPO=owner/name                  Override target repo (default: upstream remote)
   REVIEW_BANNED_COAUTHOR_RE=<regex>       Substrings filtered from Co-authored-by lines
                                           (default includes copilot/codex/cursor/claude/…)
+  REVIEW_AGENT_SAFE=1                     Run the picked agent CLI bare instead of in
+                                          its "yolo" mode. Default 0 — claude / codex /
+                                          cursor are launched with their skip-permissions
+                                          flag so headless runs don't stall on prompts.
 EOF
 }
 

scripts/shortcuts/review/coverage.sh

@@ -85,4 +85,4 @@ Additional instructions from the user:
 ${extra_prompt}"
 fi
 
-"$agent" "$prompt"
+agent_exec "$agent" "$prompt"

scripts/shortcuts/review/fix.sh

@@ -88,4 +88,4 @@ if [ -n "$extra_prompt" ]; then
 ${extra_prompt}"
 fi
 
-"$agent" "$prompt"
+agent_exec "$agent" "$prompt"

scripts/shortcuts/review/lib.sh

@@ -28,6 +28,41 @@ require() {
   done
 }
 
+# Run the picked agent CLI on a single positional prompt. Each known agent
+# is launched in its equivalent "yolo" mode so headless / detached runs
+# (CI, background tasks, tmux workers) don't stall on per-tool permission
+# prompts that have no responder. Set REVIEW_AGENT_SAFE=1 to keep the
+# prompts (e.g. an interactive local run where you want to vet each step).
+#
+# Mirrors the precedent in bin/spawn-issue, which already passes
+# --dangerously-skip-permissions to its detached claude workers, and brings
+# the claude path in line with the existing codex / cursor handling.
+agent_exec() {
+  local agent="$1"
+  local prompt="$2"
+  if [ "${REVIEW_AGENT_SAFE:-0}" = "1" ]; then
+    case "$agent" in
+      codex) codex exec "$prompt" ;;
+      *) "$agent" "$prompt" ;;
+    esac
+    return
+  fi
+  case "$agent" in
+    claude)
+      claude --dangerously-skip-permissions "$prompt"
+      ;;
+    codex)
+      codex exec --dangerously-bypass-approvals-and-sandbox "$prompt"
+      ;;
+    cursor|cursor-agent)
+      cursor-agent --yolo "$prompt"
+      ;;
+    *)
+      "$agent" "$prompt"
+      ;;
+  esac
+}
+
 gh_assign_self_issue() {
   local issue="$1"
   local repo="$2"

scripts/shortcuts/review/review.sh

@@ -69,4 +69,4 @@ if [ -n "$extra_prompt" ]; then
 ${extra_prompt}"
 fi
 
-"$agent" "$prompt"
+agent_exec "$agent" "$prompt"

scripts/shortcuts/work/start.sh

@@ -9,9 +9,13 @@
 #      repo conventions (CLAUDE.md / AGENTS.md pointers).
 #
 # --agent picks the CLI that drives the work. Default: claude.
-# `--agent codex` uses `codex exec --dangerously-bypass-approvals-and-sandbox`
+# `--agent claude` uses `claude --dangerously-skip-permissions`,
+# `--agent codex` uses `codex exec --dangerously-bypass-approvals-and-sandbox`,
 # and `--agent cursor` / `cursor-agent` use `cursor-agent --yolo`, so those
-# sessions start in their equivalent "yolo" mode.
+# sessions start in their equivalent "yolo" mode and won't stall on
+# permission prompts that have no responder in a headless context.
+# Set REVIEW_AGENT_SAFE=1 to bypass the yolo wrappers and run the agent
+# CLI bare (useful for interactive local runs where you want the prompts).
 # A trailing positional <extra-prompt> is appended to the agent prompt.
 # --no-checkout skips git sync/branch creation (use the current branch as-is).
 
@@ -169,10 +173,4 @@ ${extra_prompt}"
 fi
 
 echo "[work] handing off to ${agent} on branch ${current_branch}"
-if [ "$agent" = "codex" ]; then
-  codex exec --dangerously-bypass-approvals-and-sandbox "$prompt"
-elif [ "$agent" = "cursor" ] || [ "$agent" = "cursor-agent" ]; then
-  cursor-agent --yolo "$prompt"
-else
-  "$agent" "$prompt"
-fi
+agent_exec "$agent" "$prompt"