second-opinion

Cross-provider "phone a friend" with smart routing.

When to use

Trigger when the user says:

• "get a second opinion on …"
• "ask {grok|gemini|codex|chatgpt} what they think" (for a Claude opinion, spawn a subagent from the lead instead — see Note on Claude below)
• "fan out to everyone"
• "deep research on …"
• "what would another model say"
• The user's task fits one of the routing rules below and you'd benefit from another perspective.

Don't use it for: trivial lookups, tasks where the user wants you to answer, or any task with sensitive personal data the user hasn't authorized to leave the local session.

Routing rubric

The router script (scripts/classify.sh) picks a provider by keywords. Override with --provider when the user names one explicitly.

Task shape	Provider	Why
Frontend / UI / design critique	gemini → auto-falls-through to OpenRouter (google/gemini-2.5-pro) on error/missing key	Strong visual + UX reasoning
Humor / romance / vibe-check / casual	grok	Looser, more conversational
Tool-calling code / agent-orchestration design	codex (or spawn subagent for Claude-specific)	Codex is strong at structured engineering reasoning; for a true Claude opinion, spawn a subagent — see Note on Claude
Engineering design review / architecture	codex	OpenAI's o-series via codex CLI (rides on ChatGPT sub)
Refusal fallback / sensitive prompts that hit ethics gates	openrouter	Routes to a low-refusal model (default mistralai/mistral-large-2411); override via OPENROUTER_MODEL
Deep research / multi-source synthesis	fan-out + lead-session synthesis	Lead Claude session synthesizes the raw fan-out output
(default, no signal)	codex	Safest general-purpose, subscription-routed

How to invoke

# Auto-routed
~/.claude/skills/second-opinion/scripts/dispatch.sh "should I use Postgres or DuckDB for this"

# Explicit provider
~/.claude/skills/second-opinion/scripts/dispatch.sh --provider grok "tell me a joke about kubernetes"

# Deep research (provider's research mode if available)
~/.claude/skills/second-opinion/scripts/dispatch.sh --deep-research "what's the latest on GLP-1 + vestibular migraine"

# Fan out to all configured providers + synthesize
~/.claude/skills/second-opinion/scripts/dispatch.sh --fanout "is this migration plan safe?"

# With attached input
~/.claude/skills/second-opinion/scripts/dispatch.sh --input-file /tmp/diff.txt --mode critique "review"

The /second-opinion slash command is a thin wrapper over dispatch.sh.

Auth

Required env vars / auth (only the providers you actually use):

• Codex — no env var. Authenticates via ChatGPT subscription (Mac/Plus/Pro) through codex CLI. Install: npm install -g @openai/codex. CLI subprocess rides on the sub, no API billing.
• Claude — no adapter. The lead session IS Claude. To get a Claude second opinion, spawn a subagent (see Note on Claude below).
• XAI_API_KEY — Grok (API; xAI hasn't shipped a sub-routed CLI as of May 2026)
• GEMINI_API_KEY — Gemini (API today; the gemini CLI is installed but auth not yet sub-routed — see Note on Gemini below)
• OPENROUTER_API_KEY — OpenRouter (refusal fallback / model marketplace; separately billed at OpenRouter rates)
• OPENROUTER_MODEL — optional model override for OpenRouter (default mistralai/mistral-large-2411; append :online for web-augmented)

Note on Claude — spawn a subagent, don't call the API

The Claude adapter was dropped because calling Anthropic API from inside Claude Code would double-bill (lead session already authenticates against Claude). When you want a Claude second opinion — e.g., to compare Sonnet vs Opus, get an independent Claude perspective, or test a prompt in isolation — spawn a subagent from the lead session with Agent(subagent_type: "general-purpose", model: "sonnet"|"opus", prompt: "..."). The subagent gets its own context window, returns a fresh perspective, and reuses the harness auth.

Note on Gemini — sub-routing pending

The gemini CLI v0.27.3+ is installed but auth defaults to API key. To switch to sub-routing (rides on Google AI Pro / Code Assist plan instead of being separately billed):

1. Run gemini interactively once and choose Code Assist / Google account auth flow
2. Add export GOOGLE_GENAI_USE_GCA=1 to your shell rc
3. Once both are done, the adapter can be updated to prefer the CLI subprocess (forward-compat check). Currently uses the GEMINI_API_KEY HTTP API path.

Keys auto-load from ~/.openclaw-tgpkb/secrets/<provider>_api_key if env vars aren't set (see ~/.openclaw-tgpkb/secrets/llm_keys.sh).

Run scripts/auth-check.sh to verify each provider is reachable.

:reason — adversarial refinement subcommand

/second-opinion:reason runs a cold-start multi-agent loop modeled on /autoresearch:reason,
but with cross-provider judges instead of all-Claude subagents.

# default: 3 judges, convergence=3, mode=convergent, domain=software
~/.claude/skills/second-opinion/scripts/reason.sh "should we use event sourcing for orders"

# bounded debate (no synthesis), 5 judges
~/.claude/skills/second-opinion/scripts/reason.sh \
  --mode debate --judges 5 --iterations 4 \
  --domain product "monetize free tier vs paid-only launch"

Role mapping (defaults; degrades when fewer providers configured):

Role	Default provider	Why
Author-A	codex	Strong general reasoning, sub-routed
Critic	grok	Direct, low-hedge — fits adversarial role
Author-B	gemini	Different lineage / training distribution
Synthesizer	codex	Engineering-style merge
Judges	rotate {codex, grok, gemini, openrouter}	Cross-provider blind panel

Each invocation is a fresh adapter subprocess — no history bleed. Sequential execution
throughout (droplet 2GB constraint, no parallel calls).

Outputs land in reason-runs/{YYMMDD-HHMM}-{slug}/ with overview.md, lineage.md,
candidates.md, judge-transcripts.md, reason-results.tsv, reason-lineage.jsonl,
handoff.json. The handoff.json describes suggested chain targets but does not
auto-execute them (chains are autoresearch commands, separate skills).

Files

second-opinion/
├── SKILL.md                        # this file
├── reason-runs/                    # output dirs from /second-opinion:reason runs
└── scripts/
    ├── dispatch.sh                 # entry point for single-shot / fanout
    ├── classify.sh                 # keyword router
    ├── fanout.sh                   # parallel call + synthesize
    ├── reason.sh                   # adversarial refinement loop (:reason subcommand)
    ├── auth-check.sh               # verify each provider's auth
    └── adapters/
        ├── codex.sh                # OpenAI / Codex CLI (rides on ChatGPT sub)
        ├── grok.sh                 # xAI API (no CLI/sub available as of May 2026)
        ├── gemini.sh               # Google Gemini API (CLI sub-routing TODO — see Note on Gemini)
        └── openrouter.sh           # OpenRouter (refusal fallback / model marketplace)

Output convention

Every adapter prints to stdout in this format so fan-out + synthesis works:

=== provider: <name> ===
<response body>
=== end ===

Errors go to stderr with non-zero exit.

TODOs

• Deep-research mode is provider-specific; current adapters approximate it. Wire real APIs when available (Gemini Deep Research, OpenAI Deep Research, xAI Live Search).
• Synthesis prompt in fanout.sh is a stub — refine after first real run.
• Classifier is keyword-based; consider upgrading to a small Haiku call if accuracy matters.