Second opinion

An independent look by another model (DeepSeek) at a risky step. Your own Claude subagent would give correlated mistakes — the value is precisely in a different model.

THIS IS NOT a replacement for code review. code-reviewer stays mandatory and primary. The second opinion is an extra check on top, for the important/contested. It cancels nothing in the cycle.

When to call

• before an irreversible step (a migration with data loss, a merge into main, a wide refactor);
• on a high-stakes architectural decision (a new seam/contract, a choice of approach);
• when the agent and code-reviewer diverge or you are unsure yourself;
• by the owner's explicit command ("take a second opinion on X").

Do NOT call on routine (small edits, low cost of error) — it costs time/money.

How to call

Two ways — same logic (second_opinion()), pick by taste:

Native MCP tool (if the MCP server is connected — .mcp.json in the plugin): call the second_opinion tool with arguments question (required) + context (+ optional model). No Bash-command construction — Claude calls it itself.

CLI (always works, without MCP):

python3 "${CLAUDE_PLUGIN_ROOT}/bin/second_opinion.py" "a concrete question" <<< "relevant context (diff/plan/finding)"
# default model deepseek-reasoner; for a fast one — --model deepseek-chat

• Question — concrete ("is this seam drawn correctly?", "what am I missing in this refactor?").
• Context — the target chunk (diff, plan, a contested finding), NOT the whole repository.
• No key/an error → it says so gently and does not crash the cycle (it is an extra check) — same in CLI and MCP.

The DeepSeek key — env DEEPSEEK_API_KEY or ~/.config/retainer/deepseek_key (chmod 600), outside git.

How to accept the answer

An opinion = a hypothesis, not an authority. Read → check against the code/reality → weigh → decide yourself. Not "DeepSeek said so — I do it". [[receiving-review]] applies (verify before adopting, one at a time, unclear → clarify). The synthesis is yours.

When NOT to apply

Not on every task — only an important/irreversible step or a hard bug (the cost of error is high). Do not run routine through a second opinion. And it is an ADDITION to code-reviewer, NOT a replacement (your own reviewer first, then the cross-model one).

Related: [[task-cycle]] (an important/irreversible step), [[diagnosing]] (a second opinion on a hard bug).