Check your own hypothesis against opencode, which reads the relevant code itself — for when retelling the context would be expensive or would frame the question toward your conclusion. For high-cost calls: an architectural decision, a contested finding, a second pass on a bug with no movement. Strictly read-only.
Installs into .claude/skills of the current project.
Are you the author of Opencode Second Opinion?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/dmitry-fomin-opencode-second-opinion)
---
name: opencode-second-opinion
description: "Check your own hypothesis against opencode, which reads the relevant code itself — for when retelling the context would be expensive or would frame the question toward your conclusion. For high-cost calls: an architectural decision, a contested finding, a second pass on a bug with no movement. Strictly read-only."
when_to_use: "Triggers — \"second opinion on this code\", \"have opencode check my conclusion\", \"what does opencode say about this bug\". An explicit request is consent to launch. Not for routine or single-answer questions, and not as a substitute for your own analysis — think first. If there is no hypothesis yet and you just need the area mapped, that is /opencode:opencode-delegate; questions about a running job are /opencode:opencode-jobs. If the hypothesis fits in a paragraph and involves little code, ask a model directly instead — a whole agentic run is overkill."
context: fork
allowed-tools: Bash(${CLAUDE_PLUGIN_ROOT}/scripts/opencode-run.sh *)
---
Request: $ARGUMENTS
This is `opencode-delegate` in consultation mode: opencode reads the code itself and reaches
its own conclusion. Mechanics, waiting and exit codes are in `/opencode:opencode-delegate`
and `opencode-runtime` — below is only what differs.
1. **Your own conclusion comes first.** Nothing to compare means nothing to delegate.
2. **Phrase the question independently of your conclusion** — a supplied hypothesis nearly
always gets confirmed.
3. **Scope the area** and forbid `.env`, `*.key`, `*.pem`, `credentials.json` in the task
text.
4. **If the human did not ask for the consultation**, show them the task text and working
directory and wait for agreement.
5. **Launch in the background on the default `--permission read`:**
`run --background --label "<hypothesis topic>"`. A second opinion must not change the
working tree, and `--permission bash` would make writing technically possible.
6. **Collect, compare against your own analysis, then report.** This skill runs forked
(`context: fork`), so only your final message reaches the conversation — put the job-id
in it.
One consultation per hypothesis. A second run on the same hypothesis adds a vote, not
knowledge — and the same hypothesis on both channels is still one consultation, not two
independent opinions; the second channel is for when the first answer looks unsound.
Different hypotheses or different areas can run in parallel.
Needs a follow-up on the same hypothesis: give the run `--session <name>` and continue with
`resume --session <name>` instead of restating the context.
## Reading the answer
- Label it as one external harness's opinion and name the channel — `glm` and `deepseek`
have different blind spots.
- **Instant, complete agreement deserves scepticism.** Models share blind spots and share
your framing. The divergence is the finding — report it first.
- `deepseek` answers get stricter checking: verify each claim against the code, and don't
pass unverified ones on.
- Confirm files were actually read: `transcript <job-id>`.
- Don't turn the review into an immediate code change — report to the human first. An edit
is a separate decision and a separate `/opencode:opencode-delegate` call.
