---
name: crossfire
description: Get multi-model perspectives (Codex + Gemini) on any work product. Use at any development stage — planning, design, implementation, testing — to catch blind spots the primary agent might miss.
allowed-tools: AskUserQuestion, Bash, Glob, Grep, Read
model: sonnet
---

## Context

- Arguments: `${ARGS}` (optional: artifact type, file path, or review focus question)
- Main repo root: !`sh -c 'COMMON=$(git rev-parse --path-format=absolute --git-common-dir 2>/dev/null) && dirname "$COMMON" || echo "NOT_IN_GIT_REPO"'`
- Current branch: !`git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "NO_BRANCH"`
- Uncommitted changes: !`git status --porcelain 2>/dev/null | wc -l | xargs`
- PLAN files: !`sh -c 'COMMON=$(git rev-parse --path-format=absolute --git-common-dir 2>/dev/null); if [ -z "$COMMON" ]; then exit 0; fi; PROJECT_ROOT=$(dirname "$COMMON"); cd "$PROJECT_ROOT" && for f in PLAN__*.md; do [ -f "$f" ] && echo "$f"; done' 2>/dev/null | head -5`

# Run Crossfire Review

Detect what to review from `${ARGS}` using the Artifact Detection rules below, then proceed immediately to Orchestration.

## Artifact Detection

Parse `${ARGS}` to determine what to review. Check in this order:

### 1. Explicit Artifact Type

If args contain a recognized keyword, use that:

| Keyword | Artifact | How to gather |
|---------|----------|---------------|
| `plan` | Most recent PLAN file | Read the newest `PLAN__*.md` file from the main repo root |
| `diff` | Current code changes | `git diff` (unstaged), then `git diff --cached` (staged), then `git diff origin/$(git symbolic-ref --short HEAD)..HEAD` (unpushed) — use first non-empty |
| `tests` | Test files | Identify recently modified test files via `git diff --name-only` filtered to test patterns |

### 2. Explicit File Path

If args contain a file path (starts with `/`, `./`, or `~`), read that file as the artifact.

### 3. Review Focus Question

If args are a quoted string or natural language question (e.g., `"is this architecture sound?"`), treat it as the review focus. Gather the most relevant artifact from context:
- If there are uncommitted changes → use the diff
- If PLAN files exist → use the most recent one
- Otherwise, ask the user what to review

### 4. No Args

If no args provided:
- If uncommitted changes exist → review the diff
- If PLAN files exist → review the most recent one
- Otherwise, ask the user: "What would you like me to get a crossfire review on?"

---

Once you have the artifact and optional review focus, proceed immediately to **Orchestration** without waiting for further user input.

## Orchestration

### Step 1: Check CLI Availability

```bash
CODEX_AVAILABLE=$(command -v codex >/dev/null 2>&1 && echo "yes" || echo "no")
GEMINI_AVAILABLE=$(command -v gemini >/dev/null 2>&1 && [ -f ~/.env/gemini_cli.key ] && echo "yes" || echo "no")
```

If neither CLI is available, inform the user and stop:

```
## Crossfire Review

Crossfire unavailable: neither `codex` nor `gemini` CLI found.
```

If only one is available, proceed with that single CLI — its findings become single-model observations by default.

### Step 2: Build the Review Prompt

Construct a prompt for the external CLI agents. The prompt must include these sections in order:

1. **Preamble:**
```
You are reviewing the following artifact. Analyze it critically from your perspective as an independent reviewer. The primary AI agent has already produced this work — your job is to catch what it missed.

Before evaluating the artifact, explore the repository to understand the surrounding context:
1. Read the full content of any files that were changed (not just the diff lines)
2. Identify and read files the changes depend on (imports, parent classes, shared types, interfaces)
3. Check the project structure to understand where the changes fit architecturally

Review the artifact with this full context in mind. Do not assume that code not shown in the artifact doesn't exist — verify by reading the actual files.
```

2. **Review focus** (if one was determined from artifact detection):
```
REVIEW FOCUS: <the focus question>
```

3. **Review instructions:**
```
For each concern you identify, categorize it as:
- CRITICAL: Fundamental flaw, security risk, data loss potential, or incorrect approach
- IMPORTANT: Significant gap, missing consideration, or maintainability concern
- MINOR: Nice-to-have improvement, style issue, or alternative worth considering

Structure your response as:

## Concerns

### [CRITICAL/IMPORTANT/MINOR]: <title>
<explanation of the concern and why it matters>

### [CRITICAL/IMPORTANT/MINOR]: <title>
<explanation>

## What's Done Well
<acknowledge strengths — what should NOT be changed>

## Alternative Approaches
<if you would have taken a fundamentally different approach, describe it briefly>
```

4. **Artifact** appended at the end:
```
--- ARTIFACT ---
<artifact content>
```

### Step 3: Write Prompt and Launch CLIs

Write the full prompt to a temp file. Do NOT pass it as a command-line argument or via stdin — large artifacts exceed the OS `ARG_MAX` limit, and stdin piping causes Codex to echo the full prompt to stderr. Instead, write to a file, close stdin with `< /dev/null` (prevents non-TTY hang), and instruct the CLI to `cat` the file.

```bash
PROMPT_DIR=$(mktemp -d /tmp/crossfire-prompt-XXXXXX)
PROMPT_FILE="$PROMPT_DIR/prompt.txt"
# Write the constructed prompt from Step 2 to $PROMPT_FILE

CODEX_OUT=$(mktemp)
CODEX_ERR=$(mktemp)
GEMINI_OUT=$(mktemp)
REPO_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
# Gemini's ImportProcessor resolves @-refs relative to cwd — ENOENT errors from subdirs
cd "$REPO_ROOT"

# Clean up all temp files when shell exits (normal or interrupt)
trap 'rm -rf "$PROMPT_DIR" "$CODEX_OUT" "$CODEX_ERR" "$GEMINI_OUT"' EXIT INT TERM

# Codex (background) — only if available
if [ "$CODEX_AVAILABLE" = "yes" ]; then
  timeout 600 codex exec -C "$REPO_ROOT" \
    --dangerously-bypass-approvals-and-sandbox \
    "Run cat \"$PROMPT_FILE\" and follow the instructions in the output." \
    < /dev/null \
    > "$CODEX_OUT" 2>"$CODEX_ERR" &
  CODEX_PID=$!
fi

# Gemini (background) — only if available
if [ "$GEMINI_AVAILABLE" = "yes" ]; then
  GEMINI_API_KEY=$(cat ~/.env/gemini_cli.key) timeout 600 gemini --yolo \
    --include-directories "$PROMPT_DIR" \
    -p "Read the file at $PROMPT_FILE and follow the review instructions inside it." \
    > "$GEMINI_OUT" 2>&1 &
  GEMINI_PID=$!
fi

# Wait for both
[ -n "$CODEX_PID" ] && { wait $CODEX_PID; CODEX_EXIT=$?; }
[ -n "$GEMINI_PID" ] && { wait $GEMINI_PID; GEMINI_EXIT=$?; }

# Output results inline (parsed in Step 4 — no temp files need to persist)
echo "===CODEX_EXIT=${CODEX_EXIT:--1}==="
echo "===CODEX_OUTPUT_START==="
cat "$CODEX_OUT" 2>/dev/null || true
echo ""
echo "===CODEX_OUTPUT_END==="
echo "===CODEX_STDERR_START==="
cat "$CODEX_ERR" 2>/dev/null || true
echo ""
echo "===CODEX_STDERR_END==="
echo "===GEMINI_EXIT=${GEMINI_EXIT:--1}==="
echo "===GEMINI_OUTPUT_START==="
cat "$GEMINI_OUT" 2>/dev/null || true
echo ""
echo "===GEMINI_OUTPUT_END==="
```

### Step 4: Parse and Validate Outputs

Parse the delimited output from Step 3. Extract each model's output from the `===..._START===` / `===..._END===` markers, and read the exit codes from the `===..._EXIT=N===` lines.

Validate each model's output:
- Empty output (0 lines between markers) → mark as "Unavailable: no output produced"
- Non-empty output (>10 lines) with non-zero exit code → use the output and prepend a warning: "(Warning: CLI exited with code N — output may be partial. Stderr: [content from CODEX_STDERR section if applicable])"
- Non-zero exit code AND output is ≤10 lines → mark as "Unavailable: CLI exited with error (exit code N)"
- If Codex stderr contains only shell-init warnings (nvm, rvm, etc.) and stdout has substantive content, treat as a successful run with a warning note

Temp files are cleaned up automatically by the EXIT trap in Step 3. No manual cleanup is needed.

### Step 5: Synthesize Consensus

Apply these rules to produce the final report:

**Cross-model consensus:**
- Concern flagged by both Codex and Gemini = **agreed** (high confidence)
- Concern flagged by only one model = **potential blind spot** (note which model flagged it)

**Severity resolution:**
- If models disagree on severity for the same concern, use the highest severity (CRITICAL > IMPORTANT > MINOR)

## Output Format

Present your findings in this structure:

```
## Crossfire Review

### Codex (GPT)
[Paste Codex structured output, or "Unavailable: [reason]"]

### Gemini
[Paste Gemini structured output, or "Unavailable: [reason]"]

### Cross-Model Consensus

**Agreed concerns (2+ models flagged):**
- [CRITICAL/IMPORTANT/MINOR] [Concern]: flagged by [which models]

**Single-model concerns (potential blind spots):**
- [CRITICAL/IMPORTANT/MINOR] [Concern]: only [model] flagged this

### Strengths Confirmed
[Items that multiple models agreed are done well]

### Summary
[2-3 sentence synthesis: overall confidence level, key action items, and whether the artifact needs revision]
```

## Key Requirements

- Launch CLIs in parallel, not sequentially
- Do NOT modify the artifact — only review it
- Do NOT skip concerns because they seem minor — surface everything and let severity classification do the prioritization
- Temp files are cleaned up automatically via the EXIT trap — do not add manual cleanup
- If only one CLI is available, still run the review with that single model

## Examples

- `/crossfire` — auto-detects artifact (diff or plan), runs orchestration
- `/crossfire plan` — reviews the most recent PLAN file
- `/crossfire diff` — reviews current code changes
- `/crossfire "is this the right approach for caching?"` — reviews recent context with that focus question
- `/crossfire src/auth/middleware.py` — reviews a specific file