second-opinion

name	second-opinion
description	Get an independent cross-model second opinion (GLM-5.2 via opencode) on a specific, load-bearing technical claim or a Critical code-review finding — a different model reads the claim + code and returns AGREE / DISAGREE / UNSURE. Use to stress-test a single Critical finding, a risky assertion, or a close call before you commit to it. Auto-fired inside /pr-review on each Critical finding that survives the skeptic. Not for open-ended research (use /deep-research) or multi-lens judgement (use /llm-council).
argument-hint	<a claim to test, optionally with a file:line to pull context from>

Second Opinion (cross-model)

Input: $ARGUMENTS

Ask a different model — GLM-5.2, run through the local opencode CLI on the OpenCode Go subscription — to independently judge one concrete claim. Claude and GLM fail in different ways; a second architecture disagreeing is a real signal, an agreement is corroboration. This is the on-demand cross-model diversity the harness reserves for high-stakes calls (research.md locked decision: no standing consensus voting — the adversarial skeptic already beats that; bring GLM only when it pays).

The model is GLM-5.2 on OpenCode Go (opencode-go/glm-5.2) — the top open-weight coder as of build time, ~1/6 Claude's cost, vetted full-precision weights on the Go plan. The slug was confirmed live against opencode's own catalog when this skill was built.

Operating rules

• Interactive-only, degrades to a no-op. opencode is a local CLI, absent from CI runners (and unfunded / rate-limited plans, and offline). Every one of those returns UNAVAILABLE and the skill says so and moves on — it never blocks, never invents a verdict, never treats "couldn't ask" as "disagreed." Same CI-fallback discipline as the audit skills: when the tool isn't there, the work still completes without it.
• Never force a decision. A GLM verdict is input, not a gate. It never auto-merges, auto-drops a finding, or overrides Claude's own judgement — it surfaces a second view for a human to weigh.
• One claim per call. Feed a single, self-contained finding + its code. GLM judges only from the text you send it — no repo access — so include the cited hunk.

How it runs — the helper

The mechanics (invoke opencode, parse its JSONL event stream, extract the verdict, degrade on any failure) live in a deterministic helper so nothing is left to per-run improvisation:

node .claude/skills/second-opinion/second-opinion.mjs <<'FINDING'
<the finding dossier: title · severity · repo/path:line · the claimed defect · the cited code/diff hunk>
FINDING

It reads the dossier from stdin (no argv length / quoting limits — diffs are safe), prompts GLM as an independent skeptic, and prints one line of JSON to stdout, always exiting 0:

Field	Meaning
status	OK (a verdict was obtained) or UNAVAILABLE (opencode absent / capped / errored / unparseable)
verdict	AGREE · DISAGREE · UNSURE (only when status: OK)
confidence	high · medium · low
reasoning	≤ 2 sentences citing the specific code
reason	why it degraded (only when status: UNAVAILABLE)
model	the slug used

Options: --model <slug> (default opencode-go/glm-5.2; swap only to a live opencode slug), --timeout <ms> (default 180000). GLM is verbose/slow — the timeout is a backstop that yields UNAVAILABLE, never a hang.

Interpreting the verdict

Result	What it means	What to do
OK · AGREE	An independent model confirms the defect and the severity.	Corroborated — state that the finding is cross-model confirmed.
OK · DISAGREE	GLM argues the code is correct / the severity inflated / the claim unsupported.	Mark the finding CONTESTED; surface both verdicts (Claude's + GLM's reasoning) and let the human decide. Do not silently drop it and do not force a merge.
OK · UNSURE	The context couldn't decide it.	Note it; the finding stands as Claude's review already ruled.
UNAVAILABLE	No second opinion was obtained.	Say so in one line (with the reason); the finding stands unchanged. Never read this as agreement or disagreement.

Standalone use

For a /second-opinion <claim> invocation outside a review:

1. Build the dossier: the claim in one line, plus — if $ARGUMENTS names a file:line or a snippet — read that context and include the relevant hunk so GLM judges the real code.
2. Run the helper.
3. Report your read of the claim and GLM's verdict side by side. On DISAGREE, present both cases and recommend how to resolve; on UNAVAILABLE, answer from your own analysis and note the second opinion wasn't reachable.

Inside /pr-review

/pr-review Phase 6 fires this on each Critical finding that survives the adversarial skeptic (interactive runs only). The contract there is identical: DISAGREE → the finding is tagged CONTESTED with both verdicts shown, the recommendation is never forced, and UNAVAILABLE leaves the finding exactly as the skeptic left it.