wiggum-cli 0.17.2 → 0.18.3

package/dist/templates/scripts/feature-loop.sh.tmpl

@@ -1,12 +1,14 @@
 #!/bin/bash
 # feature-loop.sh - Full feature workflow: branch -> implement -> E2E test -> PR -> review -> merge
 # Generated by ralph-cli for {{projectName}}
-# Usage: ./feature-loop.sh <feature-name> [max-iterations] [max-e2e-attempts] [--worktree] [--resume] [--model MODEL] [--review-mode MODE]
+# Usage: ./feature-loop.sh <feature-name> [max-iterations] [max-e2e-attempts] [--worktree] [--resume] [--model MODEL] [--cli CLI] [--review-cli CLI] [--review-mode MODE]
 #
 # Options:
 #   --worktree           Use git worktree for isolation (enables parallel execution)
 #   --resume             Resume an interrupted loop (reuses existing branch/worktree)
-#   --model MODEL        Claude model to use (e.g., opus, sonnet, claude-sonnet-4-6)
+#   --model MODEL        Model to use for coding/review CLI
+#   --cli CLI            Implementation CLI: 'claude' | 'codex'
+#   --review-cli CLI     Review CLI: 'claude' | 'codex'
 #   --review-mode MODE   Review mode: 'manual' (stop at PR), 'auto' (review, no merge), or 'merge' (review + merge). Default: 'manual'
 
 set -e
@@ -23,6 +25,13 @@ if [ -f "$SCRIPT_DIR/../ralph.config.cjs" ]; then
     PROMPTS_DIR=$(node -e "console.log(require('$CONFIG_PATH').paths?.prompts || '.ralph/prompts')" 2>/dev/null || echo ".ralph/prompts")
     DEFAULT_MODEL=$(node -e "console.log(require('$CONFIG_PATH').loop?.defaultModel || 'sonnet')" 2>/dev/null || echo "sonnet")
     PLANNING_MODEL=$(node -e "console.log(require('$CONFIG_PATH').loop?.planningModel || 'opus')" 2>/dev/null || echo "opus")
+    DEFAULT_CODEX_MODEL=$(node -e "console.log(require('$CONFIG_PATH').loop?.codexModel || 'gpt-5.3-codex')" 2>/dev/null || echo "gpt-5.3-codex")
+    DEFAULT_CODING_CLI=$(node -e "console.log(require('$CONFIG_PATH').loop?.codingCli || 'claude')" 2>/dev/null || echo "claude")
+    DEFAULT_REVIEW_CLI=$(node -e "console.log(require('$CONFIG_PATH').loop?.reviewCli || require('$CONFIG_PATH').loop?.codingCli || 'claude')" 2>/dev/null || echo "claude")
+    CLAUDE_PERMISSION_MODE=$(node -e "console.log(require('$CONFIG_PATH').loop?.claudePermissionMode || 'default')" 2>/dev/null || echo "default")
+    CODEX_SANDBOX=$(node -e "console.log(require('$CONFIG_PATH').loop?.codexSandbox || 'workspace-write')" 2>/dev/null || echo "workspace-write")
+    CODEX_APPROVAL_POLICY=$(node -e "console.log(require('$CONFIG_PATH').loop?.codexApprovalPolicy || 'never')" 2>/dev/null || echo "never")
+    DISABLE_MCP_IN_AUTOMATED=$(node -e "const v=require('$CONFIG_PATH').loop?.disableMcpInAutomatedRuns; console.log(v === undefined ? 'true' : String(v))" 2>/dev/null || echo "true")
     DEFAULT_MAX_ITERATIONS=$(node -e "console.log(require('$CONFIG_PATH').loop?.maxIterations || 10)" 2>/dev/null || echo "10")
     DEFAULT_MAX_E2E=$(node -e "console.log(require('$CONFIG_PATH').loop?.maxE2eAttempts || 5)" 2>/dev/null || echo "5")
     TEST_COMMAND=$(node -e "console.log(require('$CONFIG_PATH').commands?.test || 'npm test')" 2>/dev/null || echo "npm test")
@@ -34,6 +43,13 @@ elif [ -f "$SCRIPT_DIR/../../ralph.config.cjs" ]; then
     PROMPTS_DIR=$(node -e "console.log(require('$CONFIG_PATH').paths?.prompts || '.ralph/prompts')" 2>/dev/null || echo ".ralph/prompts")
     DEFAULT_MODEL=$(node -e "console.log(require('$CONFIG_PATH').loop?.defaultModel || 'sonnet')" 2>/dev/null || echo "sonnet")
     PLANNING_MODEL=$(node -e "console.log(require('$CONFIG_PATH').loop?.planningModel || 'opus')" 2>/dev/null || echo "opus")
+    DEFAULT_CODEX_MODEL=$(node -e "console.log(require('$CONFIG_PATH').loop?.codexModel || 'gpt-5.3-codex')" 2>/dev/null || echo "gpt-5.3-codex")
+    DEFAULT_CODING_CLI=$(node -e "console.log(require('$CONFIG_PATH').loop?.codingCli || 'claude')" 2>/dev/null || echo "claude")
+    DEFAULT_REVIEW_CLI=$(node -e "console.log(require('$CONFIG_PATH').loop?.reviewCli || require('$CONFIG_PATH').loop?.codingCli || 'claude')" 2>/dev/null || echo "claude")
+    CLAUDE_PERMISSION_MODE=$(node -e "console.log(require('$CONFIG_PATH').loop?.claudePermissionMode || 'default')" 2>/dev/null || echo "default")
+    CODEX_SANDBOX=$(node -e "console.log(require('$CONFIG_PATH').loop?.codexSandbox || 'workspace-write')" 2>/dev/null || echo "workspace-write")
+    CODEX_APPROVAL_POLICY=$(node -e "console.log(require('$CONFIG_PATH').loop?.codexApprovalPolicy || 'never')" 2>/dev/null || echo "never")
+    DISABLE_MCP_IN_AUTOMATED=$(node -e "const v=require('$CONFIG_PATH').loop?.disableMcpInAutomatedRuns; console.log(v === undefined ? 'true' : String(v))" 2>/dev/null || echo "true")
     DEFAULT_MAX_ITERATIONS=$(node -e "console.log(require('$CONFIG_PATH').loop?.maxIterations || 10)" 2>/dev/null || echo "10")
     DEFAULT_MAX_E2E=$(node -e "console.log(require('$CONFIG_PATH').loop?.maxE2eAttempts || 5)" 2>/dev/null || echo "5")
     TEST_COMMAND=$(node -e "console.log(require('$CONFIG_PATH').commands?.test || 'npm test')" 2>/dev/null || echo "npm test")
@@ -45,6 +61,13 @@ else
     PROMPTS_DIR=".ralph/prompts"
     DEFAULT_MODEL="sonnet"
     PLANNING_MODEL="opus"
+    DEFAULT_CODEX_MODEL="gpt-5.3-codex"
+    DEFAULT_CODING_CLI="claude"
+    DEFAULT_REVIEW_CLI="claude"
+    CLAUDE_PERMISSION_MODE="default"
+    CODEX_SANDBOX="workspace-write"
+    CODEX_APPROVAL_POLICY="never"
+    DISABLE_MCP_IN_AUTOMATED="true"
     DEFAULT_MAX_ITERATIONS="10"
     DEFAULT_MAX_E2E="5"
     TEST_COMMAND="npm test"
@@ -59,6 +82,8 @@ USE_WORKTREE=false
 RESUME=false
 MODEL=""
 REVIEW_MODE=""
+CLI_OVERRIDE=""
+REVIEW_CLI_OVERRIDE=""
 POSITIONAL=()
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -74,6 +99,14 @@ while [[ $# -gt 0 ]]; do
             MODEL="$2"
             shift 2
             ;;
+        --cli)
+            CLI_OVERRIDE="$2"
+            shift 2
+            ;;
+        --review-cli)
+            REVIEW_CLI_OVERRIDE="$2"
+            shift 2
+            ;;
         --review-mode)
             REVIEW_MODE="$2"
             shift 2
@@ -117,9 +150,168 @@ if [ "$REVIEW_MODE" != "manual" ] && [ "$REVIEW_MODE" != "auto" ] && [ "$REVIEW_
     exit 1
 fi
 
-# Build claude commands
-CLAUDE_CMD_OPUS="claude -p --output-format json --dangerously-skip-permissions --model ${PLANNING_MODEL}"
-CLAUDE_CMD_IMPL="claude -p --output-format json --dangerously-skip-permissions --model ${MODEL:-$DEFAULT_MODEL}"
+# Resolve coding/review CLI from CLI > config > default
+CODING_CLI="${CLI_OVERRIDE:-$DEFAULT_CODING_CLI}"
+REVIEW_CLI="${REVIEW_CLI_OVERRIDE:-${DEFAULT_REVIEW_CLI:-$CODING_CLI}}"
+DISABLE_MCP_IN_AUTOMATED_NORM=$(echo "$DISABLE_MCP_IN_AUTOMATED" | tr '[:upper:]' '[:lower:]')
+
+# Validate CLI values
+if [ "$CODING_CLI" != "claude" ] && [ "$CODING_CLI" != "codex" ]; then
+    echo "ERROR: Invalid --cli value '$CODING_CLI'. Allowed values are 'claude' or 'codex'." >&2
+    exit 1
+fi
+
+if [ "$REVIEW_CLI" != "claude" ] && [ "$REVIEW_CLI" != "codex" ]; then
+    echo "ERROR: Invalid --review-cli value '$REVIEW_CLI'. Allowed values are 'claude' or 'codex'." >&2
+    exit 1
+fi
+
+is_valid_claude_permission_mode() {
+    case "$1" in
+        acceptEdits|bypassPermissions|default|dontAsk|plan|auto) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+is_valid_codex_sandbox() {
+    case "$1" in
+        read-only|workspace-write|danger-full-access) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+is_valid_codex_approval_policy() {
+    case "$1" in
+        untrusted|on-failure|on-request|never) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+if ! is_valid_claude_permission_mode "$CLAUDE_PERMISSION_MODE"; then
+    echo "ERROR: Invalid loop.claudePermissionMode '$CLAUDE_PERMISSION_MODE' in ralph.config.cjs." >&2
+    exit 1
+fi
+
+if ! is_valid_codex_sandbox "$CODEX_SANDBOX"; then
+    echo "ERROR: Invalid loop.codexSandbox '$CODEX_SANDBOX' in ralph.config.cjs." >&2
+    exit 1
+fi
+
+if ! is_valid_codex_approval_policy "$CODEX_APPROVAL_POLICY"; then
+    echo "ERROR: Invalid loop.codexApprovalPolicy '$CODEX_APPROVAL_POLICY' in ralph.config.cjs." >&2
+    exit 1
+fi
+
+case "$DISABLE_MCP_IN_AUTOMATED_NORM" in
+    true|false) ;;
+    *)
+        echo "ERROR: Invalid loop.disableMcpInAutomatedRuns '$DISABLE_MCP_IN_AUTOMATED' in ralph.config.cjs. Use true or false." >&2
+        exit 1
+        ;;
+esac
+
+is_claude_only_model() {
+    local candidate="$1"
+    case "$candidate" in
+        sonnet|opus|haiku|claude-*) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+resolve_codex_model() {
+    local candidate="${MODEL:-$DEFAULT_CODEX_MODEL}"
+    if is_claude_only_model "$candidate"; then
+        echo "$DEFAULT_CODEX_MODEL"
+    else
+        echo "$candidate"
+    fi
+}
+
+if [ -n "$MODEL" ] && { [ "$CODING_CLI" = "codex" ] || [ "$REVIEW_CLI" = "codex" ]; }; then
+    if is_claude_only_model "$MODEL"; then
+        echo "WARNING: --model '$MODEL' is Claude-specific. Codex phases will use '$DEFAULT_CODEX_MODEL'." >&2
+    fi
+fi
+
+build_cli_cmd() {
+    local cli="$1"
+    local model="$2"
+    if [[ ! "$model" =~ ^[A-Za-z0-9._:/=-]+$ ]]; then
+        echo "ERROR: Invalid model value '$model'. Only alphanumeric, dot, underscore, colon, slash, equals and hyphen are allowed." >&2
+        return 1
+    fi
+    case "$cli" in
+        claude)
+            echo "claude -p --output-format json --permission-mode ${CLAUDE_PERMISSION_MODE} --model ${model}"
+            ;;
+        codex)
+            local codex_extra=""
+            # Avoid MCP startup deadlocks in unattended loop runs.
+            if [ "${RALPH_AUTOMATED:-}" = "1" ] && [ "$DISABLE_MCP_IN_AUTOMATED_NORM" = "true" ]; then
+                codex_extra=" -c mcp_servers={}"
+            fi
+            # Keep command string token-safe. run_claude_* parses it as an argument array.
+            echo "codex --ask-for-approval ${CODEX_APPROVAL_POLICY} --sandbox ${CODEX_SANDBOX} exec --cd . --model ${model}${codex_extra}"
+            ;;
+        *)
+            echo "ERROR: Unsupported CLI '$cli'" >&2
+            return 1
+            ;;
+    esac
+}
+
+get_phase_cli() {
+    local phase="$1"
+    case "$phase" in
+        review)
+            echo "$REVIEW_CLI"
+            ;;
+        *)
+            echo "$CODING_CLI"
+            ;;
+    esac
+}
+
+get_phase_model() {
+    local phase="$1"
+    local cli
+    cli=$(get_phase_cli "$phase")
+    if [ "$cli" = "codex" ]; then
+        resolve_codex_model
+        return
+    fi
+
+    case "$phase" in
+        planning|review)
+            echo "$PLANNING_MODEL"
+            ;;
+        *)
+            echo "${MODEL:-$DEFAULT_MODEL}"
+            ;;
+    esac
+}
+
+get_phase_cmd() {
+    local phase="$1"
+    local cli
+    local model
+    cli=$(get_phase_cli "$phase")
+    model=$(get_phase_model "$phase")
+    build_cli_cmd "$cli" "$model"
+}
+
+check_cli_binary() {
+    local cli="$1"
+    local install_hint=""
+    case "$cli" in
+        claude) install_hint="npm install -g @anthropic-ai/claude-code" ;;
+        codex) install_hint="npm install -g @openai/codex" ;;
+    esac
+    if ! command -v "$cli" >/dev/null 2>&1; then
+        echo "ERROR: ${cli} CLI not found. Install with: ${install_hint}" >&2
+        exit 1
+    fi
+}
 
 # Automation footer appended to every prompt in automated mode.
 # Prevents interactive skill prompts from blocking headless sessions.
@@ -141,22 +333,69 @@ This is a fully automated session with no human operator. You MUST:
 '
 fi
 
-# Helper: pipe prompt with automation footer to claude
+# Helper: pipe prompt with automation footer to selected CLI command
 run_claude_prompt() {
     local prompt_file="$1"
     local claude_cmd="$2"
-    { cat "$prompt_file" | envsubst; echo "$AUTOMATION_FOOTER"; } | $claude_cmd
+    local -a cmd_parts=()
+    read -r -a cmd_parts <<< "$claude_cmd"
+    if [[ "${cmd_parts[0]:-}" == "codex" ]]; then
+        LAST_RUN_CLI="codex"
+        { cat "$prompt_file" | envsubst; echo "$AUTOMATION_FOOTER"; } | (cd "$APP_DIR" && "${cmd_parts[@]}" --json --output-last-message "$LAST_MESSAGE_FILE" -)
+    else
+        LAST_RUN_CLI="claude"
+        { cat "$prompt_file" | envsubst; echo "$AUTOMATION_FOOTER"; } | "${cmd_parts[@]}"
+    fi
+}
+
+# Helper: resume an existing session with a short continuation prompt
+run_claude_resume() {
+    local session_id="$1"
+    local continuation_prompt="$2"
+    local claude_cmd="$3"
+    if [[ ! "$session_id" =~ ^[A-Za-z0-9._:-]+$ ]]; then
+        echo "WARNING: Refusing to resume with unsafe session id '$session_id'" >&2
+        return 1
+    fi
+    if [[ "$claude_cmd" == codex* ]]; then
+        LAST_RUN_CLI="codex"
+        local resume_cmd="${claude_cmd/ exec / exec resume }"
+        if [ "$resume_cmd" = "$claude_cmd" ]; then
+            echo "WARNING: codex resume injection failed, exec segment not found in command" >&2
+            return 1
+        fi
+        # codex exec resume does not accept -C/--cd; resume from APP_DIR instead.
+        resume_cmd="${resume_cmd/ --cd ./}"
+        resume_cmd="${resume_cmd/ -C ./}"
+        local -a resume_parts=()
+        read -r -a resume_parts <<< "$resume_cmd"
+        { echo "$continuation_prompt"; echo "$AUTOMATION_FOOTER"; } | (cd "$APP_DIR" && "${resume_parts[@]}" "$session_id" - --json --output-last-message "$LAST_MESSAGE_FILE")
+    else
+        LAST_RUN_CLI="claude"
+        # Insert --resume before the -p flag.
+        local resume_cmd="${claude_cmd/ -p / --resume ${session_id} -p }"
+        if [ "$resume_cmd" = "$claude_cmd" ]; then
+            echo "WARNING: --resume injection failed, -p flag not found in command" >&2
+            return 1
+        fi
+        local -a resume_parts=()
+        read -r -a resume_parts <<< "$resume_cmd"
+        { echo "$continuation_prompt"; echo "$AUTOMATION_FOOTER"; } | "${resume_parts[@]}"
+    fi
 }
 
 # Token tracking
 TOKENS_FILE="/tmp/ralph-loop-${1}.tokens"
 CLAUDE_OUTPUT="/tmp/ralph-loop-${1}.output"
+LAST_MESSAGE_FILE="/tmp/ralph-loop-${1}.last-message"
 STATUS_FILE="/tmp/ralph-loop-${1}.status"
 FINAL_STATUS_FILE="/tmp/ralph-loop-${1}.final"
 PHASES_FILE="/tmp/ralph-loop-${1}.phases"
 BASELINE_FILE="/tmp/ralph-loop-${1}.baseline"
+PRE_RUN_DIRTY_FILE="/tmp/ralph-loop-${1}.dirty"
 SESSIONS_FILE="/tmp/ralph-loop-${1}.sessions"
 LOG_FILE="/tmp/ralph-loop-${1}.log"
+LAST_RUN_CLI=""
 
 # Initialize token tracking (4-field format: input|output|cache_create|cache_read)
 init_tokens() {
@@ -165,15 +404,85 @@ init_tokens() {
     > "$LOG_FILE"
 }
 
-# Extract session result from JSON output.
+# Extract session result from command output.
 # Writes human-readable result text to the .log file and captures session_id.
-# Usage: extract_session_result <raw_json_file>
+# Usage: extract_session_result <raw_file> [cli]
 # Sets: LAST_SESSION_ID variable
 extract_session_result() {
     local raw_file="$1"
+    local cli="${2:-$LAST_RUN_CLI}"
     LAST_SESSION_ID=""
     if [ ! -f "$raw_file" ]; then return; fi
 
+    if [ "$cli" = "codex" ]; then
+        local result
+        result=$(python3 -c "
+import json, sys
+session = ''
+for line in open(sys.argv[1], encoding='utf-8', errors='ignore'):
+    line = line.strip()
+    if not line:
+        continue
+    try:
+        obj = json.loads(line)
+    except Exception:
+        continue
+    stack = [obj]
+    while stack:
+        cur = stack.pop()
+        if isinstance(cur, dict):
+            for key in ('session_id', 'sessionId', 'conversation_id', 'conversationId', 'thread_id', 'threadId', 'response_id', 'responseId', 'run_id', 'runId'):
+                val = cur.get(key)
+                if isinstance(val, str) and val:
+                    session = val
+            # Newer Codex JSON can nest thread/session identifiers under typed objects.
+            node_type = cur.get('type')
+            if node_type in ('thread.started', 'session.started'):
+                val = cur.get('id')
+                if isinstance(val, str) and val:
+                    session = val
+            thread_obj = cur.get('thread')
+            if isinstance(thread_obj, dict):
+                val = thread_obj.get('id')
+                if isinstance(val, str) and val:
+                    session = val
+            for val in cur.values():
+                if isinstance(val, (dict, list)):
+                    stack.append(val)
+        elif isinstance(cur, list):
+            for val in cur:
+                if isinstance(val, (dict, list)):
+                    stack.append(val)
+print(session)
+" "$raw_file" 2>/dev/null) || true
+
+        LAST_SESSION_ID="$result"
+        if [ -n "$LAST_SESSION_ID" ]; then
+            echo "$LAST_SESSION_ID" >> "$SESSIONS_FILE"
+        fi
+
+        if [ -f "$LAST_MESSAGE_FILE" ]; then
+            cat "$LAST_MESSAGE_FILE" >> "$LOG_FILE" 2>/dev/null || true
+            echo "" >> "$LOG_FILE"
+        else
+            python3 -c "
+import json, sys
+for line in open(sys.argv[1], encoding='utf-8', errors='ignore'):
+    try:
+        obj = json.loads(line)
+    except Exception:
+        continue
+    if not isinstance(obj, dict):
+        continue
+    for key in ('output_text', 'text', 'content'):
+        val = obj.get(key)
+        if isinstance(val, str) and val.strip():
+            print(val.strip())
+" "$raw_file" >> "$LOG_FILE" 2>/dev/null || true
+        fi
+        return
+    fi
+
     local result
     result=$(python3 -c "
 import json, sys
@@ -210,29 +519,109 @@ except Exception:
 " "$raw_file" >> "$LOG_FILE" 2>/dev/null || true
 }
 
-# Accumulate tokens from a session JSONL file into the .tokens file.
-# Usage: accumulate_tokens_from_session <session_id>
+# Accumulate tokens into the .tokens file.
+# Usage: accumulate_tokens_from_session <session_id> [raw_file] [cli]
 accumulate_tokens_from_session() {
     local session_id="$1"
-    if [ -z "$session_id" ]; then return; fi
+    local raw_file="${2:-}"
+    local cli="${3:-$LAST_RUN_CLI}"
 
-    # Find the JSONL file for this session
-    local jsonl_file=""
-    for f in ~/.claude/projects/*/"${session_id}.jsonl"; do
-        if [ -f "$f" ]; then
-            jsonl_file="$f"
-            break
+    local s_input=0
+    local s_output=0
+    local s_cache_create=0
+    local s_cache_read=0
+
+    if [ "$cli" = "codex" ]; then
+        if [ -z "$raw_file" ] || [ ! -f "$raw_file" ]; then
+            return
         fi
-    done
 
-    if [ -z "$jsonl_file" ]; then
-        echo "WARNING: Could not find JSONL for session $session_id" >&2
-        return
-    fi
+        local session_tokens
+        session_tokens=$(python3 -c "
+import json, sys
+
+def to_int(v):
+    try:
+        return int(v)
+    except Exception:
+        return 0
+
+def usage_pair(usage):
+    input_tokens = (
+        to_int(usage.get('input_tokens'))
+        or to_int(usage.get('inputTokens'))
+        or to_int(usage.get('prompt_tokens'))
+        or to_int(usage.get('promptTokens'))
+    )
+    output_tokens = (
+        to_int(usage.get('output_tokens'))
+        or to_int(usage.get('outputTokens'))
+        or to_int(usage.get('completion_tokens'))
+        or to_int(usage.get('completionTokens'))
+    )
+    return input_tokens, output_tokens
+
+# Codex JSONL often contains repeated/cumulative usage in multiple events.
+# Use the highest observed values from a single run to avoid overcounting.
+max_input = 0
+max_output = 0
+
+for line in open(sys.argv[1], encoding='utf-8', errors='ignore'):
+    line = line.strip()
+    if not line:
+        continue
+    try:
+        obj = json.loads(line)
+    except Exception:
+        continue
+    stack = [obj]
+    while stack:
+        cur = stack.pop()
+        if isinstance(cur, dict):
+            if 'usage' in cur and isinstance(cur['usage'], dict):
+                usage = cur['usage']
+                u_in, u_out = usage_pair(usage)
+                if u_in > max_input:
+                    max_input = u_in
+                if u_out > max_output:
+                    max_output = u_out
+            for val in cur.values():
+                if isinstance(val, (dict, list)):
+                    stack.append(val)
+        elif isinstance(cur, list):
+            for val in cur:
+                if isinstance(val, (dict, list)):
+                    stack.append(val)
+
+print(f\"{max_input}|{max_output}|0|0\")
+" "$raw_file" 2>/dev/null) || true
+
+        if [ -n "$session_tokens" ]; then
+            s_input=$(echo "$session_tokens" | cut -d'|' -f1)
+            s_output=$(echo "$session_tokens" | cut -d'|' -f2)
+            s_cache_create=0
+            s_cache_read=0
+        fi
+    else
+        if [ -z "$session_id" ]; then return; fi
+
+        # Find the JSONL file for this session
+        local jsonl_file=""
+        for f in ~/.claude/projects/*/"${session_id}.jsonl"; do
+            if [ -f "$f" ]; then
+                jsonl_file="$f"
+                break
+            fi
+        done
+
+        if [ -z "$jsonl_file" ]; then
+            echo "WARNING: Could not find JSONL for session $session_id" >&2
+            return
+        fi
 
-    # Extract and sum token usage from all assistant messages
-    local session_tokens
-    session_tokens=$(python3 -c "
+        # Extract and sum token usage from all assistant messages
+        local session_tokens
+        session_tokens=$(python3 -c "
 import json, sys
 totals = {'input': 0, 'output': 0, 'cache_create': 0, 'cache_read': 0}
 for line in open(sys.argv[1]):
@@ -252,14 +641,19 @@ for line in open(sys.argv[1]):
 print(f\"{totals['input']}|{totals['output']}|{totals['cache_create']}|{totals['cache_read']}\")
 " "$jsonl_file" 2>/dev/null) || true
 
-    if [ -z "$session_tokens" ]; then return; fi
+        if [ -z "$session_tokens" ]; then return; fi
 
-    # Parse session tokens
-    local s_input s_output s_cache_create s_cache_read
-    s_input=$(echo "$session_tokens" | cut -d'|' -f1)
-    s_output=$(echo "$session_tokens" | cut -d'|' -f2)
-    s_cache_create=$(echo "$session_tokens" | cut -d'|' -f3)
-    s_cache_read=$(echo "$session_tokens" | cut -d'|' -f4)
+        # Parse session tokens
+        s_input=$(echo "$session_tokens" | cut -d'|' -f1)
+        s_output=$(echo "$session_tokens" | cut -d'|' -f2)
+        s_cache_create=$(echo "$session_tokens" | cut -d'|' -f3)
+        s_cache_read=$(echo "$session_tokens" | cut -d'|' -f4)
+    fi
+
+    [[ "$s_input" =~ ^[0-9]+$ ]] || s_input=0
+    [[ "$s_output" =~ ^[0-9]+$ ]] || s_output=0
+    [[ "$s_cache_create" =~ ^[0-9]+$ ]] || s_cache_create=0
+    [[ "$s_cache_read" =~ ^[0-9]+$ ]] || s_cache_read=0
 
     # Read current totals
     local current c_input c_output c_cache_create c_cache_read
@@ -396,7 +790,51 @@ detect_plan_format() {
     echo "unknown"
 }
 
-# Check if any tracked files were changed since baseline.
+# Snapshot tracked dirty files before loop execution starts.
+# Each entry stores <path>\t<working-tree blob hash or __MISSING__ marker>.
+capture_pre_run_dirty_snapshot() {
+    local baseline="$1"
+    > "$PRE_RUN_DIRTY_FILE"
+    if [ -z "$baseline" ]; then
+        return
+    fi
+    python3 - "$baseline" "$PRE_RUN_DIRTY_FILE" <<'PY' 2>/dev/null || true
+import os
+import subprocess
+import sys
+
+baseline = sys.argv[1]
+output_path = sys.argv[2]
+
+try:
+    changed = subprocess.check_output(
+        ["git", "diff", "--name-only", baseline],
+        stderr=subprocess.DEVNULL,
+    ).splitlines()
+except Exception:
+    changed = []
+
+with open(output_path, "w", encoding="utf-8", errors="surrogateescape") as fh:
+    for raw_path in changed:
+        if not raw_path:
+            continue
+        path = raw_path.decode("utf-8", errors="surrogateescape")
+        marker = "__MISSING__"
+        if os.path.exists(path):
+            try:
+                marker = subprocess.check_output(
+                    ["git", "hash-object", "--", path],
+                    stderr=subprocess.DEVNULL,
+                    text=True,
+                ).strip()
+            except Exception:
+                marker = "__HASH_ERROR__"
+        fh.write(f"{path}\t{marker}\n")
+PY
+}
+
+# Check if any tracked files were changed since baseline (excluding unchanged
+# tracked edits that already existed before the loop started).
 # Counts ALL file types (code, docs, config) — not just source code.
 count_file_changes() {
     local baseline="$1"
@@ -405,14 +843,90 @@ count_file_changes() {
         return
     fi
     local count
-    count=$(git diff --name-only "${baseline}..HEAD" 2>/dev/null | wc -l | tr -d ' ') || count=0
+    # Compare baseline commit against working tree; subtract pre-run dirty files
+    # unless that specific file changed again after loop start.
+    count=$(python3 - "$baseline" "$PRE_RUN_DIRTY_FILE" <<'PY' 2>/dev/null || echo "0"
+import os
+import subprocess
+import sys
+
+baseline = sys.argv[1]
+snapshot_path = sys.argv[2]
+snapshot = {}
+
+if os.path.exists(snapshot_path):
+    with open(snapshot_path, "r", encoding="utf-8", errors="surrogateescape") as fh:
+        for line in fh:
+            line = line.rstrip("\n")
+            if not line or "\t" not in line:
+                continue
+            path, marker = line.split("\t", 1)
+            snapshot[path] = marker
+
+try:
+    changed = subprocess.check_output(
+        ["git", "diff", "--name-only", baseline],
+        stderr=subprocess.DEVNULL,
+    ).splitlines()
+except Exception:
+    print("0")
+    raise SystemExit(0)
+
+count = 0
+for raw_path in changed:
+    if not raw_path:
+        continue
+    path = raw_path.decode("utf-8", errors="surrogateescape")
+    start_marker = snapshot.get(path)
+    if start_marker is None:
+        count += 1
+        continue
+
+    current_marker = "__MISSING__"
+    if os.path.exists(path):
+        try:
+            current_marker = subprocess.check_output(
+                ["git", "hash-object", "--", path],
+                stderr=subprocess.DEVNULL,
+                text=True,
+            ).strip()
+        except Exception:
+            current_marker = "__HASH_ERROR__"
+    if current_marker != start_marker:
+        count += 1
+
+print(count)
+PY
+)
     echo "$count"
 }
 
-# Extract review findings text from a Claude JSON output file.
+# Extract review findings text from command output.
 # Returns the result text from the last result entry.
 extract_review_findings() {
     local raw_file="$1"
+    local cli="${2:-$LAST_RUN_CLI}"
+    if [ "$cli" = "codex" ]; then
+        if [ -f "$LAST_MESSAGE_FILE" ]; then
+            cat "$LAST_MESSAGE_FILE" 2>/dev/null || echo "No review output available"
+            return
+        fi
+        python3 -c "
+import json, sys
+for line in open(sys.argv[1], encoding='utf-8', errors='ignore'):
+    try:
+        obj = json.loads(line)
+    except Exception:
+        continue
+    if isinstance(obj, dict):
+        for key in ('output_text', 'text', 'content'):
+            val = obj.get(key)
+            if isinstance(val, str) and val.strip():
+                print(val.strip())
+" "$raw_file" 2>/dev/null || echo "No review output available"
+        return
+    fi
+
     python3 -c "
 import json, sys
 try:
@@ -428,11 +942,16 @@ except Exception:
 }
 
 # Run a fix iteration based on code review findings.
-# Pipes the review output into Claude for targeted fixes.
+# Pipes the review output into the implementation CLI for targeted fixes.
 run_review_fix() {
     local findings
+    local impl_cli="$CODING_CLI"
     findings=$(extract_review_findings "${CLAUDE_OUTPUT}.raw")
-    cat <<FIXEOF | $CLAUDE_CMD_IMPL 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+    local -a impl_cmd_parts=()
+    read -r -a impl_cmd_parts <<< "$IMPL_CMD"
+    if [ "$impl_cli" = "codex" ]; then
+        LAST_RUN_CLI="codex"
+        cat <<FIXEOF | "${impl_cmd_parts[@]}" --json --output-last-message "$LAST_MESSAGE_FILE" - 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
 ## Code Review Findings
 
 The following issues were found during code review:
@@ -447,8 +966,51 @@ Fix each issue listed above. Run git diff $DEFAULT_BRANCH to see the current cha
 3. Commit and push the fixes
 Do NOT propose completion options or ask interactive questions. Just fix, test, commit, push.
 FIXEOF
-    extract_session_result "${CLAUDE_OUTPUT}.raw"
-    accumulate_tokens_from_session "$LAST_SESSION_ID"
+    else
+        LAST_RUN_CLI="claude"
+        cat <<FIXEOF | "${impl_cmd_parts[@]}" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+## Code Review Findings
+
+The following issues were found during code review:
+
+${findings}
+
+## Task
+
+Fix each issue listed above. Run git diff $DEFAULT_BRANCH to see the current changes, then:
+1. Fix each issue referenced in the review
+2. Run tests to verify fixes
+3. Commit and push the fixes
+Do NOT propose completion options or ask interactive questions. Just fix, test, commit, push.
+FIXEOF
+    fi
+    extract_session_result "${CLAUDE_OUTPUT}.raw" "$impl_cli"
+    accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$impl_cli"
+}
+
+validate_simple_command() {
+    local command="$1"
+    if [[ "$command" =~ [\;\&\|\<\>\`\$\(\)] ]]; then
+        return 1
+    fi
+    if [[ "$command" =~ [[:cntrl:]] ]]; then
+        return 1
+    fi
+    return 0
+}
+
+run_test_command() {
+    if ! validate_simple_command "$TEST_COMMAND"; then
+        echo "ERROR: commands.test contains unsupported shell operators. Use a plain command and arguments." >&2
+        return 2
+    fi
+    local -a test_cmd_parts=()
+    read -r -a test_cmd_parts <<< "$TEST_COMMAND"
+    if [ ${#test_cmd_parts[@]} -eq 0 ]; then
+        echo "ERROR: commands.test resolved to an empty command." >&2
+        return 2
+    fi
+    (cd "$APP_DIR" && "${test_cmd_parts[@]}")
 }
 
 # Normalize test failure lines: extract test name, strip timing, deduplicate.
@@ -461,7 +1023,7 @@ normalize_test_failures() {
 # Returns 0 if tests pass OR all failures are pre-existing (captured at baseline).
 check_tests_pass_or_baseline() {
     local test_output
-    test_output=$( (cd "$APP_DIR" && eval "$TEST_COMMAND") 2>&1 )
+    test_output=$(run_test_command 2>&1)
     local exit_code=$?
 
     if [ $exit_code -eq 0 ]; then
@@ -523,7 +1085,7 @@ write_phase_end() {
 # Initialize phase tracking
 > "$PHASES_FILE"
 
-FEATURE="${1:?Usage: ./feature-loop.sh <feature-name> [max-iterations] [max-e2e-attempts] [--worktree] [--resume] [--model MODEL]}"
+FEATURE="${1:?Usage: ./feature-loop.sh <feature-name> [max-iterations] [max-e2e-attempts] [--worktree] [--resume] [--model MODEL] [--cli CLI] [--review-cli CLI] [--review-mode MODE]}"
 # Sanitize feature name to prevent path traversal and shell injection when used in temp file paths
 if [[ ! "$FEATURE" =~ ^[a-zA-Z0-9][a-zA-Z0-9_-]*$ ]]; then
     echo "ERROR: Feature name must start with alphanumeric and contain only letters, numbers, hyphens, and underscores." >&2
@@ -538,6 +1100,15 @@ SPEC_FILE="$SPEC_DIR/${FEATURE}.md"
 PLAN_FILE="$SPEC_DIR/${FEATURE}-implementation-plan.md"
 BRANCH="feat/${FEATURE}"
 APP_DIR="$(pwd)"
+PLANNING_CMD=$(get_phase_cmd "planning")
+IMPL_CMD=$(get_phase_cmd "implementation")
+REVIEW_CMD=$(get_phase_cmd "review")
+
+# Fail fast if required CLIs are not installed.
+check_cli_binary "$CODING_CLI"
+if [ "$REVIEW_CLI" != "$CODING_CLI" ]; then
+    check_cli_binary "$REVIEW_CLI"
+fi
 
 echo "=========================================="
 echo "Ralph Loop: $FEATURE"
@@ -547,9 +1118,26 @@ echo "Branch: $BRANCH"
 echo "App dir: $APP_DIR"
 echo "Worktree mode: $USE_WORKTREE"
 echo "Resume mode: $RESUME"
+echo "Coding CLI (impl/e2e): $CODING_CLI"
+echo "Review CLI: $REVIEW_CLI"
 echo "Review mode: $REVIEW_MODE"
-echo "Model (planning): $PLANNING_MODEL"
-echo "Model (impl): ${MODEL:-$DEFAULT_MODEL}"
+echo "Claude permission mode: $CLAUDE_PERMISSION_MODE"
+echo "Codex sandbox: $CODEX_SANDBOX"
+echo "Codex approval policy: $CODEX_APPROVAL_POLICY"
+if [ "${RALPH_AUTOMATED:-}" = "1" ]; then
+    echo "Disable MCP in automated runs: $DISABLE_MCP_IN_AUTOMATED_NORM"
+fi
+if [ "$CODING_CLI" = "codex" ] && [ "$REVIEW_CLI" = "codex" ]; then
+    echo "Model (all phases): $(resolve_codex_model)"
+else
+    if [ "$CODING_CLI" = "claude" ] || [ "$REVIEW_CLI" = "claude" ]; then
+        echo "Model (Claude planning/review): $PLANNING_MODEL"
+        echo "Model (Claude impl/e2e): ${MODEL:-$DEFAULT_MODEL}"
+    fi
+    if [ "$CODING_CLI" = "codex" ] || [ "$REVIEW_CLI" = "codex" ]; then
+        echo "Model (Codex phases): $(resolve_codex_model)"
+    fi
+fi
 echo "Max iterations: $MAX_ITERATIONS"
 echo "Max E2E attempts: $MAX_E2E_ATTEMPTS"
 echo "=========================================="
@@ -588,7 +1176,7 @@ fi
 # tasks to be checked — the checkboxes may be stale if the work shipped under a
 # different branch name that never updated this plan file.
 if [ -f "$PLAN_FILE" ]; then
-    _DIFF_STAT=$(git diff "$DEFAULT_BRANCH..HEAD" --stat 2>/dev/null || echo "")
+    _DIFF_STAT=$(git diff "${DEFAULT_BRANCH}..HEAD" --stat 2>/dev/null || echo "")
     if [ -z "$_DIFF_STAT" ]; then
         echo "Plan exists but branch has no diff to $DEFAULT_BRANCH — work already merged."
         > "$PHASES_FILE"
@@ -612,17 +1200,18 @@ if git rev-parse --git-dir > /dev/null 2>&1; then
     if [ -n "$BASELINE_COMMIT" ]; then
         echo "$BASELINE_COMMIT" > "$BASELINE_FILE"
         echo "Baseline commit: $BASELINE_COMMIT"
+        capture_pre_run_dirty_snapshot "$BASELINE_COMMIT"
     fi
 fi
 
 # Capture baseline test failures for pre-existing failure detection
 BASELINE_FAILURES_FILE="/tmp/ralph-loop-${FEATURE}.baseline-failures"
 echo "Capturing baseline test failures..."
-if (cd "$APP_DIR" && eval "$TEST_COMMAND" 2>&1) > /dev/null 2>&1; then
+if run_test_command > /dev/null 2>&1; then
     echo "Baseline: all tests passing"
     : > "$BASELINE_FAILURES_FILE"
 else
-    (cd "$APP_DIR" && eval "$TEST_COMMAND" 2>&1) | normalize_test_failures > "$BASELINE_FAILURES_FILE" 2>/dev/null || true
+    run_test_command 2>&1 | normalize_test_failures > "$BASELINE_FAILURES_FILE" 2>/dev/null || true
     BASELINE_COUNT=$(wc -l < "$BASELINE_FAILURES_FILE" | tr -d ' ')
     echo "Baseline: $BASELINE_COUNT pre-existing test failure(s) recorded"
 fi
@@ -634,14 +1223,14 @@ echo "0|$MAX_ITERATIONS|$(date +%s)" > "$STATUS_FILE"
 if [ ! -f "$PLAN_FILE" ]; then
     echo "======================== PLANNING PHASE ========================"
     write_phase_start "planning"
-    export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR
-    run_claude_prompt "$PROMPTS_DIR/PROMPT_feature.md" "$CLAUDE_CMD_OPUS" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || {
+    export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR CODING_CLI REVIEW_CLI REVIEW_MODE
+    run_claude_prompt "$PROMPTS_DIR/PROMPT_feature.md" "$PLANNING_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || {
         echo "ERROR: Planning phase failed"
         write_phase_end "planning" "failed"
         exit 1
     }
-    extract_session_result "${CLAUDE_OUTPUT}.raw"
-    accumulate_tokens_from_session "$LAST_SESSION_ID"
+    extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+    accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
     write_phase_end "planning" "success"
 else
     echo "Plan file exists, skipping planning phase"
@@ -693,10 +1282,30 @@ while true; do
         TASKS_BEFORE=$(count_pending_tasks "$PLAN_FILE")
         echo "Legacy plan format — relying on source-file gate for completion."
     fi
-    export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR
-    run_claude_prompt "$PROMPTS_DIR/PROMPT.md" "$CLAUDE_CMD_IMPL" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
-    extract_session_result "${CLAUDE_OUTPUT}.raw"
-    accumulate_tokens_from_session "$LAST_SESSION_ID"
+    export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR CODING_CLI REVIEW_CLI REVIEW_MODE
+    # Continuation prompt for implementation loop iterations 2+
+    CONTINUATION_PROMPT="Continue implementing the remaining tasks in the implementation plan at $SPEC_DIR/${FEATURE}-implementation-plan.md.
+Check off completed tasks as you go. Skip any E2E testing tasks.
+Run validation (lint, typecheck, test) after completing tasks."
+    if [ $ITERATION -eq 1 ] || [ -z "$LAST_SESSION_ID" ]; then
+        echo "Mode: fresh"
+        run_claude_prompt "$PROMPTS_DIR/PROMPT.md" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+    else
+        echo "Mode: resume (session: $LAST_SESSION_ID)"
+        RESUME_EXIT=0
+        run_claude_resume "$LAST_SESSION_ID" "$CONTINUATION_PROMPT" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || RESUME_EXIT=$?
+        extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+        if [ $RESUME_EXIT -ne 0 ] || [ -z "$LAST_SESSION_ID" ]; then
+            if [ $RESUME_EXIT -ne 0 ]; then
+                echo "Resume failed (resume_exit_nonzero: exit=$RESUME_EXIT). Fallback: using fresh prompt"
+            else
+                echo "Resume failed (resume_no_session_id). Fallback: using fresh prompt"
+            fi
+            run_claude_prompt "$PROMPTS_DIR/PROMPT.md" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+        fi
+    fi
+    extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+    accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
 
     # Check if any progress was made
     TASKS_AFTER=$(count_pending_tasks "$PLAN_FILE")
@@ -752,6 +1361,20 @@ if [ "$IMPL_SUCCESS" = true ]; then
     fi
 fi
 
+# Stop early when implementation failed to avoid wasting E2E/review cycles.
+if [ "$IMPL_SUCCESS" != true ]; then
+    echo "Implementation phase failed. Skipping remaining phases."
+    write_phase_start "e2e_testing"
+    write_phase_end "e2e_testing" "skipped"
+    write_phase_start "verification"
+    write_phase_end "verification" "skipped"
+    write_phase_start "pr_review"
+    write_phase_end "pr_review" "skipped"
+    echo "$ITERATION|$MAX_ITERATIONS|$(date +%s)|failed" > "$FINAL_STATUS_FILE"
+    rm -f "$STATUS_FILE" 2>/dev/null || true
+    exit 1
+fi
+
 # Phase 5: E2E Testing
 echo "======================== E2E TESTING PHASE ========================"
 E2E_TOTAL=$({ grep "^- \[.\].*E2E:" "$PLAN_FILE" 2>/dev/null || true; } | wc -l | tr -d ' ')
@@ -763,14 +1386,32 @@ else
     write_phase_start "e2e_testing"
     E2E_SUCCESS=false
     E2E_ATTEMPT=0
+    E2E_SESSION_ID=""
+    E2E_CONTINUATION_PROMPT="Continue remaining E2E scenarios. Check the implementation plan for unchecked \`- [ ] E2E:\` entries and implement/run those tests. Run validation after completing each scenario."
     while [ $E2E_ATTEMPT -lt $MAX_E2E_ATTEMPTS ]; do
         E2E_ATTEMPT=$((E2E_ATTEMPT + 1))
+        echo "$E2E_ATTEMPT|$MAX_E2E_ATTEMPTS|$(date +%s)" > "$STATUS_FILE"
         echo "------------------------ E2E Attempt $E2E_ATTEMPT of $MAX_E2E_ATTEMPTS ------------------------"
 
-        export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR
-        run_claude_prompt "$PROMPTS_DIR/PROMPT_e2e.md" "$CLAUDE_CMD_IMPL" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
-        extract_session_result "${CLAUDE_OUTPUT}.raw"
-        accumulate_tokens_from_session "$LAST_SESSION_ID"
+        export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR CODING_CLI REVIEW_CLI REVIEW_MODE
+        if [ $E2E_ATTEMPT -eq 1 ] || [ -z "$E2E_SESSION_ID" ]; then
+            echo "E2E attempt $E2E_ATTEMPT: using full prompt"
+            run_claude_prompt "$PROMPTS_DIR/PROMPT_e2e.md" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+        else
+            echo "E2E attempt $E2E_ATTEMPT: using resume session $E2E_SESSION_ID"
+            E2E_RESUME_EXIT=0
+            run_claude_resume "$E2E_SESSION_ID" "$E2E_CONTINUATION_PROMPT" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || E2E_RESUME_EXIT=$?
+            extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+            if [ $E2E_RESUME_EXIT -ne 0 ] || [ -z "$LAST_SESSION_ID" ]; then
+                echo "E2E attempt $E2E_ATTEMPT: resume unavailable, using full prompt"
+                run_claude_prompt "$PROMPTS_DIR/PROMPT_e2e.md" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+            fi
+        fi
+        extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+        if [ -n "$LAST_SESSION_ID" ]; then
+            E2E_SESSION_ID="$LAST_SESSION_ID"
+        fi
+        accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
 
         # Check if all E2E tests passed
         E2E_FAILED=$({ grep "^- \[ \].*E2E:.*FAILED" "$PLAN_FILE" 2>/dev/null || true; } | wc -l | tr -d ' ')
@@ -784,9 +1425,24 @@ else
 
         if [ $E2E_ATTEMPT -lt $MAX_E2E_ATTEMPTS ]; then
             echo "E2E tests have failures. Running fix iteration..."
-            run_claude_prompt "$PROMPTS_DIR/PROMPT.md" "$CLAUDE_CMD_IMPL" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
-            extract_session_result "${CLAUDE_OUTPUT}.raw"
-            accumulate_tokens_from_session "$LAST_SESSION_ID"
+            if [ -n "$E2E_SESSION_ID" ]; then
+                echo "E2E fix: using resume session $E2E_SESSION_ID"
+                E2E_FIX_EXIT=0
+                run_claude_resume "$E2E_SESSION_ID" "Fix the failing E2E tests identified above. Run validation after fixing." "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || E2E_FIX_EXIT=$?
+                extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+                if [ $E2E_FIX_EXIT -ne 0 ] || [ -z "$LAST_SESSION_ID" ]; then
+                    echo "E2E fix: resume unavailable, using full prompt"
+                    run_claude_prompt "$PROMPTS_DIR/PROMPT_e2e_fix.md" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+                fi
+            else
+                echo "E2E fix: resume unavailable, using full prompt"
+                run_claude_prompt "$PROMPTS_DIR/PROMPT_e2e_fix.md" "$IMPL_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+            fi
+            extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+            if [ -n "$LAST_SESSION_ID" ]; then
+                E2E_SESSION_ID="$LAST_SESSION_ID"
+            fi
+            accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
         fi
     done
 
@@ -797,21 +1453,16 @@ else
     fi
 fi
 
-# Phase 6: Spec Verification
-echo "======================== SPEC VERIFICATION PHASE ========================"
+# Phase 6: Spec Verification (merged into review phase)
+# Verification responsibilities (spec status, acceptance criteria, README updates)
+# are now handled in Step 0 of the review prompt templates.
+# This no-op marker preserves backward compatibility for TUI phase tracking.
 write_phase_start "verification"
-export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR
-VERIFY_STATUS="success"
-if ! run_claude_prompt "$PROMPTS_DIR/PROMPT_verify.md" "$CLAUDE_CMD_OPUS" 2>&1 | tee "${CLAUDE_OUTPUT}.raw"; then
-    VERIFY_STATUS="failed"
-fi
-extract_session_result "${CLAUDE_OUTPUT}.raw"
-accumulate_tokens_from_session "$LAST_SESSION_ID"
-write_phase_end "verification" "$VERIFY_STATUS"
+write_phase_end "verification" "skipped"
 
 # Guard B: Skip PR phase if branch has no diff to default branch
 # Safety net for cases where implementation ran but produced no net diff.
-_PR_DIFF_STAT=$(git diff "$DEFAULT_BRANCH..HEAD" --stat 2>/dev/null || echo "")
+_PR_DIFF_STAT=$(git diff "${DEFAULT_BRANCH}..HEAD" --stat 2>/dev/null || echo "")
 if [ -z "$_PR_DIFF_STAT" ]; then
     echo "No diff between $BRANCH and $DEFAULT_BRANCH — skipping PR phase."
     write_phase_start "pr_review"
@@ -824,12 +1475,14 @@ if [ -z "$_PR_DIFF_STAT" ]; then
     exit 0
 fi
 
-# Phase 7: PR and Review
+# Phase 7: PR and Review (includes spec verification via Step 0 in review prompts)
 echo "======================== PR & REVIEW PHASE ========================"
 write_phase_start "pr_review"
-export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR
+export FEATURE APP_DIR SPEC_DIR PROMPTS_DIR CODING_CLI REVIEW_CLI REVIEW_MODE
 PR_STATUS="success"
 MAX_REVIEW_ATTEMPTS=3
+REVIEW_SESSION_ID=""
+REVIEW_CONTINUATION_PROMPT="The issues from the previous review have been fixed. Re-run the code review, checking only for remaining issues. Report your verdict."
 
 # Short-circuit: skip review if no files exist in diff
 _REVIEW_FILE_CHANGES=$(count_file_changes "$BASELINE_COMMIT")
@@ -860,14 +1513,7 @@ check_review_approved() {
         fi
     fi
 
-    # Secondary: check if PR was already merged (Claude may merge before verdict is captured)
-    local pr_state
-    pr_state=$(gh pr view "$BRANCH" --json state --jq '.state' 2>/dev/null || echo "")
-    if [ "$pr_state" = "MERGED" ]; then
-        return 0
-    fi
-
-    # Tertiary: check the latest PR comment for approval signal
+    # Secondary: check the latest PR comment for approval signal
     local latest_comment
     latest_comment=$(gh pr view "$BRANCH" --json comments --jq '.comments[-1].body' 2>/dev/null || echo "")
     if echo "$latest_comment" | grep -qi "VERDICT:.*APPROVED" 2>/dev/null; then
@@ -879,12 +1525,60 @@ check_review_approved() {
     return 1
 }
 
+# Wait for CI checks to finish and pass.
+# Returns 0 when checks pass (or no checks exist), 1 on failure.
+wait_for_ci_checks() {
+    local pr_ref="$1"
+    echo "Waiting for CI checks on $pr_ref..."
+
+    local checks_output=""
+    checks_output=$(gh pr checks "$pr_ref" --watch --interval 10 2>&1)
+    local checks_exit=$?
+
+    echo "$checks_output"
+
+    if [ $checks_exit -eq 0 ]; then
+        echo "CI checks passed."
+        return 0
+    fi
+
+    # Some repos have no checks configured for certain PRs.
+    if echo "$checks_output" | grep -qiE "no checks|no status checks"; then
+        echo "No CI checks found for $pr_ref. Continuing."
+        return 0
+    fi
+
+    echo "ERROR: CI checks failed or did not complete successfully." >&2
+    return 1
+}
+
+# Merge PR after all gates are green.
+# Returns 0 on success, 1 on failure.
+merge_pr_after_ci_gate() {
+    local pr_ref="$1"
+    local pr_state
+    pr_state=$(gh pr view "$pr_ref" --json state --jq '.state' 2>/dev/null || echo "")
+
+    if [ "$pr_state" = "MERGED" ]; then
+        echo "PR already merged."
+        return 0
+    fi
+
+    echo "Merging PR after CI gate..."
+    if gh pr merge "$pr_ref" --squash --delete-branch; then
+        return 0
+    fi
+
+    echo "ERROR: Failed to merge PR $pr_ref after CI gate." >&2
+    return 1
+}
+
 if [ "$REVIEW_MODE" = "manual" ]; then
-    if ! run_claude_prompt "$PROMPTS_DIR/PROMPT_review_manual.md" "$CLAUDE_CMD_OPUS" 2>&1 | tee "${CLAUDE_OUTPUT}.raw"; then
+    if ! run_claude_prompt "$PROMPTS_DIR/PROMPT_review_manual.md" "$REVIEW_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw"; then
         PR_STATUS="failed"
     fi
-    extract_session_result "${CLAUDE_OUTPUT}.raw"
-    accumulate_tokens_from_session "$LAST_SESSION_ID"
+    extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+    accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
 
 elif [ "$REVIEW_MODE" = "merge" ]; then
     # Merge mode: create PR, iterate review+fixes until approved, then merge
@@ -893,22 +1587,53 @@ elif [ "$REVIEW_MODE" = "merge" ]; then
     while [ $REVIEW_ATTEMPT -lt $MAX_REVIEW_ATTEMPTS ]; do
         REVIEW_ATTEMPT=$((REVIEW_ATTEMPT + 1))
         echo "--- Review attempt $REVIEW_ATTEMPT of $MAX_REVIEW_ATTEMPTS ---"
-        run_claude_prompt "$PROMPTS_DIR/PROMPT_review_merge.md" "$CLAUDE_CMD_OPUS" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
-        extract_session_result "${CLAUDE_OUTPUT}.raw"
-        accumulate_tokens_from_session "$LAST_SESSION_ID"
+        if [ $REVIEW_ATTEMPT -eq 1 ] || [ -z "$REVIEW_SESSION_ID" ]; then
+            echo "Review attempt $REVIEW_ATTEMPT: using full prompt"
+            run_claude_prompt "$PROMPTS_DIR/PROMPT_review_merge.md" "$REVIEW_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+        else
+            echo "Review attempt $REVIEW_ATTEMPT: using resume session $REVIEW_SESSION_ID"
+            REVIEW_RESUME_EXIT=0
+            run_claude_resume "$REVIEW_SESSION_ID" "$REVIEW_CONTINUATION_PROMPT" "$REVIEW_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || REVIEW_RESUME_EXIT=$?
+            extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+            if [ $REVIEW_RESUME_EXIT -ne 0 ] || [ -z "$LAST_SESSION_ID" ]; then
+                echo "Review attempt $REVIEW_ATTEMPT: resume unavailable, using full prompt"
+                run_claude_prompt "$PROMPTS_DIR/PROMPT_review_merge.md" "$REVIEW_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+            fi
+        fi
+        extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+        if [ $REVIEW_ATTEMPT -eq 1 ] && [ -n "$LAST_SESSION_ID" ]; then
+            REVIEW_SESSION_ID="$LAST_SESSION_ID"
+        fi
+        accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
 
         # Check stdout and PR comment for approval
         if check_review_approved "${CLAUDE_OUTPUT}.raw"; then
             echo "Review approved! Running post-approval test gate..."
             if check_tests_pass_or_baseline; then
                 echo "Post-approval test gate passed."
+                if ! wait_for_ci_checks "$BRANCH"; then
+                    PR_STATUS="failed"
+                    break
+                fi
+                if ! merge_pr_after_ci_gate "$BRANCH"; then
+                    PR_STATUS="failed"
+                    break
+                fi
                 REVIEW_APPROVED=true
                 break
             else
                 echo "WARNING: Tests failing after review approval. Running fix iteration..."
                 run_review_fix
                 if check_tests_pass_or_baseline; then
-                    echo "Tests pass after fix. Proceeding with merge."
+                    echo "Tests pass after fix. Running CI and merge gates."
+                    if ! wait_for_ci_checks "$BRANCH"; then
+                        PR_STATUS="failed"
+                        break
+                    fi
+                    if ! merge_pr_after_ci_gate "$BRANCH"; then
+                        PR_STATUS="failed"
+                        break
+                    fi
                     REVIEW_APPROVED=true
                     break
                 else
@@ -936,9 +1661,24 @@ else
     while [ $REVIEW_ATTEMPT -lt $MAX_REVIEW_ATTEMPTS ]; do
         REVIEW_ATTEMPT=$((REVIEW_ATTEMPT + 1))
         echo "--- Review attempt $REVIEW_ATTEMPT of $MAX_REVIEW_ATTEMPTS ---"
-        run_claude_prompt "$PROMPTS_DIR/PROMPT_review_auto.md" "$CLAUDE_CMD_OPUS" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
-        extract_session_result "${CLAUDE_OUTPUT}.raw"
-        accumulate_tokens_from_session "$LAST_SESSION_ID"
+        if [ $REVIEW_ATTEMPT -eq 1 ] || [ -z "$REVIEW_SESSION_ID" ]; then
+            echo "Review attempt $REVIEW_ATTEMPT: using full prompt"
+            run_claude_prompt "$PROMPTS_DIR/PROMPT_review_auto.md" "$REVIEW_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+        else
+            echo "Review attempt $REVIEW_ATTEMPT: using resume session $REVIEW_SESSION_ID"
+            REVIEW_RESUME_EXIT=0
+            run_claude_resume "$REVIEW_SESSION_ID" "$REVIEW_CONTINUATION_PROMPT" "$REVIEW_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || REVIEW_RESUME_EXIT=$?
+            extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+            if [ $REVIEW_RESUME_EXIT -ne 0 ] || [ -z "$LAST_SESSION_ID" ]; then
+                echo "Review attempt $REVIEW_ATTEMPT: resume unavailable, using full prompt"
+                run_claude_prompt "$PROMPTS_DIR/PROMPT_review_auto.md" "$REVIEW_CMD" 2>&1 | tee "${CLAUDE_OUTPUT}.raw" || true
+            fi
+        fi
+        extract_session_result "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
+        if [ $REVIEW_ATTEMPT -eq 1 ] && [ -n "$LAST_SESSION_ID" ]; then
+            REVIEW_SESSION_ID="$LAST_SESSION_ID"
+        fi
+        accumulate_tokens_from_session "$LAST_SESSION_ID" "${CLAUDE_OUTPUT}.raw" "$LAST_RUN_CLI"
 
         # Check stdout and PR comment for approval
         if check_review_approved "${CLAUDE_OUTPUT}.raw"; then
@@ -1016,6 +1756,8 @@ fi
 rm -f "$STATUS_FILE" 2>/dev/null || true
 rm -f "/tmp/ralph-loop-${FEATURE}.output" 2>/dev/null || true
 rm -f "/tmp/ralph-loop-${FEATURE}.output.raw" 2>/dev/null || true
+rm -f "/tmp/ralph-loop-${FEATURE}.last-message" 2>/dev/null || true
+rm -f "$PRE_RUN_DIRTY_FILE" 2>/dev/null || true
 
 # Print final token usage
 if [ -f "/tmp/ralph-loop-${FEATURE}.tokens" ]; then