$npx -y skills add Undertone0809/rudder --skill agent-work-reviewer-maintainerUse when reviewing Rudder agent work, Codex sessions, PRs, commits, UI, releases, regressions, proposals, or agent outcomes for product correctness, evidence quality, scope, architecture, and handoff trust.
| 1 | # Agent Work Reviewer Maintainer |
| 2 | |
| 3 | Review completed or in-progress Rudder agent work. This is a reviewer workflow, |
| 4 | not an implementation workflow. |
| 5 | |
| 6 | The core question is: |
| 7 | |
| 8 | > Did the agent solve the right product problem, with the right object model, |
| 9 | > complete behavior, credible validation, and a clean handoff? |
| 10 | |
| 11 | Reviewer is not the same role as verifier. Reviewer may inspect the running |
| 12 | product when that is needed for judgment, but its durable responsibility is to |
| 13 | judge the diff, architecture, scope, test strategy, handoff safety, and the |
| 14 | credibility of acceptance evidence. For black-box acceptance of final product |
| 15 | behavior, use `product-acceptance-verifier-maintainer`. |
| 16 | |
| 17 | Default to Chinese when the user asks in Chinese. Keep the verdict early and |
| 18 | ground every judgment in evidence. |
| 19 | |
| 20 | ## Use When |
| 21 | |
| 22 | Use this skill when the user asks to review: |
| 23 | |
| 24 | - a Codex session, Rudder agent run, task, or transcript |
| 25 | - a local branch, commit, diff, pull request, or implementation |
| 26 | - a product proposal, plan doc, feature spec, or UI direction |
| 27 | - a release verification, Desktop install path, npm publish, or canary/stable |
| 28 | handoff |
| 29 | - a screenshot, browser state, visual interaction, or workflow that "feels |
| 30 | wrong" |
| 31 | |
| 32 | Common trigger phrases: |
| 33 | |
| 34 | - "review 一下 codex session id ..." |
| 35 | - "as a 专业产品经理 review" |
| 36 | - "第一性原理思考一下" |
| 37 | - "作为 reviewer" |
| 38 | - "这个实现是不是产品上对" |
| 39 | - "这个 PR 本地跑起来看一下有没有问题" |
| 40 | - "这次 release 验证做完了吗" |
| 41 | - "我觉得这个功能之前有,现在没了" |
| 42 | |
| 43 | ## Do Not Use When |
| 44 | |
| 45 | Do not use this skill for: |
| 46 | |
| 47 | - fixing the findings during the same reviewer pass, unless the user explicitly |
| 48 | asks to implement fixes |
| 49 | - generic line-by-line code review where product judgment is irrelevant |
| 50 | - debugging a failed Rudder agent run root cause only; use the run transcript |
| 51 | debugging workflow first, then return here for product/workflow judgment |
| 52 | - summarizing logs without verdict or acceptance criteria |
| 53 | |
| 54 | If the user asks to fix findings after the review, switch to normal |
| 55 | implementation mode and follow repository validation, commit, and push rules. |
| 56 | |
| 57 | ### Read-Only Guard |
| 58 | |
| 59 | Reviewer mode is read-only by default. Do not edit files, stage changes, |
| 60 | restore files, commit, push, start destructive cleanup, or "just fix" findings |
| 61 | while reviewing unless the user explicitly changes the task from review to |
| 62 | implementation. |
| 63 | |
| 64 | If the user says "do not implement", "review only", or assigns a reviewer role, |
| 65 | that instruction is binding for the whole reviewer pass. Use tools only to |
| 66 | inspect evidence. |
| 67 | |
| 68 | Direct UI inspection with Browser, Desktop, or Computer Use is still reviewer |
| 69 | work when it only observes or exercises low-risk local/dev flows. If a realistic |
| 70 | scenario requires mutating Rudder data, prefer an isolated dev org, disposable |
| 71 | test records, or an existing preview instance, and report what was created or |
| 72 | changed. Do not delete data, publish, submit external communication, install |
| 73 | software, change system settings, or perform other risky UI actions without the |
| 74 | appropriate user confirmation. |
| 75 | |
| 76 | ## Evidence Packet |
| 77 | |
| 78 | Never start with opinion. Build the smallest evidence packet that can support a |
| 79 | real judgment. |
| 80 | |
| 81 | ### 1. Identify The Target |
| 82 | |
| 83 | Resolve what is being reviewed: |
| 84 | |
| 85 | - Codex session id or prefix |
| 86 | - Rudder run id or transcript |
| 87 | - PR number or URL |
| 88 | - branch name, commit hash, staged/unstaged diff |
| 89 | - plan doc, proposal, screenshot, or browser URL |
| 90 | - release version, tag, workflow run, npm dist-tag, or Desktop asset set |
| 91 | |
| 92 | If the user is vague, infer from current branch, recent commits, open browser |
| 93 | state, or named files before asking. |
| 94 | |
| 95 | ### 2. Collect Task Intent |
| 96 | |
| 97 | For Codex sessions, search: |
| 98 | |
| 99 | ```bash |
| 100 | rg "<session-id-or-prefix>" ~/.codex/session_index.jsonl ~/.codex/sessions ~/.codex/archived_sessions |
| 101 | ``` |
| 102 | |
| 103 | Extract real user requests and corrections. Ignore injected `AGENTS.md`, |
| 104 | environment context, skill bodies, and system/developer text. |
| 105 | |
| 106 | For branches, PRs, commits, or diffs, inspect: |
| 107 | |
| 108 | ```bash |
| 109 | git status --short --branch |
| 110 | git log --oneline --decorate -12 |
| 111 | git diff --stat |
| 112 | git diff |
| 113 | git show --stat <commit> |
| 114 | git show <commit> |
| 115 | ``` |
| 116 | |
| 117 | For commit or session reviews, also compare the changed-file set against the |
| 118 | stated task. Classify every surprising file as one of: |
| 119 | |
| 120 | - required for the requested behavior |
| 121 | - test, docs, or contract evidence for the requested behavior |
| 122 | - pre-existing unrelated dirty work |
| 123 | - unrelated change mixed into the reviewed commit or handoff |
| 124 | |
| 125 | Unrelated changes mixed into a product fix are review findings, not cleanup |
| 126 | details. If they change skills, release state, generated files, dependencies, |
| 127 | or broad runtime behavior outside the task, usually treat that as at least a |
| 128 | `conditional accept` blocker until the scope is split or justified. |
| 129 | |
| 130 | For PRs, read the PR description, changed files, review comments, and CI status |
| 131 | when available. |
| 132 | |
| 133 | ### 3. Read Product Context |
| 134 | |
| 135 | For |