nova-codex-review-fix

Inputs

Parameter	Required	Default	Notes
REVIEW_MODE	No	branch	Review scope: branch, staged, or full. Maps to default, --only-staged, or --full.
BASE	No	Auto-detect	Baseline branch passed to review/verify scripts as --base.
OUTPUT_DIR	No	Script default	Review/verify artifact directory passed to Codex scripts as --output-dir.
GOAL	No	Fix high-confidence blockers	Natural-language goal for the closure loop.
FIX_SCOPE	No	high-confidence	Policy scope for selecting review findings to fix; does not widen script behavior.

Parameter Resolution

• Parse natural-language payload, explicit KEY=value, --flag value, and --flag=value forms from $ARGUMENTS.
• Normalize parameter names to uppercase snake case and map known mode words before assigning remaining text to GOAL.
• Explicit values win over inferred values only when they do not conflict with another explicit value.
• Apply documented defaults only when unambiguous; probe Git status, base branches, and latest artifacts only for context parameters.
• Safety-boundary parameters for this skill: none for this skill.
• In non-interactive mode, fail before side effects when required or safety-boundary parameters are missing.
• Full policy: nova-plugin/skills/_shared/parameter-resolution.md.

Safety Preflight

• This skill declares side-effect-capable tools: Bash, Edit, Write.
• Resolve parameters and present a preflight card before writing artifacts, editing project files, or running Bash.
• Show files or artifacts that may be written, scripts or commands that may run, disallowed operations, and the proceed condition.
• Do not infer missing safety-boundary values; ask once in interactive mode or fail in non-interactive mode.
• Preserve repository constraints: no destructive Git cleanup, no branch deletion, no push/merge/rebase, no editing archived agents as active agents.
• Full policy: nova-plugin/skills/_shared/safety-preflight.md.

Codex Script Argument Mapping

• REVIEW_MODE=branch runs codex-review.sh without scope flags.
• REVIEW_MODE=staged adds --only-staged; REVIEW_MODE=full adds --full.
• BASE is passed as --base <BASE> to both review and verify scripts.
• OUTPUT_DIR, when provided, is passed as --output-dir <OUTPUT_DIR> to review and verify.
• If OUTPUT_DIR is provided, use <OUTPUT_DIR>/review.md for the fix step and write local checks to <OUTPUT_DIR>/artifacts/checks.txt so codex-verify.sh can auto-detect them when --checks-file is not passed.
• If OUTPUT_DIR is not provided, use the script's latest artifacts paths under .codex/codex-review-fix/latest-artifacts/.
• GOAL and FIX_SCOPE guide Claude Code's fix selection and summary; they are not script flags.

Outputs

• Follow the skill-specific output rules below and the shared output contract.
• For written artifacts, report the path and a short executive summary instead of pasting the full artifact into chat.
• For reviews and verification, lead with findings or verdicts and state residual risk.
• Full policy: nova-plugin/skills/_shared/output-contracts.md.
• Artifact policy: nova-plugin/skills/_shared/artifact-policy.md.

Workflow

1. Resolve parameters using the shared policy and this skill's input table.
2. Read only the context needed for the requested scope.
3. Apply the skill-specific guidance and migrated slash command contract below.
4. Respect safety preflight before any side effects.
5. Produce the required output and report validation or skipped validation honestly.

Failure Modes

• Required payload is missing or ambiguous.
• A safety-boundary parameter is missing, conflicting, or unsafe to infer.
• Required files, scripts, CLIs, credentials, or runtime dependencies are unavailable.
• Existing user changes overlap the intended write scope and cannot be merged safely.
• Repository policy conflicts with the requested action.

Examples

• Use /codex-review-fix as the full Codex review -> Claude Code fix -> checks -> verify loop.
• Explicit parameters may use KEY=value or --flag value; natural-language payload is accepted when unambiguous.

Common Rationalizations

Rationalization	Required Response
"This is small enough to skip validation."	Run the focused check or state why it is unavailable.
"The existing output contract is obvious."	Follow the shared output contract and the skill-specific output format exactly.
"The nearby cleanup is harmless."	Keep scope to the requested execution basis and note unrelated cleanup separately.
"A plausible result is enough."	Report command evidence, artifact paths, or an explicit skipped-check reason.

Red Flags

• Scope expands beyond the requested execution basis.
• Validation is claimed without command evidence, artifact evidence, or a skipped-check reason.
• Existing user changes are overwritten, normalized, or reformatted without being part of the task.
• The output omits required residual risk, deviations, or follow-up notes.
• The skill uses tools outside its declared safety boundary.

Verification

• Inputs were resolved through the shared parameter policy.
• Safety preflight was respected before side effects.
• Relevant checks were run or explicitly skipped with reason.
• Existing user changes were preserved unless explicitly in scope.
• Output follows the shared output contract and skill-specific format.

Skill-Specific Guidance

目的

执行一个面向当前分支的半自动闭环：

1. 运行 Codex review 脚本产出结构化问题报告。
2. 阅读 必须修、建议修 中高置信问题。
3. 由 Claude Code 修改代码，必要时补测试。
4. 运行本地检查脚本。
5. 运行 Codex verify 脚本确认问题是否解决。
6. 输出闭环结论、剩余风险、下一步建议。

何时使用

• 当前分支已经有改动，需要先 review 再修复。
• 希望让 Claude Code 负责修复与编排，Codex 只做 reviewer / verifier。
• 需要一套可沉淀到仓库、可复制到其他项目的 review-fix workflow。

资源导航

• 主说明：README.md
• Prompt 模板：prompts/*.prompt.md
• 外部脚本：scripts/*.sh

只有在需要具体调用方式、参数细节或排障说明时再读取 README.md。

执行步骤

1. 先确认当前目录是 Git 仓库，并识别默认基线分支。
2. 运行 scripts/codex-review.sh。
3. 读取输出目录中的 review.md，只提取高置信、高优先级问题。
4. 参考 prompts/claude-fix.prompt.md 组织修复，不做与问题无关的大改。
5. 如问题涉及行为缺口，优先补单测或集成校验。
6. 运行 scripts/run-project-checks.sh --all --report-file <checks-file>。默认使用 .codex/codex-review-fix/latest-artifacts/checks.txt；显式 OUTPUT_DIR 时使用 <OUTPUT_DIR>/artifacts/checks.txt。
7. 运行 scripts/codex-verify.sh --review-file <review.md> --checks-file <checks-file>。
8. 读取 verify.md，按“已解决 / 未解决 / 不确定 / 新增高风险问题”总结。

优先级策略

• 先修 必须修 中会导致错误行为、数据风险、异常路径缺口、测试缺失的条目。
• 再处理 建议修 中高置信、低改动成本、能显著降低回归风险的条目。
• 可忽略 不默认处理，除非顺手即可消除且不会扩大改动范围。

安全边界

• 不执行 git reset --hard、git clean -fd、批量删除等危险命令。
• 不把 Codex 产物当成绝对真理；对低置信问题保持克制。
• 不因为 review 报告而主动重构无关模块。
• 本技能默认只操作当前仓库工作区，不负责发布、部署或远程环境变更。

输出要求

• Review 阶段：保留 review.md。
• 修复阶段：总结已修改内容、未修改原因、补充测试情况。
• 校验阶段：保留 checks 输出摘要。
• Verify 阶段：保留 verify.md。
• 收尾：输出本轮是否建议合并、剩余阻塞项、后续建议。

变体

• nova-codex-review-only：只做 review，不进入修复闭环。
• nova-codex-verify-only：基于已有 review.md 做 verify。

Migrated Slash Command Contract

Migrated from the pre-thin slash command contract for /codex-review-fix (nova-plugin/commands/codex-review-fix.md).

CODEX REVIEW -> FIX -> VERIFY LOOP

你是 Claude Code，扮演 fixer / orchestrator。

本命令用于当前分支的半自动闭环：

1. 运行 Codex review
2. 读取 review 结果
3. 只修复高置信、高优先级问题
4. 运行本地 checks
5. 运行 Codex verify
6. 输出本轮闭环结论

---

输入参数

从 $ARGUMENTS 中提取以下信息：

• BASE：可选，review/verify 基线分支，默认自动识别
• GOAL：可选，本轮修复目标，例如“修到可合并”
• REVIEW_MODE：可选，branch / staged / full

如果未提供，按当前仓库状态做最保守且最合理的选择。

---

强制执行流程

第一步：运行 review 脚本

先确认 CLAUDE_PLUGIN_ROOT 可用；如果不可用，必须停止并提示插件未正确启用。

执行：

bash "${CLAUDE_PLUGIN_ROOT}/skills/nova-codex-review-fix/scripts/codex-review.sh"

根据参数决定是否追加：

• --base <BASE>
• --only-staged
• --full

第二步：读取 review 结果

review 脚本会把结果写入 .codex/codex-review-fix/latest-artifacts/review.md（同时保留时间戳目录下的副本）。从该路径读取，只提取：

• 必须修
• 建议修 中高置信且高收益的问题

忽略低置信、纯风格、会扩大改动范围的问题。

第三步：实施修复

• 只改与问题直接相关的代码
• 必要时补测试
• 不做无谓大改或额外重构

第四步：运行本地验证

执行：

bash "${CLAUDE_PLUGIN_ROOT}/skills/nova-codex-review-fix/scripts/run-project-checks.sh" --all --report-file .codex/codex-review-fix/latest-artifacts/checks.txt

第五步：运行 verify 脚本

执行：

bash "${CLAUDE_PLUGIN_ROOT}/skills/nova-codex-review-fix/scripts/codex-verify.sh" --review-file .codex/codex-review-fix/latest-artifacts/review.md --checks-file .codex/codex-review-fix/latest-artifacts/checks.txt

第六步：输出总结

必须输出：

• 已修复问题
• 未修复问题及原因
• 本地验证结果
• Verify 结论
• 是否建议合并
• 剩余阻塞项
• 下一步建议

---

严格限制

你必须：

• 优先修复高风险 / 高置信问题
• 保持改动最小闭环
• 如实报告验证结果

你不得：

• 把 Codex 的所有建议都强行落地
• 伪造检查通过
• 执行高风险 Git 清理操作
• 因 review 报告而主动重构无关模块

---

结束条件

当以下条件满足时，本轮可视为闭环完成：

1. review 已生成
2. 代码修复已完成或明确受阻
3. 本地 checks 已执行
4. verify 已生成
5. 最终结论已输出