docs(checklist-test): a QA security finding gets no card and goes privately to the PM session, runners only test, and runner dispatches state effort by card shape

.claude/agents/os-dev.md

@@ -1,19 +1,19 @@
 ---
 name: os-dev
 description: >
-  Developer agent for exactly ONE GitHub issue, dispatched by the /pm-dispatch
+  Developer agent for exactly ONE GitHub issue (or one no-card order), dispatched by the /pm-dispatch
   PM loop. Implements the issue end-to-end in a dedicated worktree — branch,
   code, tests, changeset, push, draft PR — and returns a structured JSON
-  report to the PM. Use only with a single fully-specified issue as input;
+  report to the PM. Use only with a single fully-specified issue or order as input;
   never for open-ended or multi-issue work.
 model: opus
 ---
 
 <!-- `model: opus` 是刻意钉死的下限,不是上限 —— PM 的逐单定档优先。完整解析顺序、
 批量压死事故与两个陷阱(环境变量、允许名单)住 check:agent-model-declared 门禁脚本头部。 -->
 
-你是 ObjectStack 开发 agent,由 PM 派发,恰好带一张 GitHub issue。
-交付物两件:该 issue 实现完毕并推成 draft PR,加下方的 JSON 报告。
+你是 ObjectStack 开发 agent,由 PM 派发,恰好带一张 GitHub issue;无卡时派发令即卡并点名分支。
+交付物两件:该 issue 实现完毕并推成 draft PR,加下方的 JSON 报告;无卡则无 `Fixes #`,仅交终报。
 报告交付两次,GitHub 优先:同一段 JSON 先作 issue 评论,首行单独一行是纯文本 `os-dev-report`。
 ⛔ 不用 HTML 注释写标记。再作为终报消息:PM 机械解析它,终报消息就是 JSON 本身。
 仓库根的 AGENTS.md 有约束力,第一次编辑前先读它。
@@ -359,9 +359,9 @@ model: opus
 
 ```json
 {
-  "issue": <n>,
+  "issue": <n | null — null only on a no-card dispatch>,
   "status": "done | rework | blocked | needs_decision | reading",
-  "branch": "claude/issue-<n>-<slug>",
+  "branch": "claude/issue-<n>-<slug> | the branch a no-card dispatch order names",
   "pr": "<url or null>",
   "session": "session_<id> — this run's harness-stamped id (transcript Claude-Session: line; subagent = parent's)",
   "premise_still_valid": true,

.claude/skills/checklist-test/SKILL.md

@@ -102,8 +102,16 @@ Tier 1 批量清掉,剩下的预算才对得起 Tier 2。
   `node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --tier <paths>` 的输出、
   ⛔ 不凭记忆(floor sonnet · default opus · ceiling fable);每个给:该项 JSON、RUNNER.md、dogfood 技
   能、自己的端口/DB、结果不进仓规则(§4)。
+- **effort 按卡形状定,派发写明,⛔ 不留空。** runner 卡(Tier 1 pin 批、手工 API/CLI 与浏览器
+  组、P0 冒烟、UX 走查)一律 `high`;`xhigh` 只给修安全发现的 `os-dev` 与争议判定的对抗复核。
+- **按后端传档。** `Agent` 子代理显式传 `effort`(本条即传参的依据)。云卡(`create_session`)不收
+  effort,卡的会话记录也不带 `effort_level`,跑在哪档读不回(实测):派发记录写「未指定」并附
+  发起会话 `get_session` 读到的档位;它高于卡形状时照 pm-dispatch `dispatch-runbook.md`
+  〈云卡五课〉五之 ② 行办,或改开项目线程、开时给档位(实测按所给,不随父会话)。
+- **花费行。** 波次锚卡的派发记录与每个 runner 终报带各卡报告时的花费(`get_session` 的
+  `external_metadata.usage.cost_usd`),供下一波校准。
 - **QA 运行不是车道卡。** runner(子代理或云卡)由发起本轮的会话直接派,⛔ 不走车道认领
-  与座位贴;pm-dispatch 只管写入门(§4)与安全修复派卡(Guardrails)。
+  与座位贴;pm-dispatch 管写入门(§4)、安全修复派发与 PR 持有(Guardrails)。
 - **没有子代理工具时,顺序跑 —— 并在运行记录里声明这一轮是顺序执行的。** 规则与它
   的论证住在 `checklist-author` 技能(维护者所定,一处成文);⛔ 不留第二份拷贝。
 - 忠实执行每项的 `steps`,按各自声明的 `oracle` 判定每条 `acceptance` 与 `negative`
@@ -126,10 +134,11 @@ expected-fail 探针 —— 永不在真实缺陷上把条款打成绿。
 **issue 是纯文本 —— 永不放图。** 截图只为让你和子代理**现场**得出判定;它们是判断辅
 助,随运行环境一起丢弃。持久报告需要的是**复现规则,不是图片**。
 
-GitHub 写入只走 pm-dispatch 写入门:仓根直调 `scripts/pm/issue-create.mjs`、`post-stamped.mjs`、
-`close-cards.mjs`,批量先 `write-pace.mjs --announce-batch N`,云容器内自走 fleet relay;⛔ 永不 MCP
-`issue_write`(锁 1 已拒)或裸 `curl`。写入被会话权限拦下 ⛔ 不绕行:待发文本落文件、终报给
-出路径与正文,放行后照序补发 —— 写入未落地即未收口。记录这样写:
+一轮的每次 GitHub 写入与 issue / PR 处理(抽取、查重、收口)只走 pm-dispatch 写入门:仓根直调
+`scripts/pm/issue-create.mjs`、`post-stamped.mjs`、`close-cards.mjs`,批量先 `write-pace.mjs --announce-batch N`,云
+容器自走 fleet relay;⛔ 永不 MCP `issue_write`(锁 1 已拒)或裸 `curl`。写入门被本会话权限拦下
+⛔ 不绕行:待发文本落盘,路径与正文随终报或会话间消息交派发会话,由它照序经原写入门
+补发;其中 ⛔ 无安全复现 —— 写入未落地即未收口。记录这样写:
 
 - **标题** —— `QA run · <selector> (<已判定>/<总数>) · <sha8> · <date> · <counts>`:唯一
   形状、严格解析、分数**强制**;语法与退役写法见 RUNNER.md「The record title」。
@@ -149,7 +158,7 @@ GitHub 写入只走 pm-dispatch 写入门:仓根直调 `scripts/pm/issue-create.
   取为独立 issue:标题自含、复现与机制条目化、指回本 run issue 取全量证据;缺陷卡不
   挂 `qa-run`,正常进分诊首触。清单准确性发现与 fixture 缺口归波次锚卡(sweep 跟踪
   issue)收口,⛔ 不抽取;环境阻塞记录在案即可。逐条 `fail` 的四分法处置、收口评论的
-  机器行与抽取卡的溯源行见 §5。⛔ 认证 / 授权漏洞的抽取卡不写复现,见 Guardrails。
+  机器行与抽取卡的溯源行见 §5。⛔ 安全发现不抽取、不立任何卡,见 Guardrails。
 - **完成判据** —— 一个区只有在**每一项都有判定**时才算跑完。跑不完的一轮照常立
   单:交出已有的判定,**外加一份点名未触达项的交接清单**,标题按上面的分数如实写。
   `not-run` 是诚实判定 —— 判 1 项、11 项没跑就写 `(1/12) · … · 11 NOT-RUN`。
@@ -175,9 +184,9 @@ GitHub 写入只走 pm-dispatch 写入门:仓根直调 `scripts/pm/issue-create.
 | **过期规格** | 两边都按各自的声明行事,条款钉的是已被裁决改掉的旧形状 | 立 re-point 卡:把条款重新指向现行规格,引裁决 |
 | **配对无效** | 该项与这个 subject / fixture 根本不成立(错配的 revision、不适用的环境) | 记 `blocked` 说明,⛔ 不立卡 |
 
-四选一是**穷尽的**:落不进任何一格的 `fail` 说明你还没判完 —— ⛔ 永不把它留成一条无
-处
-置的红行。
+除安全发现外,四选一是**穷尽的**:落不进任何一格的 `fail` 说明你还没判完 —— ⛔ 永不把
+它留成一条无处置的红行。安全发现那一行本身即处置:⛔ 不立卡、不写 `Extracted:` 行,修复
+由派发它的 PM 会话派(Guardrails)。
 
 与 §4「清单准确性发现归波次锚卡」的分界,是**有没有一条 `fail` 要清**:由 `fail` 判出
 的断言缺陷 / 过期规格各自立卡(那条红行得有人负责变绿),运行中顺带看到的清单准确性问
@@ -235,12 +244,16 @@ QA-source: #运行记录号 · area.item · 条款
   或记 `blocked(environment)`;证到一半的项是 `partial`,不是 `pass`。带证据的
   blocked 判定是一次成功运行;伪造的 pass 不是。
 - **每个产品问题同一轮经 §4 立 issue。** `fail` 条款与条款外探测同样,⛔ 不只留在运行记
-  录或聊天;认证 / 授权类按下一条。
-- **认证 / 授权漏洞的复现,永不发布到 GitHub 的任何地方** —— 不进运行 issue,不进跟踪
-  卡,不进评论。卡只载项 id、条款与 `detail withheld pending maintainer`,复现留在会话里;修复直
-  接经 pm-dispatch 派卡:席内 `mode:subagent` 派 `os-dev`,复现只进派发令,⛔ 不等维护者;runner 无
-  派发通道时由派它的会话接手。⚠ 这与 RUNNER 规则 2 是**同一条规则写在两处**,⛔ 两边
-  都不是对另一边的优先级主张;完整论证与它的维护者裁定在 RUNNER。
+  录或聊天;安全发现按下一条。
+- **安全发现不立卡,复现不上 GitHub**(安全发现 = 认证 / 授权漏洞,或复现即利用配方的发
+  现):⛔ 不立任何卡,锚卡、跟踪卡不提;复现经会话间消息私报给**派发它的 PM 会话**(无此
+  通道则留本会话,终报报 blocked),由它经 pm-dispatch 派修复:席内 `os-dev` 子代理、`xhigh`,复现只
+  进派发令,PR、提交与 changeset 只写加固;要定契约的交维护者。运行记录只留一行
+  `area.item · 条款 · security finding — detail withheld, reported to PM`,算已完成判定、标题计 `FAIL`。
+- **runner 只测试。** 记录、缺陷卡、评论与 issue / PR 的每步处理都走 §4 写入门;⛔ 不派
+  `os-dev`,⛔ 不驱动 PR:一轮产出的每个 PR 归派发的 PM 会话复核、欠时契约复审、翻 ready 与
+  落地。⚠ 这两条与 RUNNER 规则 2 是**同一条规则写在两处**,⛔ 两边都不是对另一边的优先
+  级主张;完整论证与它的维护者裁定在 RUNNER。
 - **不把 blocked 项当可运行的跑** —— 解析器隐藏它们正为此。
 - **一个选择器、一次运行、一张 issue;验证重测不立 issue,证据进 PR 正文(§5)。** release
   sweep 把 `since:vN` 与 `priority:P0` 作为分开的两轮跑 → 两张 issue,不要糊在一起。

docs/qa/platform-checklist/RUNNER.md

@@ -78,22 +78,38 @@ selector handed you has a verdict; planned ids are listed, not chased.
      issue. A `fail` with no reproduction rule in its issue is not a completed verdict.
      (The screenshot that convinced you is a live judgment aid, not report content —
      describe what it showed in one line; never attach it.)
-   - **Carve-out — authentication and authorization findings only.** ⛔ Never publish a
-     reproduction for an authentication or authorization hole **anywhere on GitHub**: not
-     the run issue, not a tracking or extracted card, not a comment. Relocating it is not a
-     mitigation — a tracking card is a public issue in a public repo just the same. Such a
-     `fail` **is a completed verdict** when it records the item id, the clause, and `detail
-     withheld pending maintainer`; the reproduction stays in the runner session. The fix is
-     **dispatched directly through `pm-dispatch`**, never waiting on the maintainer: a card
-     carrying only the item id, the clause and `detail withheld pending maintainer`, then an
-     in-seat `mode:subagent` `os-dev` dispatch (a disclosure-discipline card) whose prompt
-     alone carries the reproduction. A one-shot runner with no dispatch channel hands that to
-     the session that dispatched it. Existence published, recipe withheld, is a complete and
-     actionable report — getting a defect fixed never requires handing anyone a working
-     exploit. (Maintainer disclosure ruling, 2026-08-18, recorded on #9387; fix-dispatch
-     ruling, 2026-10-27, recorded on #25807; the `checklist-test` skill states the same
-     guardrail in the same terms — one rule written in both places, not a precedence claim
-     by either.)
+   - **Carve-out — security findings.** A security finding is an authentication or
+     authorization hole, or any finding whose reproduction is an exploit recipe. ⛔ It gets
+     **no GitHub card at all**: no withheld-detail card, no tracking or extracted card, and
+     no line about it on the wave's anchor. A card that withholds the recipe still announces,
+     in a public repository, that a named item carries a hole before its fix lands. ⛔ Its
+     reproduction appears nowhere on GitHub: not the run issue, not a card, not a comment,
+     not a PR. The run record carries one row,
+     `area.item · clause · security finding — detail withheld, reported to PM`; that row
+     **is a completed verdict**, and the title counts it as a `FAIL`. The runner reports the
+     reproduction **privately** to the session that dispatched it, in a session-to-session
+     message, and keeps it nowhere else; a runner with no such channel keeps it in its own
+     session and reports the hand-off as blocked, without the reproduction. **That
+     dispatching PM session** decides the fix and dispatches it through `pm-dispatch`: an
+     in-seat `os-dev` subagent at `xhigh` whose dispatch prompt alone carries the
+     reproduction, and whose PR, commit and changeset describe a hardening without the
+     exploit steps. A finding that needs a contract decision first goes to the maintainer
+     instead. Getting a defect fixed never requires handing anyone a working exploit.
+   - **The runner only tests.** Its record, its defect cards, its comments, and every issue
+     or PR handling step of the run (extraction, dedupe, close-out) go through the
+     `pm-dispatch` write doors (`scripts/pm/`). ⛔ It never spawns `os-dev` and never drives
+     a PR: the dispatching PM session owns every PR the run produces, through review,
+     contract review where it is owed, readiness and landing. When the runner's own
+     permission check refuses a write door, it does not route around it: it stages the
+     writes in a file and hands off to the dispatching session with the path **and the
+     staged text itself**, in its final report or by session-to-session message, because
+     the dispatching session cannot read a file in a cloud runner's container. That session
+     posts them through the same doors in the staged order. ⛔ The staged text never carries
+     a security reproduction; that travels only in the private report. (Maintainer
+     disclosure ruling, 2026-08-18, recorded on #9387; fix-dispatch ruling, 2026-10-27,
+     recorded on #25807; no-card, runner-only-tests and PM-dispatches-the-fix rulings,
+     2026-10-29, recorded on #27377; the `checklist-test` skill states the same guardrail in
+     the same terms — one rule written in both places, not a precedence claim by either.)
      **No other failure class is softened by this**: everything else owes its reproduction
      rule in full.
 3. **Classify blockers honestly.** Missing seed/persona/fixture → `blocked(fixture)`,
@@ -763,10 +779,11 @@ its README.
 reach a verdict — never report artifacts. What the report carries for a defect is the
 **reproduction rule**: ordered steps / API calls (method · path · body) / the
 ref-targeted selector path + expected-vs-actual from the oracle, enough to re-hit it on a
-fresh boot with no picture — **except under rule 2's authentication/authorization
-carve-out**, where the report carries the item, the clause and `detail withheld pending
-maintainer`, and nothing else. A clause whose oracle was a screenshot is recorded as a
-one-line text description of what it showed, not a link.
+fresh boot with no picture — **except under rule 2's security-finding carve-out**, where
+the report carries the one row
+`area.item · clause · security finding — detail withheld, reported to PM` and nothing else.
+A clause whose oracle was a screenshot is recorded as a one-line text description of what it
+showed, not a link.
 
 The in-environment JSON scratch (RUNNER shape, never committed):
 
@@ -860,12 +877,11 @@ closing out the report includes the extraction, owed by the runner:
 - **Product defects found during the run**: at close-out, extract each one into its own
   standalone issue — self-contained title, reproduction and mechanism itemized in the
   card, a pointer back to the run record for the full evidence chain. A defect card does
-  not carry `qa-run`; it enters triage first-touch normally. ⛔ An authentication or
-  authorization defect is extracted under rule 2's carve-out — item, clause, `detail
-  withheld pending maintainer`, no recipe — and its fix is dispatched at once (rule 2). The
-  extracted card is public too. Every product defect the run meets, a clause `fail` or an
-  off-clause probe finding alike, is extracted in the same run — ⛔ never left only in the
-  run record or the chat.
+  not carry `qa-run`; it enters triage first-touch normally. ⛔ A security finding is never
+  extracted: it gets no card at all, and its reproduction goes privately to the dispatching
+  session, which dispatches the fix (rule 2's carve-out). Every other product defect the
+  run meets, a clause `fail` or an off-clause probe finding alike, is extracted in the same
+  run — ⛔ never left only in the run record or the chat.
 - **Checklist-accuracy findings and fixture gaps** close out through the wave's anchor
   card (the sweep's tracking issue) — ⛔ not extracted.
 - **Environment blockers**: recorded in the run record is enough.

docs/qa/platform-checklist/areas/access-security.json

@@ -3045,7 +3045,7 @@
       "title": "OS_PLATFORM_OWNER_EMAIL platform-admin anchor: only a VERIFIED stored-email match confers PLATFORM_ADMIN — unset, unverified match, non-match and malformed lists each confer nothing, fail closed",
       "since": "v17",
       "status": "active",
-      "revision": 11,
+      "revision": 12,
       "priority": "P1",
       "surface": "api",
       "personas": [
@@ -3109,7 +3109,7 @@
         }
       ],
       "negative": [
-        "no runtime write reaches the anchor: OS_PLATFORM_OWNER_EMAIL is declared a NON-TABLE derivation input (admin-standing-surface.ts) — revocation is a configuration change plus a process roll, by design. Any in-product surface found mutating who is a platform administrator is an authorization finding: record it under RUNNER rule 2's carve-out (item, clause, detail withheld pending maintainer), never with a reproduction",
+        "no runtime write reaches the anchor: OS_PLATFORM_OWNER_EMAIL is declared a NON-TABLE derivation input (admin-standing-surface.ts) — revocation is a configuration change plus a process roll, by design. Any in-product surface found mutating who is a platform administrator is an authorization finding: record it under RUNNER rule 2's carve-out (no card; one run-record row `area.item · clause · security finding — detail withheld, reported to PM`), never with a reproduction",
         "the derivation reads the caller's OWN STORED sys_user row, never the session-seeded grants.email — the ⭐ P1 pin of the leg. That property is asserted by the unit pin (the '⭐ … reads the STORED row, never the seeded email' describe); do not attempt to re-prove it against a live server",
         "run every refused probe as the seeded admin too: a 200 proves each refusal keyed on the member's standing, not on a broken route (wrong-persona, both sides)",
         "out of scope here: the WALLED-posture boot refusal when the variable is unset (#11184) is plugin-auth's separate door — do not score it on this item"
@@ -3209,6 +3209,12 @@
           "date": "2026-10-27",
           "change": "automated.kind unit -> dogfood (#26571 slice 3; the slice-2 ACCEPT noted it): revision 10's ref already leads with a dogfood pin over the in-process HTTP stack, with the unit pins beneath. No ref, clause or status changed",
           "ref": "#26571"
+        },
+        {
+          "revision": 12,
+          "date": "2026-10-29",
+          "change": "RUNNER rule 2's security-finding carve-out re-worded (#27382; maintainer rulings of 2026-10-29, recorded on #27377): a security finding gets no GitHub card at all, the run record carries the one row `area.item · clause · security finding — detail withheld, reported to PM`, and the reproduction goes privately to the dispatching PM session. negative[0] quoted the retired 'detail withheld pending maintainer' row and now names the new one. The same PR re-words every item string that quoted the retired row. No clause, oracle or status changed",
+          "ref": "#27642"
         }
       ]
     },