docs(pm-dispatch): governing-text step and protocol-baseline rule in the decision frame

[os-steve]

Fixes #15929

Session: session_019RfFHiRCSs3JXLK4cwcfox · branch claude/issue-15929-governing-text-step · worktree objectstack-issue-15929.

What changed

Two rules enter the decision-card frame, each written once per carrier in that carrier's register — rules only; the two misses and the case stay on the card, and the maintainer's verbatim quote stays there too.

Rule A — Governing text. A decision card (落卡) and a director presentation (呈报) carry a Governing text: item: the repo AGENTS.md section / ADR / spec docblock / lint rule that already constrains the option space, quoted with the grep that found it, or none found with the searches run. A card whose options contradict that text is not a decision card. Leaving the decision box cites the item alongside the ruling id (the 鲜度门 comment re-read stays — this adds the missing carrier search, it does not replace anything).

Rule B — the protocol is the baseline. When a card measures "the spec declares X, the code does Y", the default direction is code → spec; a protocol change is its own card, never a side option inside a defect card. The presenter's first two lines on such a card are what the protocol declares (quoted, with the file) and whether the card asks to change it — yes ⇒ re-file as a protocol card; no ⇒ the option space is how the code conforms.

carrier	register	what it now says
SKILL.md 〈升级与决策〉	the rule, one line each	落卡与呈报必带 Governing text: (carriers enumerated; conflicting option space ⇒ not a decision card) · 协议为基准 (default code → spec; protocol change is its own card, ⛔ not a defect-card option) · 出决策箱 now cites the ruling id and the Governing text: item
references/decision-analysis.md 落卡分析模板	the template skeleton	段落骨架 gains Governing text(引文 + grep,或 none found + 检索式) and 协议声明、是否改协议 as segments; the six body items are untouched
references/lanes/director.md 职责二	the presenter's duty	re-run the grep per card per repo before presenting (git grep -n -iE 'TERMS' origin/main -- AGENTS.md docs/adr packages/spec/src); on a spec-vs-code card the first two lines are the protocol's text (with file) and whether the card asks to change it — 改 ⇒ 重立协议卡,否则只议对齐

core-rules.md is not touched, as ruled. The four-axis frame block (SKILL.md 733–751) is byte-identical and still at the same line numbers; check:skill-frame-sync confirms 2 copies isomorphic. No model identifier appears in the commits, this body, or the files.

Merge with origin/main, and the proof

PR #15908 landed on main at 97afcb5cc while this flight was open, and so did the rules-only rewrite of the lane job descriptions; origin/main (a2051fa4e9) was merged into the branch as merge commit 3ba4752b31 (a merge, never a rebase of the pushed branch; no stash). SKILL.md auto-merged — main's removal of the old 四棱 must-carry line sits below this hunk. director.md conflicted on the rewrite and was resolved on main's rewritten content: the two lines were re-applied under 职责二, and the file's new 72-line ceiling is honoured (see the accounting). Proof on the pushed head:

git merge-tree --write-tree --name-only origin/main HEAD
3fc609b6cddfa5b7fbd810f820cb1ae3cb98fffe
exit=0

Exit 0, only the merged tree oid printed, no conflicting path named. (The pre-merge proof against PR #15908's own branch head 3b86d07a7b, taken on 2a149b3165, was also exit 0 with tree 3f8ea1a66d and no conflicting path — the later conflict came from the lane rewrite, not from that PR.)

Line accounting (ratchet, measured with LC_ALL=C awk)

• SKILL.md: 811/811 — line-neutral. Two new rule lines are paid by two compressions inside 升级与决策, each keeping both facts on one ≤120-byte line: 711+712 (the second escalation threshold + PM discretion) → one line, 117 bytes; 727+728 (③ 到手 + the four same-stroke items) → one line, 116 bytes. Measured first: 723+724 (102+117), 730+731 (118+69), 715–717 (80/96/102) cannot fold under 120 bytes without dropping a rule, so they were not chosen. The compression is the ratchet header's own sanctioned payment ("compressing in place"), not a re-wrap.
• decision-analysis.md: 50/50 — line-neutral. The eight-segment skeleton does not fit one 120-byte line in any spelling (measured 125–156 bytes), so it spans two lines (112 + 95 bytes); the pay-fold is 适用面 8+9 → one line (113 bytes), chosen because no pair among the 正文六项 (27–33: 108/108/82/103/83/93/88 bytes) fits 120 without dropping a clause — a stated deviation from the brief's suggested fold location, same line count.
• director.md: 72/72 — line-neutral on main's rewritten file. The brief's "+2 within 77" no longer exists: the lane rewrite took the file to 72 lines with a 72-line ceiling, so the two presenter lines (110 and 112 bytes) are paid by two in-place compressions: 61+62 (INCOMPLETE 照实报 + 清单并入决裁批次) → one line, 101 bytes, nothing dropped; 69+70 (the skills-seat / director-seat boundary) → one line, 110 bytes, dropping only the two grammatical fillers 的工作 / 的勤务. Stated as the brief asked.
• Widest line written: 119 bytes (the SKILL.md Governing-text line); every other written line is at most 117.

Gates run on 3ba4752b31 (every exit captured before any pipe)

• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 16 commands from the merged tree (3 paths vs merge base a2051fa4e); all 16 run, all exit 0; --ran reconcile: ✓ dispatch-gates --ran: 16 derived famil(ies) accounted for — 16 run, 0 NOT-MEASURED. (The same 16 had been run on 2a149b3165 before the merge; one, check:doc-formula-expressions, first exited 3 PREREQUISITE NOT MET — formula/lint unbuilt — read as NOT MEASURED, built through the verify lock, then green: ✓ check:doc-formula-expressions self-test: 58 cases passed and the three sweeps clean.)
• check-skill-line-ratchet (the version now on main): --self-test 155 cases pass; live: SKILL.md 811 (ceiling 811), decision-analysis.md 50 (ceiling 50), director.md 72 (ceiling 72); pin 342 unchanged; every line within 120 bytes.
• check-skill-id-lint: ✓ 24 file(s) clean. check-skill-frame-sync: ✓ 2 copies of the decision frame are structurally isomorphic across 2 files. check-governed-prose: ✓ 2 instruction surface(s) name all 5 registered governed surfaces … and claim no others.
• Whole-repo pnpm lint through bash scripts/pm/os-verify-lock.sh (slot issue-15929), on the merged tree: VERDICT command-exit 0 · held the lock 91s · waited 0s (shared-box seconds, not idle-box figures).
• Control-character scan over the three files: none.

Report-only, not done here

• The summon feat: Comprehensive CRM example demonstrating all ObjectStack protocol features #14 batch-ledger sampling suggested on the card was not performed; the PR is not gated on it.
• core-rules.md digests 升级与决策 (its 落卡/出决策箱 lines mirror SKILL.md 723–732); it now owes one line for the Governing-text item and the protocol baseline — reported for the skills seat, not edited, as ruled.

维护者速读(草稿)

改了什么:给决策卡和总监呈报加了两条规则,写进 PM 技能的三个文件:①每张决策卡先写明「已有的规则文本」(仓库规则、架构决策记录、协议注释、lint 规则里已经定过这件事的那一段),找不到就写明搜过哪里;②「协议为基准」——协议和代码不一致时默认改代码,想改协议要单独立卡,不能混在缺陷卡的选项里。不改任何产品功能和代码行为。

为什么改:同一天两张卡的推荐都和您已经定过的规矩相反,而那些规矩就写在仓库的规则文件和协议注释里——现有流程只在卡片评论里找旧裁决,所以两个席位都没看见,最后靠您凭记忆纠正。加上这一步,呈到您面前的卡不再和已定规则打架;协议问题也有了正确的入口(单独立卡改协议),而不是被当成缺陷卡里的一个选项绕过去。

风险与代价(含回滚):每张卡多一次 grep,几乎零成本。三个技能文件行数都不增:每处新增都靠把两条相邻规则压成一行付账,压缩后的措辞更紧,读者需要结合上下文。回滚 = revert 这两个提交(一个改动、一个合并 main),不涉及数据、接口或客户可见行为。

席位意见:

你要做的(一个动作):受管面 PR,请您人工合并;或否决。

---

Generated by Claude Code

[os-steve]

ACCEPT — lands #15929 as ruled (the maintainer's 插队), rules A and B in the three carriers. Governed (.claude/skills/**, predicate exit 3), so this PR stays draft at the governed terminal: both approvers requested; needs-user-decision goes on the PR with the final 「维护者速读」 posted beneath this verdict.

What the seat verified, at head 3ba4752b31, in its own compare worktree (merge base a2051fa4e9, origin/main at 46803fa69e)

• Three files, +12/−12, every ceiling held: SKILL.md 811/811 (pin 342), decision-analysis.md 50/50, lanes/director.md 72/72 — the seat's own ratchet run exits 0 with those three lines.
• Rule A in SKILL.md names the four carriers and the "conflicting option space is not a decision card" consequence; the 出决策箱 line now requires the Governing text: item beside the ruling id; the template skeleton carries the item's shape (quote + grep, or none found + the searches) and the protocol question; director.md carries the per-card, per-repo grep re-run before presenting. Rule B in SKILL.md (code → spec by default; a protocol change is its own card, ⛔ never a defect-card option) and its presenter form in director.md (first two lines: the protocol's text with file, whether the card asks to change it).
• The four folds were read for lost content: 711+712 keep both escalation facts and the PM-discretion clause; 727+728 keep ③ and all four same-stroke items; 适用面 8+9 keep both scope rules; director.md 61+62 keep both, 69+70 drop only two grammatical fillers. No rule lost.
• The four-axis frame block is byte-identical to origin/main (it sits at 734–752 on both after PR pm-dispatch: a governed PR at ACCEPT gets needs-user-decision + a Chinese 「维护者速读」 brief (dev draft, seat final) #15908's +1 shift above it; the seat's line-by-line diff against main differs only at the PR's own 出决策箱 line). check-skill-id-lint 24 files clean; check-governed-prose green; no issue number in the three files; widest written line 119 bytes.
• Merge state: the branch carries origin/main at a2051fa4e9 as a merge commit (PR pm-dispatch: a governed PR at ACCEPT gets needs-user-decision + a Chinese 「维护者速读」 brief (dev draft, seat final) #15908 and the lane rewrite included); the dev's merge-tree proof on the pushed head reads exit 0 with no conflicting path, and the seat's diff against origin/main shows no other SKILL.md movement since.
• CI, seat's read 16:2xZ: 28 check runs — 17 success, 9 skipped, 2 cancelled (concurrency duplicates of Part-of PR must not also close its card and No other open PR may claim the same single-writer path, each with a same-named success twin on this head), none failing, none pending.
• Body line 1 Fixes #15929; skip-changeset present; the body carries the 草稿 section.

Open questions answered: A and A. Keep the director.md folds — the lane rewrite moved that file's ceiling to 72 under this flight, and in-place compression with no rule lost is the ratchet's sanctioned payment. Accept the 适用面 fold for the skeleton — the measured 正文六项 pairs do not fit 120 bytes without dropping a clause.

Report-only items taken by the seat: references/core-rules.md owes one digest line for the two rules — filed as a lane card; the fresh-PR no-runs observation goes to the readings ledger only if reproduced.

Implemented-by: os-dev executor, flight #15929, branch claude/issue-15929-governing-text-step
Reviewed-by: pm-dispatch skills seat, https://claude.ai/code/session_019RfFHiRCSs3JXLK4cwcfox

---

Generated by Claude Code

[os-steve]

维护者速读

改了什么:决策卡与总监呈报的框架加两条规则,写进 PM 技能的三个文件(手册的「升级与决策」、决策分析模板、总监章程),三个文件行数都不变。①「Governing text」:每张决策卡、每次呈报必带一项「已经约束这个选项空间的规则文本」——仓库 AGENTS.md 章节、ADR、引用裁决的协议注释、lint 规则——引原文并附找到它的 grep,找不到就写 none found 并附搜过的检索式;选项与既有文本冲突的卡不是决策卡。②「协议为基准」:协议声明与代码不一致默认改代码对齐;要改协议必须单独立卡,不能混进缺陷卡的选项里;协议之争呈报前两行 = 协议原文(带文件)、卡是否要改协议。

为什么改:今天一批里两个推荐都和你早已定下的规矩相反——规矩就写在 hotcrm 的 AGENTS.md 和 spec 的注释里,可现有流程只在卡评论和总监台账里找旧裁决,两个席位都没看见,最后靠你凭记忆推翻。这是检索缺环,不是权衡出错:读到规则文本后两个推荐都自动翻转。第二条是你在 #15828 上的原话「本项目以协议为基准……协议有问题应该立卡修改协议」。

风险与代价(含回滚):每张卡多一次 grep,近零成本;不碰产品代码。三处新增靠把相邻两条规则各压成一行付账,席位逐句核过无规则丢失(总监章程里只删了两个虚词)。四轴框架块逐字节不变。flight 期间 main 上落了两个同文件 PR,dev 以合并提交并入并给出 merge-tree 无冲突证明;席位复跑棘轮、编号 lint、受管面判定全绿;CI 17 成功 / 9 跳过 / 2 个并发重复被取消(同名成功孪生在)。回滚 = revert 两个提交。

席位意见:建议批准。四轴:业务——两次误裁是实测的,规则文本三种载体也是实测的;长远——把「先读已有规则」变成框架的一步,而不是靠人记;防 AI 错——正是防 agent 在已裁问题上重新提案的那道门,协议为基准更把「放宽协议迁就代码」这类最危险的选项从缺陷卡里拿掉;创业阶段——零新增行、零新面。附带:核心条款摘要(core-rules.md)欠一行同步,席位另立小卡。

你要做的:批准并入队,或人工直合。一字:是/否。

---

Generated by Claude Code