opencode-review
Use when the user asks to review code/a PR with opencode or with a non-Claude model (e.g., "review with opencode", "review using gpt-5.6-sol", "/opencode-review"). Runs a fresh opencode session in a subagent, tells it to use the review-code skill, and relays its review.
How do I install this agent skill?
npx skills add https://github.com/lwlee2608/agent-skills --skill opencode-reviewIs this agent skill safe to install?
- Gen Agent Trust Hubwarn
This skill delegates code reviews to a separate subagent using the 'opencode' tool. It constructs shell commands by directly inserting user-provided inputs, such as branch names or model identifiers, into a script template. This pattern creates a risk of command injection if the inputs are not properly sanitized. Additionally, the skill is subject to indirect prompt injection risks from the code it reviews.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
Code Review via opencode
Delegate code review to a fresh opencode session running a VelociRouter model (default velocirouter/gpt-5.6-sol). The goal is a second opinion from a different model — don't duplicate the review yourself.
Assume opencode is installed and the review-code skill is available to it. Don't probe for either — if opencode run fails, surface the error and stop.
Rules
-
Resolve the review target. Hand
review-codeits own vocabulary; don't resolve the diff yourself.- User named a PR number →
pr <N> - User named a ref range (e.g.,
main...HEAD,abc123..def456) → use it verbatim - Otherwise →
diff
- User named a PR number →
-
Resolve the model. Default
velocirouter/gpt-5.6-sol. If the user names another model without a provider prefix, usevelocirouter/<name>(e.g.,gpt-5.6-sol-fast→velocirouter/gpt-5.6-sol-fast). If they give a fullprovider/model, use it verbatim. -
Run opencode inside a subagent — never from the main session.
opencode runstreams the model's reasoning and tool logs to stdout; that whole transcript lands in whichever context invoked it. Spawn one general-purpose subagent whose only job is to run the command and hand back the review. -
Start a fresh opencode session. Never pass
-c,-s, or--fork— reusing a prior session pollutes the review with unrelated context. That's the whole point of this skill. -
Give the subagent this task (substituting target and model):
Run the command below exactly as written, from the repo root. Do not review the code yourself, do not read the diff, do not edit files.
set -o pipefail OUT=$(mktemp /tmp/opencode-review-XXXXXX.md) opencode run \ -m <provider/model> \ --dir "$(pwd)" \ --title "review: <target>" \ "Use the review-code skill to review <target> in this repo. Do not modify any files." \ | tee "$OUT" rc=$? echo "raw output: $OUT (exit $rc)" exit $rcset -o pipefailis load-bearing: without it the pipeline reportstee's exit status and an opencode failure looks like success.Do not rename
rctostatus—statusis read-only in zsh, so the assignment aborts the script after opencode has already run, making a successful review look like a failure.If the exit status is non-zero, return the error output and nothing else. Otherwise return two things: the path printed on the last line, and the review itself copied verbatim — findings, ratings, and
path:linecitations exactly as opencode wrote them. Drop only the surrounding tool logs and reasoning chatter.--dir "$(pwd)"is required so opencode can rungitand read repo files.- Do not pass
--dangerously-skip-permissions— this is a read-only review. - Don't inline the diff into the prompt. The
review-codeskill resolves the target itself, and shell-escaping a large diff breaks quoting.
-
Relay the subagent's review verbatim to the user. Don't summarize, re-review, or second-guess it — the user asked for a different model's take, not yours. Mention the raw-output path so they can read the full transcript if they want it.
-
Surface errors loudly. If the command exited non-zero, print the error and stop. Don't silently fall back to a Claude-side review.
Verification procedure
- The command ran in a subagent, not the main session.
- If opencode rejects the model, check
opencode models— VelociRouter models are listed asvelocirouter/<name>. - The relayed review contains concrete
path:linecitations — if it doesn't, opencode probably never loaded the skill or read the diff, and it should be re-run.
Common mistakes to watch for
- Running
opencode runfrom the main session: floods the context with the model's full transcript. Always delegate. - Letting the subagent do the reviewing: it's a courier, not a reviewer. Its prompt must say so.
- Duplicating the review: Claude re-reviews after opencode returns. Don't — just relay.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/lwlee2608/agent-skills/opencode-review">View opencode-review on skillZs</a>