ワンクリックでManusで任意のスキルを実行

adversary-code-review-opencode

スター1

フォーク3

更新日2026年5月31日 02:52

Use when code changes need independent adversarial review from a different model. Accepts git SHA range and optional spec file for compliance checking.

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

Manusで実行

ソース

ronsanzone

ronsanzone/agent-essentials

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Manusで実行

Adversary Code Review (OpenCode)

Route code changes to opencode for adversarial review. The reviewing model assumes the implementation has defects and independently verifies correctness — not trusting claims, commit messages, or test results at face value.

This skill composes on opencode-consult. It constructs the adversarial prompt; opencode-consult handles the transport.

When to Use

After implementation — before merging or creating a PR
When you want a genuinely independent code review (different model, different biases)
When Claude already reviewed the code and you want a second opinion
With --spec when you need to verify implementation matches a spec or plan

When NOT to use:

For quick reviews — use quick-review or pr-review (faster, no server dependency)
For plan review — use adversary-plan-review-opencode instead
When the diff is very large (>500 changed lines) — split into logical chunks first

Arguments

Parse $ARGUMENTS for:

SHA range — in base..head format (e.g., main..HEAD, abc123..def456)
--spec flag — optional path to a plan or spec file for compliance checking

If no arguments provided, default to main..HEAD.

Examples:

/adversary-code-review-opencode -> reviews main..HEAD
/adversary-code-review-opencode abc123..def456 -> reviews that range
/adversary-code-review-opencode main..HEAD --spec docs/plan.md -> reviews with spec compliance

Process

Step 1: Resolve git range

Parse the SHA range from arguments. Default to main..HEAD if not provided.

Split on the .. delimiter (not individual dots — refs can contain dots):

BASE="${RANGE%%%..*}"
HEAD="${RANGE##*..}"

Verify both refs are valid:

git rev-parse --verify "$BASE" >/dev/null 2>&1
git rev-parse --verify "$HEAD" >/dev/null 2>&1

If invalid: "Invalid git range $RANGE. Use format base..head." Stop.

Get the diff stat to understand scope:

git diff --stat "$BASE".."$HEAD"

If empty diff: "No changes found in range $RANGE." Stop.

Step 2: Resolve spec file (if provided)

If --spec was passed, verify the file exists. If not: "Spec file not found at $SPEC_FILE." Stop.

Step 3: Build the inline context string

The adversarial rubric (review categories, calibration rules, severity, output format) lives in the static file prompt.md alongside this skill. The inline string provides the dynamic context — what to review and how.

Replace <BASE> and <HEAD> with actual resolved values.

Base context (always included):

You are performing an adversarial code review. The implementation may be incomplete,
incorrect, or subtly broken. Do not assume it works — verify independently.

DO NOT: Trust commit messages. Assume tests prove correctness. Take any claim at face value.
DO: Run git diff <BASE>..<HEAD>. Read changed files in full. Trace logic paths. Check edge cases.

Spec compliance addendum (appended when --spec is provided):

ADDITIONAL: Read the spec at <SPEC_FILE>. Verify every requirement has a corresponding
implementation, nothing was silently dropped, and nothing extra was added beyond spec.
Add a Spec Compliance section to your output between Files Reviewed and Critical Issues.

Step 4: Dispatch via opencode-consult

Spawn a general-purpose subagent that follows the opencode-consult dispatch pattern (server check, --dir, --file, model env var, output noise filtering).

The subagent must:

Follow opencode-consult Step 2 to verify the server and resolve PROJECT_DIR and MODEL.

Run the consultation with the static rubric file:

SKILL_DIR="$HOME/.claude/skills/adversary-code-review-opencode"
opencode run --attach "$SERVER_URL" --dir "$PROJECT_DIR" -m "$MODEL" \
  "<inline context string from Step 3>" \
  --file "$SKILL_DIR/prompt.md"

Filter tool call noise from the output (per opencode-consult guidance) and return:

## OpenCode Adversarial Code Review
**Model:** <model used>
**Range:** <base>..<head>
**Spec:** <spec file path, or "none">

### Review Result
<Structured review output only — verdict, findings, strengths. No tool call traces.>

Step 5: Present result

Display the subagent's structured result to the user. Add guidance based on verdict:

APPROVE: "OpenCode adversarial review found no blocking issues."
REQUEST CHANGES: "OpenCode found issues that should be addressed. Review findings above and decide how to proceed."

The user decides what to do with the findings. This skill does not modify anything.

Configuration

Setting	Default	Override
Model	`github-copilot/gpt-5.4`	`OPENCODE_REVIEW_MODEL` env var
Port	`4096`	First argument to `check-server.sh`

Common Mistakes

Mistake	Fix
Reviewing a huge diff (500+ lines)	Split into logical chunks. Large diffs overwhelm the reviewing model — findings get shallow.
Omitting `--spec` when a spec exists	Without spec compliance, the review only checks code quality — not whether you built the right thing.
Treating opencode findings as authoritative	OpenCode is advisory. Findings may be wrong — the reviewing model lacks Claude's full context. Verify before acting.
Using `main..HEAD` when main is stale	Make sure main is up to date (`git fetch origin main`) or the diff will include changes already merged.

Failure Handling

Failure	What to Tell the User
`check-server.sh` exits 1	"OpenCode server is not running on port 4096. Start it before retrying."
`git diff --stat` returns empty	"No changes found in range. Verify the SHA range is correct."
`opencode run` returns empty output	"OpenCode returned no response. The server may have timed out. Check http://localhost:4096 for session details."
`opencode run` returns truncated output	"OpenCode response appears incomplete. The diff may be too large. Try a smaller SHA range or split the review."
Subagent fails for any other reason	"Consultation failed. Check that opencode is configured correctly (`opencode auth list`) and the server is healthy."

このリポジトリの他の Skills

同じリポジトリ

bruh

ronsanzone/agent-essentials

Restate the last message in plain human language, with no jargon.

2026-07-121

facts

ronsanzone/agent-essentials

Give a capable engineer a concise briefing on the critical facts and implications of an effort before it is built or after it is completed. Use when the human explicitly invokes $facts, and never otherwise; do not invoke this skill automatically.

2026-07-121

vendor-skills-repo

ronsanzone/agent-essentials

Vendor skills from a GitHub repository into this repo. Use when asked to import, vendor, list, add, or update skills from another skills repository while preserving provenance and detecting local edits.

2026-07-061

investigate-and-fix

ronsanzone/agent-essentials

Use when given a well-scoped bug fix or small feature ticket from Jira. Investigates the ticket, researches the codebase, proposes a fix, then plans and implements on approval. Lighter alternative to the full deep-work pipeline for tickets that don't need multi-session design discussion.

2026-07-061

refine-ticket

ronsanzone/agent-essentials

Use when starting work on a ticket that needs refinement before entering the deep-work pipeline. Reads from Jira, pasted text, or a file, then interactively refines it into a structured ticket.md.

2026-07-061

session-retrospective

ronsanzone/agent-essentials

Use when a session is ending and you want to analyze process efficiency. Triggers on requests to review how a session went, evaluate context usage, or identify workflow improvements.

2026-07-061

name	adversary-code-review-opencode
description	Use when code changes need independent adversarial review from a different model. Accepts git SHA range and optional spec file for compliance checking.

Adversary Code Review (OpenCode)

This skill composes on opencode-consult. It constructs the adversarial prompt; opencode-consult handles the transport.

When to Use

After implementation — before merging or creating a PR
When you want a genuinely independent code review (different model, different biases)
When Claude already reviewed the code and you want a second opinion
With --spec when you need to verify implementation matches a spec or plan

When NOT to use:

For quick reviews — use quick-review or pr-review (faster, no server dependency)
For plan review — use adversary-plan-review-opencode instead
When the diff is very large (>500 changed lines) — split into logical chunks first

Arguments

Parse $ARGUMENTS for:

SHA range — in base..head format (e.g., main..HEAD, abc123..def456)
--spec flag — optional path to a plan or spec file for compliance checking

If no arguments provided, default to main..HEAD.

Examples:

/adversary-code-review-opencode -> reviews main..HEAD
/adversary-code-review-opencode abc123..def456 -> reviews that range
/adversary-code-review-opencode main..HEAD --spec docs/plan.md -> reviews with spec compliance

Process

Step 1: Resolve git range

Parse the SHA range from arguments. Default to main..HEAD if not provided.

Split on the .. delimiter (not individual dots — refs can contain dots):

BASE="${RANGE%%%..*}"
HEAD="${RANGE##*..}"

Verify both refs are valid:

git rev-parse --verify "$BASE" >/dev/null 2>&1
git rev-parse --verify "$HEAD" >/dev/null 2>&1

If invalid: "Invalid git range $RANGE. Use format base..head." Stop.

Get the diff stat to understand scope:

git diff --stat "$BASE".."$HEAD"

If empty diff: "No changes found in range $RANGE." Stop.

Step 2: Resolve spec file (if provided)

If --spec was passed, verify the file exists. If not: "Spec file not found at $SPEC_FILE." Stop.

Step 3: Build the inline context string

Replace <BASE> and <HEAD> with actual resolved values.

Base context (always included):

You are performing an adversarial code review. The implementation may be incomplete,
incorrect, or subtly broken. Do not assume it works — verify independently.

DO NOT: Trust commit messages. Assume tests prove correctness. Take any claim at face value.
DO: Run git diff <BASE>..<HEAD>. Read changed files in full. Trace logic paths. Check edge cases.

Spec compliance addendum (appended when --spec is provided):

ADDITIONAL: Read the spec at <SPEC_FILE>. Verify every requirement has a corresponding
implementation, nothing was silently dropped, and nothing extra was added beyond spec.
Add a Spec Compliance section to your output between Files Reviewed and Critical Issues.

Step 4: Dispatch via opencode-consult

Spawn a general-purpose subagent that follows the opencode-consult dispatch pattern (server check, --dir, --file, model env var, output noise filtering).

The subagent must:

Follow opencode-consult Step 2 to verify the server and resolve PROJECT_DIR and MODEL.

Run the consultation with the static rubric file:

SKILL_DIR="$HOME/.claude/skills/adversary-code-review-opencode"
opencode run --attach "$SERVER_URL" --dir "$PROJECT_DIR" -m "$MODEL" \
  "<inline context string from Step 3>" \
  --file "$SKILL_DIR/prompt.md"

Filter tool call noise from the output (per opencode-consult guidance) and return:

## OpenCode Adversarial Code Review
**Model:** <model used>
**Range:** <base>..<head>
**Spec:** <spec file path, or "none">

### Review Result
<Structured review output only — verdict, findings, strengths. No tool call traces.>

Step 5: Present result

Display the subagent's structured result to the user. Add guidance based on verdict:

APPROVE: "OpenCode adversarial review found no blocking issues."
REQUEST CHANGES: "OpenCode found issues that should be addressed. Review findings above and decide how to proceed."

The user decides what to do with the findings. This skill does not modify anything.

Configuration

Setting	Default	Override
Model	`github-copilot/gpt-5.4`	`OPENCODE_REVIEW_MODEL` env var
Port	`4096`	First argument to `check-server.sh`

Common Mistakes

Mistake	Fix
Reviewing a huge diff (500+ lines)	Split into logical chunks. Large diffs overwhelm the reviewing model — findings get shallow.
Omitting `--spec` when a spec exists	Without spec compliance, the review only checks code quality — not whether you built the right thing.
Treating opencode findings as authoritative	OpenCode is advisory. Findings may be wrong — the reviewing model lacks Claude's full context. Verify before acting.
Using `main..HEAD` when main is stale	Make sure main is up to date (`git fetch origin main`) or the diff will include changes already merged.

Failure Handling

Failure	What to Tell the User
`check-server.sh` exits 1	"OpenCode server is not running on port 4096. Start it before retrying."
`git diff --stat` returns empty	"No changes found in range. Verify the SHA range is correct."
`opencode run` returns empty output	"OpenCode returned no response. The server may have timed out. Check http://localhost:4096 for session details."
`opencode run` returns truncated output	"OpenCode response appears incomplete. The diff may be too large. Try a smaller SHA range or split the review."
Subagent fails for any other reason	"Consultation failed. Check that opencode is configured correctly (`opencode auth list`) and the server is healthy."