triage: make Stage 1-pre closer tree total over closer shapes, with fixture tests

[yiliang114]

Summary

Stage 1-pre's closer query in .qwen/skills/triage/references/pr-workflow.md fetches three closer shapes but the decision tree consumes only two:

1. Commit closers are dropped. The query explicitly fetches ... on Commit { oid } (GitHub sets closer to a Commit when an issue is auto-closed by a commit pushed to the default branch), but the jq filter's select(. != null and .number != null) discards it. A commit-closed issue is then indistinguishable from a manual close and escalates forever — even though the fix is on main by construction and the deterministic subsumption comparison could have run.
2. Unmerged-PR closers have no branch. The filter can emit "N false" (an unmerged PR closed the issue), which no bullet consumes.
3. No executable test. The filter is one of only two executable fragments in triage that decide a gh pr close, yet it ships untested, while scripts/tests/qwen-triage-workflow.test.js already has an established fixture-execution pattern for exactly this content class (the re-run-classifier test runs embedded jq under real jq against fixtures).

Proposed scope

• Make the tree total over the filter's output space: emit commit closers (e.g. "C \(.oid)") and route "closed by a commit on the default branch" into the same subsumption comparison; add an explicit merged == false branch (ambiguity → escalate).
• Add fixture-driven tests: extract the closer --jq filter from pr-workflow.md, run it against recorded GraphQL payloads, and pin merged → "7777 true", commit-only → the commit route, unmerged → the escalate route, so removing any route goes red.

Source

Review threads PRRT_kwDOPB-92c6c3VLn and PRRT_kwDOPB-92c6c3VLg on #10292 (Stage 1-pre spec review). Deferred from that PR to keep its diff bounded.

[qwen-code-ci-bot]

✅ Qwen Triage finished — view run. See the stage comments in this thread for the result.

✅ Qwen Triage 已完成 —— 查看运行。结果见本线程中的各阶段评论。

[qwen-code-ci-bot]

Triage

• Type: feature
• Labels: type/enhancement, category/development, scope/ci-cd, priority/P2
• Next: Implement the total closer tree + fixture tests as a follow-up to feat(triage): add duplicate / already-fixed gate (Stage 1-pre) #10292 (Stage 1-pre lives only on that open PR's branch, not on main)

Analysis (Stage 2)

Direction check: aligned — accept for exploration.

All three claims verified statically against #10292's head (57ee13a; Stage 1-pre is not on main yet):

• Commit closers dropped — confirmed. The GraphQL query fetches ... on Commit { oid }, but the --jq filter ends in select(. != null and .number != null); a Commit closer has no .number, so it is discarded and falls into "closer unresolved → escalate". An issue auto-closed by a commit on the default branch would escalate forever, even though the fix is on main by construction.
• "N false" unhandled — confirmed. An unmerged-PR closer passes the filter and emits "<number> false", but the decision tree only has bullets for "closed by a merged PR" and "closed manually / unresolved".
• Untested — confirmed. scripts/tests/qwen-triage-workflow.test.js (~6.8k lines at that head) never references the closer filter, while it already runs embedded jq under real jq against fixtures elsewhere (re-run classifier, it.skipIf(spawnSync('jq', …)) pattern). The subsumption close is the only gh pr close in the triage skill at that head, so this untested filter gates the one destructive action triage can take.

The proposed scope is minimal and matches the established fixture pattern. Ordering: land with or after #10292.

中文说明

• 类型: feature
• 标签: type/enhancement、category/development、scope/ci-cd、priority/P2
• 下一步: 作为 feat(triage): add duplicate / already-fixed gate (Stage 1-pre) #10292 的后续实现完备的 closer 决策树与 fixture 测试（Stage 1-pre 仅存在于该未合并 PR 的分支上，尚未进入 main）

方向核对：与方向一致——接受探索。

对 #10292 head（57ee13a）静态逐条核实：

• 提交类 closer 被丢弃 —— 属实。 GraphQL 查询取了 ... on Commit { oid }，但 --jq 过滤器以 select(. != null and .number != null) 结尾；Commit 类 closer 没有 .number，会被过滤掉并落入「closer 无法解析 → 升级人工」。默认分支上的提交自动关闭的 issue 会永远被升级，尽管修复按构造已在 main 上。
• "N false" 无分支 —— 属实。 未合并的 PR 关闭 issue 时过滤器会输出 "<number> false"，而决策树只有「由已合并 PR 关闭」与「手动关闭/无法解析」两个分支。
• 无测试 —— 属实。 该 head 上的 scripts/tests/qwen-triage-workflow.test.js（约 6.8k 行）未引用 closer 过滤器，而同文件已有在真实 jq 下用 fixture 运行内嵌 jq 的成熟模式（re-run 分类器，it.skipIf(spawnSync('jq', …))）。在该 head 的 triage skill 中，subsumption 关闭是唯一一处 gh pr close，即这个无测试的过滤器把守着 triage 唯一可执行的破坏性动作。

建议范围最小且与既有 fixture 模式一致。落地顺序：随 #10292 或在其合并后。

--- Qwen Code