fix(code): make codex-review run on current codex-cli (ISS-11082)

CHANGELOG.md

@@ -4,6 +4,16 @@ All notable changes to the claude-plugins project will be documented in this fil
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Entries are listed newest-first; each plugin section is treated as released when merged to `main`.
 
+### code v1.16.4
+
+#### Fixed
+- `codex-review`'s `run_codex_review.sh` no longer passes `--full-auto`, which codex-cli 0.147 removed, so every review round exited 2. It now passes `-c sandbox_mode=read-only`, which both `codex exec` and `codex exec resume` accept; `-s read-only` would have broken every resumed round. The reviewer's sandbox narrows from `--full-auto`'s workspace-write to read-only.
+- `run_codex_review.sh` keeps codex's stderr instead of discarding it. `CODEX_FAILED` now carries one line after the exit code saying why codex failed: the last `turn.failed` (else `error`) message from the JSON stream, otherwise the first stderr line starting with `error`, otherwise the last stderr line, skipping the `Reading ... from stdin...` banner. The full stderr goes to the script's stderr, and a failed session resume names its cause before falling back to a fresh session.
+- `debate-loop.sh` prints the `CODEX_FAILED` reason with `printf '%s'` instead of `echo -e`, so backslashes in codex's message are printed as-is instead of being read as escapes.
+- `run_codex_review.sh` also passes `-c approval_policy=never`. `codex exec` defaults approval to `never` but drops that default when the user's config sets `approvals_reviewer = "auto_review"`, which left the user's `approval_policy` in effect.
+- `run_codex_review.sh` writes codex's stderr to the script's stderr when it reports `CODEX_EMPTY`, as it already does for `CODEX_FAILED`.
+- `hooks/plan-review.sh` passes `-c sandbox_mode=read-only -c approval_policy=never` instead of `--full-auto`, and appends codex's stderr to its debug log instead of discarding it.
+
 ### code-review v3.10.3
 
 #### Fixed

plugins/code/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "code",
   "description": "Code and planning framework plugin",
-  "version": "1.16.3",
+  "version": "1.16.4",
   "author": {
     "name": "ClosedLoop",
     "email": "support@closedloop.ai"

plugins/code/hooks/plan-review.sh

@@ -57,7 +57,7 @@ EOF
 
 # Get Codex's review using stdin to avoid shell escaping issues
 log "Calling codex exec..."
-REVIEW=$(codex exec --full-auto -m "gpt-5.3-codex-spark" < "$TMPFILE" 2>/dev/null)
+REVIEW=$(codex exec -c sandbox_mode=read-only -c approval_policy=never -m "gpt-5.3-codex-spark" < "$TMPFILE" 2>>"$LOG_FILE")
 
 # If codex failed, exit silently
 if [ -z "$REVIEW" ]; then

plugins/code/scripts/debate-loop.sh

@@ -501,7 +501,7 @@ while [[ $round -le $MAX_ROUNDS ]]; do
 
     # Handle failures and empty responses
     if [[ "$CODEX_VERDICT" == FAILED:* ]]; then
-      echo -e "${RED}Error: Codex failed: ${CODEX_VERDICT#FAILED:}${NC}" >&2
+      printf '%bError: Codex failed: %s%b\n' "$RED" "${CODEX_VERDICT#FAILED:}" "$NC" >&2
       exit 1
     fi
 

plugins/code/skills/codex-review/scripts/run_codex_review.sh

@@ -84,6 +84,7 @@ tmp_dir=$(mktemp -d)
 trap 'rm -rf "$tmp_dir"' EXIT
 
 codex_json="$tmp_dir/codex_output.json"
+codex_stderr="$tmp_dir/codex_stderr.txt"
 prompt_file="$tmp_dir/prompt.txt"
 
 # ── Build the review prompt ──────────────────────────────────────────────────
@@ -256,20 +257,57 @@ sys.stdout.write('\n'.join(lines))
 " "$json_file" > "$output_file" 2>/dev/null
 }
 
+# Extract the failure message from the JSON stream: the last turn.failed error,
+# else the last top-level error event. Prints it on one line, or nothing.
+parse_failure_message() {
+  python3 -c "
+import json, sys
+turn = err = ''
+for line in open(sys.argv[1]):
+    try:
+        e = json.loads(line.strip())
+        if e.get('type') == 'turn.failed':
+            turn = e['error']['message'] or turn
+        elif e.get('type') == 'error':
+            err = e['message'] or err
+    except Exception:
+        pass
+print(' '.join(str(turn or err).split()))
+" "$1" 2>/dev/null || true
+}
+
 # ── Run codex ────────────────────────────────────────────────────────────────
 
 run_codex_cmd() {
   local json_out="$1"; shift
   # Log round header
   printf '\n--- Round %s | %s ---\n' "$ROUND" "$(date -u +%Y-%m-%dT%H:%M:%SZ)" >> "$LOG_FILE"
   # Tee raw JSON stream to both the capture file and the persistent log
-  codex "$@" 2>/dev/null | tee -a "$LOG_FILE" > "$json_out"
+  codex "$@" 2>"$codex_stderr" | tee -a "$LOG_FILE" > "$json_out"
+}
+
+# One line saying why the last codex run failed. Turn failures arrive in the JSON
+# stream; CLI and config errors only on stderr, where the first line starting with
+# "error" wins, else the last line, skipping the "Reading ... stdin..." banner.
+codex_failure_reason() {
+  local reason lines
+  reason=$(parse_failure_message "$codex_json")
+  if [[ -n "$reason" ]]; then
+    echo "$reason"
+    return
+  fi
+  lines=$(tr -d '\r' < "$codex_stderr" 2>/dev/null | grep -v -e '^[[:space:]]*$' -e '^Reading .*stdin\.\.\.$') || true
+  grep -i -m1 '^error' <<<"$lines" || tail -n1 <<<"$lines"
 }
 
 effective_session_id="$SESSION_ID"
 codex_exit=0
 
-base_args=(--full-auto --json -m "$CODEX_MODEL" -c model_reasoning_effort=high)
+# `-c sandbox_mode=` rather than `--full-auto` (removed in codex-cli 0.147) or
+# `-s` (rejected by `codex exec resume`), so one arg set serves both calls.
+# approval_policy is explicit because exec drops its own `never` default when the
+# user's config sets `approvals_reviewer = "auto_review"`.
+base_args=(--json -m "$CODEX_MODEL" -c sandbox_mode=read-only -c approval_policy=never -c model_reasoning_effort=high)
 prompt_content=$(cat "$prompt_file")
 
 # Attempt session resume if we have a prior session ID
@@ -292,7 +330,8 @@ if [[ -n "$SESSION_ID" ]]; then
     # Resume succeeded -- skip to verdict extraction
     :
   else
-    echo "Codex session resume failed, starting fresh session..." >&2
+    resume_reason=$(codex_failure_reason)
+    echo "Codex session resume failed${resume_reason:+ ($resume_reason)}, starting fresh session..." >&2
     effective_session_id=""
     rm -f "$codex_json"
 
@@ -330,14 +369,17 @@ feedback_content=$(cat "$FEEDBACK_FILE" 2>/dev/null || echo "")
 
 # Handle failures
 if [[ $codex_exit -ne 0 ]] && [[ -z "$feedback_content" ]]; then
-  echo "CODEX_FAILED:codex exited with code $codex_exit"
+  cat "$codex_stderr" >&2 2>/dev/null || true
+  reason=$(codex_failure_reason)
+  echo "CODEX_FAILED:codex exited with code $codex_exit${reason:+: $reason}"
   echo "CODEX_SESSION:${effective_session_id:-none}"
   echo "LOG_ID:$LOG_ID"
   exit 0
 fi
 
 # Handle empty response
 if [[ -z "$feedback_content" ]]; then
+  cat "$codex_stderr" >&2 2>/dev/null || true
   echo "CODEX_EMPTY"
   echo "CODEX_SESSION:${effective_session_id:-none}"
   echo "LOG_ID:$LOG_ID"
@@ -354,6 +396,7 @@ elif echo "$feedback_content" | grep -q "^### Finding"; then
   echo "VERDICT:NEEDS_CHANGES"
 else
   # No verdict AND no findings -- likely truncated response, not a real review
+  cat "$codex_stderr" >&2 2>/dev/null || true
   echo "CODEX_EMPTY"
 fi