gemini-peer-review

# Gemini Peer Review

Peer validation system using Google Gemini CLI. Validates Claude's designs and code reviews through structured discussion before presenting to user.

**Core principle:** Two AI perspectives catch more issues than one. When they disagree, structured discussion resolves most issues. External research (Perplexity if available, otherwise WebSearch) arbitrates persistent disagreements.

## CRITICAL: Gemini CLI Differences from Codex

| Aspect | Codex CLI | Gemini CLI |
|--------|-----------|------------|
| JSON output | `--json` | `-o json` |
| Model selection | `--model X` | `-m X` |
| Review command | `codex review --base X` | **NONE** - use prompts with git diff |
| Session resume | `resume $SESSION_ID` | **UNRELIABLE** - re-inject context instead |
| Tool prevention | Not needed | **MANDATORY** prompt suffix |
| Sandbox mode | Not needed | `-s` flag required |

### MANDATORY: Tool Prevention

**EVERY Gemini prompt MUST end with:**
```
CRITICAL: Do not use any tools. Output text only.
```

Without this, Gemini may attempt to use tools and produce unexpected results.

### Session Continuity Strategy

Gemini's `--resume latest` is index-based (not ID-based) and unreliable. For multi-round discussions:
- **DO NOT** use `--resume`
- **DO** re-inject the full context from previous rounds into each new prompt

### Robust JSON Parsing (IMPORTANT)

Gemini's `-o json` output is sometimes wrapped in Markdown code blocks. **Always parse output** to extract clean JSON:

```bash
# Function to extract JSON from potentially wrapped output
parse_gemini_output() {
  local input="$1"
  # Strip markdown code fences if present
  echo "$input" | sed 's/^```json//; s/^```//; s/```$//' | \
    # Remove CLI status messages
    grep -v '^Loaded cached' | \
    # Extract response field, or return raw if not JSON
    jq -r '.response // .' 2>/dev/null || echo "$input"
}

# Usage:
RAW_OUTPUT=$(gemini -s -m gemini-3-pro-preview -o json "$(cat "$PROMPT_FILE")" 2>&1)
RESPONSE=