Repository navigation
The guard for "no drafts in search" reads a field its own search never returns — and drafts are returned #2270
Description
一句话
tests/test_mcp_server.py::test_no_drafts_in_search 声称检查"搜索结果里没有 draft 课程"。它读的是 r.get("status")，而 status 在默认档（compact）与 summary 档里根本不存在——只有 detail=full 才带。所以这个断言无论如何都是 0，永远无法失败。而 draft 课程确实出现在搜索结果里。
实测（2026-09-25，本机，真实 worker + 真实语料）
query: "search quota exhausted false zero results"
detail=compact status key: ABSENT keys: id,title,problem,freshness,evidence_level,kind
detail=summary status key: ABSENT keys: id,title,problem,freshness,evidence_level,domain,tags,fix,kind
detail=full status key: draft keys: id,title,domain,tags,path,status,description,evidence_level,score,kind
语料里那篇课的状态：status = draft（命中结果就是它：search-quota-exhaustion-false-zero）
被断言的代码（tests/test_mcp_server.py）：
resp = rpc("tools/call", {"name": "misakanet_search",
"arguments": {"query": "draft test lesson", "top": 10}}) # 默认 detail=compact
results = result.get("results", [])
draft_count = sum(1 for r in results if r.get("status") == "draft")
check("no drafts in results", draft_count == 0, ...)r.get("status") 在 compact 档恒为 None → draft_count 恒为 0 → 这条 check 从来没有能力失败。它旁边的 check() 只是把布尔记进 PASS/FAIL 计数器，所以从 CI 的角度看它一直"通过"。
两个独立的问题，别混为一谈
(A) 一个不能失败的门禁。 这是本仓反复出现的那一类（#1822 的 --kv-only 提前 return、#2183 吞掉的退出码、#2525 之前的 identity aura 死读）。修法：断言真实属性——要么把结果里的 status 取出来（在 detail=full 下查），要么直接断言"结果里不存在 draft 课程 id"（不依赖投影暴露了什么字段）。
(B) draft 到底该不该出现在检索里——这是产品决定，不是 bug。 现状是：
- 24 篇
status=draft的课程在索引里，因此可被检索到（上面就是证据）； - 它们只在
detail=full才带status，所以吃 compact 的 agent 无法知道自己拿到的是 draft； - 而这 24 篇的证据等级：23 篇 E0、1 篇 E1；其中 17 篇的
## Verification是grep/echo/wc占位符而不是真的修复校验（scripts/lesson_gate.py --existing的警告，24 篇 0 阻塞项、46 条劝告）。
也就是说，本仓的信任语义是"标出证据等级而不是藏起内容"（compact 里就带 evidence_level），但唯一能说明"这是未定稿"的那个字段恰好不在默认档。这与今天刚修的 B25（evidence_level 只在 full 有、compact 没有）是同一个形状。
建议（两步，互不依赖）
-
无争议的一半：把
status（或仅draft: true）在 compact 档按需暴露——只对 draft 课程加键，非 draft 的响应逐字节不变（与summary_plain的加法式约定一致，那套约定由workers/d1-lesson-service.test.mjs的 "keeps the exact legacy shape" 逐字节钉住）。这样 agent 至少能知道。 -
需要 owner 定的那一半：draft 是否应被检索。三种取法——
(a) 保持可检索 + 标出 draft（= 上面第 1 步）；
(b) 完全排除 draft（那么 24 篇就是死重，要决定发布或删除）；
(c) 排除E0 且 Verification 为占位符的子集（17 篇 fanuc 系），其余保留。注意 (b) 会降低召回：
search-quota-exhaustion-false-zero是 C-07 的动机课程，里面写着"不要再执行旧版本的rm misakanet/.quota.json"，现在它是 draft——它正是最需要被搜到的那一类更正。
复跑方式
# 命中一片 draft 课程（默认档）
# misakanet_search(query="search quota exhausted false zero results")
node --test workers/d1-lesson-service.test.mjs # 逐字节的旧形状约定
python3 tests/test_mcp_server.py # 那条 check 所在的 smoke 脚本
python3 scripts/lesson_gate.py --existing --json $(python3 - <<'PY'
import json;print(" ".join(e["url"] for e in json.load(open("data/lessons.json")) if e.get("status")=="draft"))
PY
) # 24 篇 draft 的 0 阻塞 / 46 劝告发现路径（留个记录，因为它是"分类积压"顺出来的）：在做开放 issue 的机械五分类时，为了用生产检索判定"真盲区 vs 伪盲区"，逐条驱动真实 worker 检索 issue 的技术签名，撞见 draft 课程被返回，进而发现那条 check 读了一个不存在的字段。
This repo is using Opire - what does it mean? 👇
💵 Everyone can add rewards for this issue commenting
/reward 100 (replace 100 with the amount).🕵️♂️ If someone starts working on this issue to earn the rewards, they can comment
/try to let everyone know!🙌 And when they open the PR, they can comment
/claim #2270 either in the PR description or in a PR's comment.🪙 Also, everyone can tip any user commenting
/tip 20 @Ikalus1988 (replace 20 with the amount, and @Ikalus1988 with the user to tip).📖 If you want to learn more, check out our documentation.
Activity
fix(test): the guard for the search scope read a field that level nev…
zsxh1990 commentedon Sep 28, 2026
!claim
fix(test): remove vacuous draft-status check from compact search (#2270)
Ikalus1988 commentedon Sep 28, 2026
Live measurement, 2026-09-29 (this issue is closed; the additive half did not land)
Closing this as completed is right for its first finding — the vacuous guard is gone (#2306, #2378).
Its step 1 ("暴露
status… 只对 draft 课程加键，非 draft 的响应逐字节不变") is not in the code, and thegap is live on production today. Measured minutes ago,
detailbydetail, for the same draft lesson:So the finding in the original report stands unchanged: a draft lesson is the top hit for its own title
(
fanuc-alarm-code-referencefirst), it is astatus: draftfile inlessons/contrib/, and both defaultdetail levels — plus
misakanet_get_lesson— are silent about it.misakanet_get_lessonis the secondpath, and it was not in the original report: its payload has no
statusat all, for any detail.The corpus is the other half of the measurement: production indexes 426 rows (
GET /api/search-index→
docCount 426, andcanonical_lessons(lessons/)is exactly 426) while the repository's managed count is418 published lessons. Those 8 rows are the difference between "what the repo publishes" and "what
search returns", which is also why the retrieval bench scores lower against production than against
data/lessons.json— extra documents change IDF and displace expected lessons (measured inscripts/bench_production_recall.py).What this comment is not
It is not a request to reopen the product question. Whether drafts should be searchable at all — the
issue's own (a)/(b)/(c) — remains an owner decision, and the note in the original report is worth keeping in
front of whoever decides it:
search-quota-exhaustion-false-zerois a correction that is currently adraft, i.e. exactly the kind of lesson that needs to be findable.
The additive half needs no such decision, so I have opened a PR for it: it appends
statusto a resultonly when the lesson is not published, so every published response stays byte-identical (the projection
contract
workers/d1-lesson-service.test.mjspins), and it coversmisakanet_get_lessontoo.fix(mcp): mark a non-published lesson in the default detail levels an…
test(search): one CJK floor row's only answer is a draft lesson, and …
fix(site): a non-published lesson page must say so, like the API alre…