ABOUTME: Three independent second opinions from isolated Docker containers (Claude + Gemini + DeepSeek)

ABOUTME: Same prompt/context to all, zero config contamination, four-way synthesis

Second Opinion (Isolated)

All reviewers run in Docker containers with NO access to your ~/.claude/ config, memories, rules, or settings. This ensures genuinely independent opinions. The three reviewers span three labs (Anthropic, Google, DeepSeek), so agreement across them is a strong signal rather than shared-model bias.

Prerequisites

See references/setup.md.

Execution Flow

Step 1: Understand the problem

Before calling reviewers, clearly articulate:

1. The problem - what is being analyzed or decided
2. Your current analysis - what you think so far
3. Relevant context - code snippets, error messages, architecture constraints
4. Specific question - what you want a second opinion on

Step 2: Gather context

Collect all relevant material:

• Read the files involved (full content, not summaries)
• Include error messages/stack traces if debugging
• Include project constraints (from CLAUDE.md, architecture docs)
• Include what has been tried and why it didn't work (if applicable)

Step 3: Build the prompt

Load the prompt template:

PROMPT_TEMPLATE=$(cat ~/.claude/skills/second-opinion/prompts/default.md)

Compose a FULL_PROMPT combining the template with a ## Problem Context section containing:

• Problem description
• Your current analysis/hypothesis
• All gathered context (code, errors, docs)
• The specific question for the reviewer

The FULL_PROMPT must be identical for all three reviewers. Write it to a temp file:

PROMPT_FILE=$(mktemp)
cat > "$PROMPT_FILE" <<'PROMPT_EOF'
<the composed prompt here>
PROMPT_EOF

Step 4: Call all reviewers in parallel

Launch all containers simultaneously using parallel Bash tool calls in a single message.

Set an explicit Bash-tool timeout per call. These docker run commands are inline, not the .sh scripts, so the only wall-clock bound is the Bash-tool timeout. The default (120s) will kill deepseek-reasoner mid-reasoning. Pass timeout: 600000 (600s) for the DeepSeek call and at least timeout: 300000 for Claude and Gemini.

Credentials never touch the host filesystem: Claude auth is mounted directly from the Docker volume, Gemini and DeepSeek API keys are read in-memory from files.

Call 1 - Isolated Claude:

docker run --rm \
  -v claude-reviewer-auth:/home/node/.claude:ro \
  -v <PROJECT_ROOT>:/workspace:ro \
  claude-reviewer:latest \
  --print \
  --model fable \
  "$(cat <PROMPT_FILE>)"

Call 2 - Isolated Gemini:

docker run --rm \
  -e GEMINI_API_KEY="$(cat ~/.config/gemini-api-key)" \
  -v <PROJECT_ROOT>:/workspace:ro \
  gemini-reviewer:latest \
  -p "$(cat <PROMPT_FILE>)" \
  -m gemini-3.6-flash \
  --sandbox false \
  2>&1 | grep -v "^\[WARN\] Skipping unreadable" | grep -v "^Warning: Could not read"

Call 3 - Isolated DeepSeek (pi):

docker run --rm \
  -e DEEPSEEK_API_KEY="$(cat ~/.config/deepseek-api-key)" \
  -v <PROJECT_ROOT>:/workspace:ro \
  deepseek-reviewer:latest \
  --provider deepseek \
  --model deepseek-reasoner \
  -p \
  -t read \
  --no-session \
  "$(cat <PROMPT_FILE>)"

Cleanup. Set the trap right after creating the temp file in Step 3, so it is removed even if a call times out before this point is reached:

trap 'rm -f "$PROMPT_FILE"' EXIT

Degrade gracefully on reviewer failure. This is a procedure, not a suggestion. A reviewer can fail independently: expired Claude OAuth in the volume (401), a missing/invalid API key, a rate limit, or a timeout.

1. Classify each reviewer as OK or FAILED before synthesizing. A reviewer is FAILED if its Bash call exited non-zero OR its output matches 401, timed out, API Error, No API key, or is empty. Do NOT treat a 401 body or an error trace as if it were an opinion.
2. Never abort if at least one reviewer is OK. Synthesize from the survivors. A two-of-three (or one-of-three) synthesis is useful; a silent drop is not.
3. Account for every reviewer in the output via the status line in Step 5.
4. If Claude returns 401, surface the re-login command from Troubleshooting.

Step 5: Synthesize

Required first line: reviewer status. Before the table, state which reviewers responded, so a missing one can never be glossed over:

Reviewer status: Claude ✅ | Gemini ✅ | DeepSeek ❌ (401, expired OAuth; see Troubleshooting)

Then present a four-way analysis (drop the column of any FAILED reviewer):

Final recommendation: Your updated position, incorporating all independent opinions. Explain what changed (or didn't) and why. Flag any point where the independent reviewers agree against your original analysis (the more of the three that agree, the stronger the signal to reconsider).

When to Use

• Stuck on a debugging problem
• Weighing architectural trade-offs
• Unsure about root cause analysis
• Want to validate an approach before implementing
• Complex decisions with multiple valid paths

Auto-trigger (no user prompt needed)

Call this skill on your own initiative when any of these holds:

• Complexity = complex (orchestrator step 0b): validate the approach before planning
• Debugging with 2+ failed root-cause attempts: get a fresh perspective before trying again
• 🔴 decisions (rewrites, core logic, security): challenge the assumptions before deciding
• Review loop stuck (score < 80 after 2 fix rounds): break out of the cycle

When NOT to Use

• Code review before commit (use gemini-review)
• 🟢 tasks, simple or moderate complexity, docs and config changes
• Simple, well-understood problems
• When you just need to read more code/docs first

Troubleshooting

See references/setup.md.