ci-triage

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

ci-triage

Shared logic for classifying a failed CI workflow run and recording the outcome. Used by humans (or Claude in an interactive session) when triaging a red

main

workflow or a red check on an open PR — the latter via the consumer repo's

<repo>-pr-lifecycle

skill (e.g.

onsager-pr-lifecycle

lean-spec-pr-lifecycle

duhem-pr-lifecycle

This skill owns the taxonomy, the de-dup rules for the

main-red

issue, and the issue template. Workflow-specific reproduction steps live in the consumer repo's

<repo>-pr-lifecycle

skill (for build-tool-specific detail) and in

web-testing

(for e2e).

用于分类失败的CI工作流运行并记录结果的共享逻辑。当处理

main

分支故障工作流或开放PR上的故障检查时，由人工（或交互式会话中的Claude）使用——后者通过消费仓库的

<repo>-pr-lifecycle

技能调用（例如

onsager-pr-lifecycle

、

lean-spec-pr-lifecycle

、

duhem-pr-lifecycle

）。

该技能负责分类体系、

main-red

议题的去重规则以及议题模板。工作流特定的复现步骤存放在消费仓库的

<repo>-pr-lifecycle

技能中（针对构建工具的细节）和

web-testing

技能中（针对e2e测试）。

Taxonomy

分类体系

Every failure lands in exactly one bucket. Be explicit — "unclear" is not a bucket,

needs-human

is.

Bucket	Signal	Default action
`regression`	Reproduces deterministically on `main` at HEAD	File/update `main-red` , @-mention suspect author
`flake`	Same workflow passed on the previous main commit without code change	Comment on the existing `main-red` issue if one is open; otherwise skip
`infra`	Postgres service didn't come up, rustup 403, pnpm registry down	File/update `main-red` labelled `infra` ; do not @-mention authors
`needs-human`	Logs truncated, auth failed, classification genuinely ambiguous	Open a `main-red` issue with raw log excerpt, label `needs-human`

Do not invent a fifth bucket. If the signal you see fits nothing above, it's

needs-human

每个故障都属于且仅属于一个分类。分类需明确——“不明确”不属于分类，

needs-human

才是。

Bucket	信号描述	默认操作
`regression`	在 `main` 分支HEAD节点可确定性复现	创建/更新 `main-red` 议题，@提及可疑提交的作者
`flake`	未修改代码的情况下，相同工作流在上一个main分支提交时运行成功	若已有开放的 `main-red` 议题则添加评论；否则跳过
`infra`	Postgres服务未启动、rustup 403错误、pnpm registry宕机等	创建/更新带 `infra` 标签的 `main-red` 议题；不@提及作者
`needs-human`	日志被截断、认证失败、分类结果确实不明确	创建包含原始日志片段的 `main-red` 议题，添加 `needs-human` 标签

请勿新增第五个分类。如果遇到的信号不符合上述任何一种，归为

needs-human

。

Suspect commit identification

可疑提交识别

The GitHub

workflow_run

payload includes

head_sha

. That's the commit that triggered this run — on a

push: main

event, it is the merge commit.

Fetch the commit via
```
mcp__github__get_commit
```
.
If the commit message matches
```
Merge pull request #N
```
, the suspect PR is
```
#N
```
; its author is the suspect author.
Otherwise (direct push to main, squash-merge), use the commit author directly.

Never blame more than one commit per failure. If the previous main commit's CI was also red, link that issue rather than opening a new one.

GitHub的

workflow_run

负载包含

head_sha

，即触发本次运行的提交——在

push: main

事件中，它是合并提交。

通过
```
mcp__github__get_commit
```
获取提交信息。
如果提交信息匹配
```
Merge pull request #N
```
，则可疑PR为
```
#N
```
；其作者为可疑作者。
其他情况（直接推送到main分支、 squash合并），直接使用提交作者。

每次故障最多指向一个提交。如果上一个main分支提交的CI也失败，链接到该议题而非创建新议题。

The rolling

main-red

issue

滚动

main-red

议题

One open
main-red
issue at a time. If main is broken for three days, that's one issue accumulating comments — not twelve.

Before filing:

mcp__github__list_issues

with

labels: main-red, state: open

If one exists, append a comment:
Run #<run-id> also failed. Workflow
```
<name>
```
, bucket
```
<bucket>
```
, suspect <sha-short> (#<pr-or-none>, @<author>). <one-line-cause>.
Only open a new issue if none is open. Title:
```
main is red: <workflow> (<bucket>)
```
. Labels:
```
main-red
```
, plus
```
infra
```
or
```
needs-human
```
if applicable.

When main goes green again (the next successful run on the same workflow), close the issue with a comment naming the green run id. This close step is manual.

同一时间仅保留一个开放的
main-red
议题。如果main分支故障持续三天，应在同一个议题中累积评论——而非创建12个议题。

创建前：

使用

mcp__github__list_issues

，参数为

labels: main-red, state: open

。

如果存在开放议题，添加评论：
运行#<run-id>同样失败。工作流
```
<name>
```
，分类
```
<bucket>
```
，可疑提交<sha-short> (#<pr-or-none>, @<author>)。<故障简要说明>。
仅当没有开放议题时才创建新议题。标题：
```
main is red: <workflow> (<bucket>)
```
。标签：
```
main-red
```
，若适用则添加
```
infra
```
或
```
needs-human
```
。

当main分支恢复正常（同一工作流的下一次成功运行）时，关闭该议题并在评论中注明恢复正常的运行ID。关闭操作需手动执行。

Issue body template

议题正文模板

markdown

**Workflow**: <workflow-name>
**Run**: <run-url>
**First failed step**: <step-name>
**Bucket**: <regression|flake|infra|needs-human>
**Suspect**: <sha-short> — <commit-subject> (PR #<n>, @<author>)

markdown

**Workflow**: <workflow-name>
**Run**: <run-url>
**First failed step**: <step-name>
**Bucket**: <regression|flake|infra|needs-human>
**Suspect**: <sha-short> — <commit-subject> (PR #<n>, @<author>)

Failure excerpt

Reproduction

<one of:> - Deterministic: `<exact command from the workflow yaml>` - Flake: passed on <prev-sha-short>; rerun button: <run-url>/attempts/2 - Infra: <service name> — <symptom> - Needs human: <why the logs are ambiguous>

Next action

<one line — "revert #N and reland", "rerun", "fix <specific thing>", "human eyes">


Keep the excerpt tight. Dumping the full log helps nobody.

<one line — "revert #N and reland", "rerun", "fix <specific thing>", "human eyes">


日志片段需精简。粘贴完整日志毫无帮助。

Reproducing locally

本地复现

A human invoking this skill via the consumer repo's

<repo>-pr-lifecycle

skill should reproduce before filing, using the commands in that skill's CI-triage section. For a

main

failure caught from outside a PR, check out

main

at the suspect SHA and run the same commands locally before filing.

For

e2e

failures specifically, delegate classification to

web-testing

's triage mode — it handles regression-vs-flake for browser-driven tests (the ambiguous case).

通过消费仓库

<repo>-pr-lifecycle

技能调用本技能的人员，应在创建议题前使用该技能CI处理部分的命令进行复现。对于在PR外发现的

main

分支故障，应检出可疑SHA对应的

main

分支，在创建议题前本地运行相同命令。

针对

e2e

故障，需将分类工作委托给

web-testing

的处理模式——它负责处理浏览器驱动测试的回归vs偶发故障（模糊场景）。

Log access

日志访问

WebFetch

cannot read authenticated GitHub Actions logs (403). The GitHub MCP gives you:

mcp__github__pull_request_read

with

method: get_check_runs

— step names, status, timings (no log body).

The workflow run's
```
jobs_url
```
via
```
mcp__github__get_commit
```
+
```
check_suite
```
traversal — same metadata.

Log bodies are not reliably accessible from the GitHub MCP. When the log body is unavailable, classify from step names + exit codes + the workflow yaml, and bias toward

needs-human

rather than guessing.

WebFetch

无法读取需要认证的GitHub Actions日志（403错误）。GitHub MCP提供以下能力：

```
mcp__github__pull_request_read
```
，参数
```
method: get_check_runs
```
——步骤名称、状态、耗时（无日志内容）。
通过
```
mcp__github__get_commit
```
+
```
check_suite
```
遍历获取工作流运行的
```
jobs_url
```
——仅元数据。

GitHub MCP无法可靠获取日志内容。当无法获取日志内容时，根据步骤名称、退出码和工作流yaml进行分类，优先归为

needs-human

而非猜测。

Flake-detection heuristic

偶发故障（flake）检测启发式

A failure is

flake

only if both hold:

The same workflow on the previous main commit passed (check via
```
mcp__github__list_commits
```
+ check runs on the prior sha).
The failing step's logs do not contain a symbol name, file path, or assertion message that appears in the suspect commit's diff.

One of those alone is not enough. A deterministic regression can pass on the prior commit; a real flake can mention a touched file by coincidence.

仅当同时满足以下两个条件时，故障才归为

flake

：

上一个main分支提交的相同工作流运行成功（通过
```
mcp__github__list_commits
```
+ 前一个sha的检查运行验证）。
失败步骤的日志中不包含可疑提交diff中出现的符号名、文件路径或断言消息。

仅满足其中一个条件不足以判定。确定性回归故障也可能在上一个提交时运行成功；真实的偶发故障也可能巧合提及被修改的文件。

Constraints

约束

Never open a PR from this skill. Triage is read-only on the codebase.
Never @-mention for
```
infra
```
or
```
flake
```
buckets — alert fatigue kills the signal.
Never close a
```
main-red
```
issue without a green run id to cite.
Scope: any GitHub-Actions–driven repo. The taxonomy and rolling
```
main-red
```
pattern are repo-agnostic; the consumer repo's CLAUDE.md can override triggers and label conventions.

切勿通过本技能创建PR。本技能仅对代码库进行只读处理。
切勿针对
```
infra
```
或
```
flake
```
分类@提及人员——告警疲劳会削弱信号。
切勿在未引用恢复正常的运行ID的情况下关闭
```
main-red
```
议题。
适用范围：任何GitHub Actions驱动的仓库。分类体系和滚动
```
main-red
```
模式与仓库无关；消费仓库的CLAUDE.md可覆盖触发条件和标签约定。

Relationship to other surfaces

与其他技能的关系

Surface	Role
`<repo>-pr-lifecycle` (consumer-repo skill, e.g. `onsager-pr-lifecycle` )	Interactive caller; humans use this when triaging a red PR check.
`web-testing`	Delegated to for `e2e` workflow classification.

Surface	角色
`<repo>-pr-lifecycle` （消费仓库技能，例如 `onsager-pr-lifecycle` ）	交互式调用者；人员处理PR故障检查时使用本技能。
`web-testing`	负责 `e2e` 工作流的分类委托。
",

ci-triage

Original

Translation

ci-triage

ci-triage

Taxonomy

分类体系

Suspect commit identification

可疑提交识别

The rolling
`main-red`
issue

滚动
`main-red`
议题

Issue body template

议题正文模板

Failure excerpt

Failure excerpt

Reproduction

Reproduction

Next action

Next action

Reproducing locally

本地复现

Log access

日志访问

Flake-detection heuristic

偶发故障（flake）检测启发式

Constraints

约束

Relationship to other surfaces

与其他技能的关系

ci-triage

Original

Translation

ci-triage

ci-triage

Taxonomy

分类体系

Suspect commit identification

可疑提交识别

The rolling main-red issue

滚动main-red议题

Issue body template

议题正文模板

Failure excerpt

Failure excerpt

Reproduction

Reproduction

Next action

Next action

Reproducing locally

本地复现

Log access

日志访问

Flake-detection heuristic

偶发故障（flake）检测启发式

Constraints

约束

Relationship to other surfaces

与其他技能的关系

The rolling
`main-red`
issue

滚动
`main-red`
议题