Second Opinion Command

Get comprehensive multi-model AI feedback on your code, design decisions, or bug analysis.

🚀 Uses command-line approach to bypass 25K token limits and send full PR context!

Usage

# Get second opinion on current PR branch
/secondo

# Specify feedback type
/secondo design
/secondo code-review
/secondo bugs

# With specific question
/secondo "Should I use Redis or in-memory caching for rate limiting?"

How It Works

This command uses a direct approach with auth-cli.mjs for secure token management:

1. Authentication (Auto-Refresh):

   • Uses ~/.claude/scripts/auth-cli.mjs token for secure token retrieval
   • Automatically refreshes expired ID tokens using refresh token (silent, no browser popup)
   • ID tokens expire after 1 hour, refresh tokens enable 30+ day sessions
   • Only opens browser for initial login or if refresh token expires

2. Gather Full PR Context (automated via build_second_opinion_request.py):

   • Resolve branch + base reference (prefers SECOND_OPINION_BASE_REF, falls back to origin/main → main → master)
   • Capture git status --short, git diff --stat, and recent commits for orientation
   • Generate a full diff versus the base ref (truncated to stay under token budgets with explicit notices)
   • Attach per-file patches (up to configurable limit) so the MCP tool sees real code snippets instead of summaries

3. Build Comprehensive Request:

   • Create detailed analysis request (optimized to stay under 25K tokens)
   • Include all relevant code context
   • Add production context and testing requirements

4. Direct MCP Server Call:

   • Uses HTTPie with auto-refreshed Bearer token
   • Sends to: https://ai-universe-backend-dev-114133832173.us-central1.run.app/mcp
   • Handles streaming responses properly
   • Saves results locally

5. Multi-Model Analysis:

   • Gemini (Primary model)
   • Perplexity (Secondary)
   • OpenAI (Secondary)
   • Synthesis with 50+ authoritative sources

6. Results Display:

   • Save comprehensive report to tmp/secondo_analysis_[timestamp].md
   • Display verdict and key findings
   • Show token usage and cost breakdown

Implementation Protocol

When executing /second_opinion or /secondo:

Step 0: Authentication Setup (Auto-Refresh)

# Verify auth-cli.mjs is installed
if [ ! -f "$HOME/.claude/scripts/auth-cli.mjs" ]; then
  echo "❌ auth-cli.mjs not found. Run /localexportcommands to install"
  exit 1
fi

# Get token (auto-refreshes if expired using refresh token)
# This is silent - only prompts for login if refresh token is invalid/missing
TOKEN=$(node ~/.claude/scripts/auth-cli.mjs token)

# If this fails, user needs to authenticate
if [ $? -ne 0 ]; then
  echo "❌ Authentication failed. Please run:"
  echo "   node ~/.claude/scripts/auth-cli.mjs login"
  exit 1
fi

Key Behavior:

• Seamless Auto-Refresh: Automatically renews ID tokens using refresh token (no browser popup)
• 30+ Day Sessions: Refresh tokens enable long-lived sessions without re-authentication
• Browser Only When Needed: Only opens browser for initial login or if refresh token expires
• Same Token File: Uses ~/.ai-universe/auth-token.json (exact same as AI Universe repo)

Step 1: Gather PR Context (now automated)

skills/second_opinion_workflow/scripts/request_second_opinion.sh calls build_second_opinion_request.py to harvest the full PR delta. The helper:

• Resolves the comparison base (env override SECOND_OPINION_BASE_REF, then origin/main → main → master, finally HEAD^).
• Captures branch name, repo root, remote URL, git status --short, git diff --stat, recent commits, and the full diff.
• Extracts per-file patches for up to SECOND_OPINION_MAX_FILES (default 20) with truncation markers when limits are hit.
• Annotates the payload with gitContextNotices so the MCP tool sees exactly what was trimmed.

Manual usage if you want to inspect the payload directly:

python3 skills/second_opinion_workflow/scripts/build_second_opinion_request.py \
  /tmp/secondo_request.json \
  "What should I double-check before merging?" \
  3 \
  origin/main

Tune the capture limits with environment variables:

export SECOND_OPINION_BASE_REF=origin/main   # override comparison base
export SECOND_OPINION_MAX_FILES=25           # number of per-file patches to attach
export SECOND_OPINION_MAX_DIFF_CHARS=32000   # full diff char budget
export SECOND_OPINION_MAX_PATCH_CHARS=8000   # per-file diff char budget

Step 2: Build Analysis Request

The generated payload now includes the git context automatically:

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "agent.second_opinion",
    "arguments": {
      "question": "Should I land this patch set as-is?",
      "maxOpinions": 3,
      "gitContext": {
        "branch": "feature/refactor",
        "base": "origin/main",
        "diffstat": "…",
        "recentCommits": "…",
        "changedFiles": [
          {"status": "M", "path": "$PROJECT_ROOT/api/routes.py"},
          {"status": "A", "path": "$PROJECT_ROOT/services/cache.py"}
        ],
        "patches": {
          "$PROJECT_ROOT/api/routes.py": "@@ -42,6 +42,15 @@ …",
          "$PROJECT_ROOT/services/cache.py": "@@ -0,0 +1,200 @@ …"
        },
        "limits": {
          "maxFiles": 20,
          "diffCharLimit": 24000,
          "patchCharLimit": 6000
        }
      },
      "gitContextNotices": [
        "git diff origin/main...HEAD truncated by 1520 characters (limit 24000)."
      ]
    }
  },
  "id": 1
}

💡 You can still tailor the natural-language question, but you no longer need to paste diff snippets manually—the helper attaches them for you.

Step 3: Execute Request

# Call MCP server with HTTPie (matches request_second_opinion.sh)
http POST "https://ai-universe-backend-dev-114133832173.us-central1.run.app/mcp" \
  "Accept:application/json, text/event-stream" \
  "Authorization:Bearer $TOKEN" \
  < /tmp/secondo_request.json \
  --timeout=180 \
  --print=b > /tmp/secondo_response.json

# Check if successful
if [ $? -eq 0 ] && [ -s /tmp/secondo_response.json ]; then
  echo "✅ Analysis complete"
else
  echo "❌ Request failed"
  exit 1
fi

Step 4: Parse and Display Results

Extract from /tmp/secondo_response.json:

1. Parse the JSON response (look for result.content[0].text)
2. Extract verdict, token counts, costs, sources
3. Save formatted report to tmp/secondo_analysis_[timestamp].md
4. Display key findings to user

Example parsing:

# Extract the main response text
jq -r '.result.content[0].text' /tmp/secondo_response.json > /tmp/secondo_parsed.txt

# Extract metrics if available
TOKENS=$(jq -r '.result.content[0].text' /tmp/secondo_response.json | grep -o 'Total Tokens: [0-9,]*' | head -1)
COST=$(jq -r '.result.content[0].text' /tmp/secondo_response.json | grep -o 'Total Cost: \$[0-9.]*' | head -1)

Token Optimization

Maximum token budget: 24,900 tokens (stay under 25K limit)

Allocation strategy:

• Request overhead: ~100 tokens (JSON structure)
• Question/context: ~500-1000 words (optimize for clarity)
• Git diff stats: Full output (~100-500 tokens)
• Critical code sections: ~15K tokens (selective, not full files)
• Commit messages: ~200 tokens
• Metadata: ~100 tokens

Tips to maximize context:

1. Include git diff --stat (compact, informative)
2. Select critical changed sections (not full files)
3. Prioritize files with complex logic changes
4. Keep question focused (~500 words)
5. Skip unchanged context

Authentication & Rate Limits

Authentication: Required via Firebase OAuth (exact same as AI Universe repo)

• Initial Login: node ~/.claude/scripts/auth-cli.mjs login (browser-based OAuth, run outside Claude Code)
• Check Status: node ~/.claude/scripts/auth-cli.mjs status (view current auth status)
• Token Location: ~/.ai-universe/auth-token.json (ID token + refresh token)
• Session Duration: 30+ days (refresh tokens auto-renew ID tokens silently)
• Auto-Refresh: /secondo automatically refreshes expired ID tokens (no browser popup needed)

Token Lifecycle:

• ID Token: Expires after 1 hour (Firebase security policy)
• Refresh Token: Enables automatic ID token renewal for 30+ days
• Auto-Renewal: Silent, seamless - only opens browser if refresh token expires

When You'll Need to Re-authenticate:

• Initial setup (no token file exists)
• After 30+ days (when refresh token expires)
• If token file is corrupted or manually deleted

Rate Limits: Applied per authenticated user based on Firebase account

Practical limits:

• Server timeout: 180 seconds max
• Token limit: 25,000 tokens per response
• Cost: ~$0.10 per full PR analysis (3 models)

Output Format

Display results in markdown with:

• 📊 Summary: Models used, tokens, cost, sources
• 🎯 Verdict: Unanimous consensus or majority opinion
• ⚠️ Critical Issues: Security, correctness, production safety
• 💡 Model Perspectives: Individual model insights
• ✅ Validation: What was confirmed as correct
• 🔗 Sources: Authoritative references (50+)

Save to: tmp/secondo_analysis_[timestamp].md

Success Criteria

✅ Request completes successfully (curl exit code 0) ✅ Response file is non-empty ✅ Response contains valid JSON with result.content ✅ Analysis report saved to tmp directory ✅ User sees verdict and key findings

Error Handling

Common issues:

1. Token limit exceeded: Reduce code context, keep question focused
2. Timeout: Complex analysis may take 60-120 seconds, use --max-time 180
3. Empty response: Check curl exit code, verify network connectivity
4. JSON parse error: Response may be streaming format, extract text content first

Example Execution Flow

User: /secondo

1. Gather PR context:
   ✓ Branch: fix_mcp
   ✓ Changed files: 18 files (+2448/-799)
   ✓ Git diff saved to /tmp/secondo_diff_full.txt
   ✓ Commits: 5 commits analyzed

2. Build analysis request:
   ✓ Question: 487 words
   ✓ Code context: 14,250 tokens
   ✓ Total estimated: 23,890 tokens (95.6% of limit)

3. Execute request:
   ✓ curl -X POST [MCP endpoint]
   ✓ Response: 73KB received in 47 seconds
   ✓ Status: HTTP 200 OK

4. Parse results:
   ✓ Models: 3 (Gemini, Perplexity, OpenAI)
   ✓ Total tokens: 24,964
   ✓ Total cost: $0.10193
   ✓ Sources: 52 authoritative references

5. Display verdict:
   🎯 UNANIMOUS VERDICT: CORRECT (with caveats)

   ✅ Fix is safe for production
   ⚠️ Array initialization discipline required
   📊 Report saved: tmp/secondo_analysis_20251031_1847.md