[jasontopia]

Intent

继续 Learn FirstMate 教程，写第五章。前四章已写完（第一、二章已合入主分支，第三、四章正在走质检流水线）。README 的学习路线图里，第五章的定位是：为什么默认必须船长点头才合并，yolo 姿态放开了什么、没放开什么，以及红色 CI 为什么依然不能合。第五章也是「沿着一件活的真实旅程往下走」这条线（第 2 到 5 章）的最后一章。

What Changed

• 新增 chapters/05-pr-merge-authority.md：讲解为什么合并默认需要船长明确点头、yolo 姿态放开的是合并权限而非质检标准或红色 CI 拦截、以及销毁性/不可逆/安全相关合并为何仍需上报，内容依据 AGENTS.md 第 1 节硬规矩二与第 7 节「Selected delivery path and merge authority」小节。
• 章节结合本仓库自身的 PR 合并记录（已合并 vs. CI 绿但未合并的对比）说明合并需要 CI 绿与船长同意两个条件同时满足。
• 提供两个只读动手练习，分别推演不同 yolo/CI 组合下的合并判断，以及核对本仓库真实 PR 状态，并给出前后 git 状态对比的验证方法。
• 更新 README.md 路线图表格，将第 5 章状态从「待写」改为「已写」并加上章节链接。

🤖 Generated with Claude Code

Risk Assessment

✅ Low: All three round-1 findings (private-data sourcing, false chapter-3 merge claim, mislabeled AGENTS.md section) are correctly fixed: the chapter now sources its examples from learn-firstmate's own public PR/CI history (verified via gh-axi: PR #1 and #3 merged with green CI, PR #4 still open with green CI but no merge), which is public data distinct from the FirstMate checkout's captain-private paths; all AGENTS.md quotes and the section-7/subsection citation were independently verified against ~/firstmate/AGENTS.md and match exactly; the exercises remain read-only and consistent with the corrected claims.

Testing

Baseline python3 scripts/check-docs.py already passed (6 files, no problems) and I re-ran it to confirm. Since this change only edits tutorial prose in chapters/05-pr-merge-authority.md and a README table entry, there is no running application for this test phase to drive live; the tutorial's own exercises call for a live FirstMate/大副 session that this test phase has no credential or environment to start. As extra diligence I fact-checked the corrected claims against real GitHub state via gh-axi pr view 3 and gh-axi pr view 4: PR #3 (chapter 2) is confirmed merged and PR #4 (chapter 3) is confirmed still open with green CI, exactly matching the rewritten chapter text - so the fixes for the false merge claim and the private-data sourcing read as accurate on manual review, though this is a document fact-check rather than a live scenario. Overall: no-surface, nothing to park.

• Live validation: ⚠️ no-surface - 0 of 5 scenarios driven live against the product

Scenario	Result	Live	Evidence
reader-follows-readme-to-ch5	⏸️ untested	no	This repo has no runtime application - chapters are static Markdown read by a human. There is no product process to launch or drive; the only executable check is the docs linter, which already ran as…
exercise-1-yolo-merge-scenarios-with-real-firstmate	⏸️ untested	no	Exercising this requires a live FirstMate supervisor session (the 大副) that this test phase has no credential or environment to start, and the tutorial's own rule marks the FirstMate checkout read-only…
exercise-2-verify-pr2-and-pr3-status-with-real-firstmate	⏸️ untested	no	The exercise itself asks the reader's own FirstMate session to look up and narrate this state; no such live agent session is available to this test phase. The underlying facts were independently verif…
guard-no-captain-private-data-quoted	⏸️ untested	no	Confirming the absence of private-data quoting is a document-content check, not something with a live product surface to drive - there is no application behavior to exercise here, only prose to read.
guard-no-false-ch3-merged-claim	⏸️ untested	no	Same as above - this is a factual-accuracy read of static tutorial prose, not a live product interaction.

• Outcome: ⚠️ 1 warning across 1 run (1m22s)

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

🔧 **Review** - 3 issues found → auto-fixed ✅

• 🚨 chapters/05-pr-merge-authority.md:17 - Chapter 5 quotes and derives content directly from this checkout's captain-private state to build 'authoritative' tutorial material, violating this repo's own AGENTS.md rule: ".env, data/, state/, config/, projects/, and .no-mistakes/ in that checkout are captain-private. Never use their contents as tutorial material." Line 17 quotes the exact registry line from data/projects.md ('learn-firstmate [no-mistakes-prod-only] - ...') verbatim, presenting it as something the reader can look up. The 'two real examples' section built on top of that (lines 42-48) is also sourced from captain-private files: the force-push example (line 46) reproduces the captain's quoted authorization from data/backlog.md:49 and the harness-classifier explanation from data/learnings.md:11; the ch2/ch3 merge example (line 48) is likewise drawn from data/backlog.md. Exercise 2 (lines 83-102) then has the reader ask their own firstmate to recall these same private-sourced events. Remedy: rewrite these passages so the illustrative facts are stated without quoting or reproducing content lifted from data/, state/, config/, projects/, .no-mistakes/, or .env — the project's own writing rule requires public/tracked sources only (AGENTS.md, README.md, docs/, bin/ header comments).
• 🚨 chapters/05-pr-merge-authority.md:48 - Line 48 (and the exercise-2 prompt/expectations built on it at lines 87, 92-93) states that chapter 3's PR was already merged after the captain said '合并吧', citing it as a settled precedent for 'yolo 关着时最朴素的合并流程'. But per data/backlog.md, learn-fm-ch3 is still listed under 'In flight' (not 'Done'), its PR has no merged marker, and the live PR list shows PR docs(chapters): write chapter 3 on crewmates and worktree isolation #4 (chapter 3) still open with no review — matching the user intent's own framing that 'chapters 3 and 4 are still going through the quality pipeline'. The claim is fabricated/premature, not traceable to any authoritative source, and directly undercuts the chapter's own instruction to the reader not to invent unverifiable history ('不要编'). A reader running exercise 2 to verify against 'PR 链接或提交记录' will find this false. Fix by dropping the chapter-3 half of this example (keep only the chapter-2 precedent) until chapter 3 has actually merged.
• ℹ️ chapters/05-pr-merge-authority.md:9 - Lines 9 and 21 cite "AGENTS.md 第 7 节「Selected delivery path and merge authority」" as if that were section 7's title. In the actual AGENTS.md, section 7 is titled 'Task lifecycle'; 'Selected delivery path and merge authority' is a ### subsection nested inside it. A reader trying to locate '第 7 节' by that name in AGENTS.md would not find a section with that title. The quoted content itself is accurate — only the section label is off.

🔧 Fix applied.
✅ Re-checked - no issues remain.

⚠️ **Test** - 1 warning

• ⚠️ this change has no live-validatable surface; proceed without live validation? (0 of 5 scenarios were driven live against the product); reader-follows-readme-to-ch5: This repo has no runtime application - chapters are static Markdown read by a human. There is no product process to launch or drive; the only executable check is the docs linter, which already ran as the baseline command.; exercise-1-yolo-merge-scenarios-with-real-firstmate: Exercising this requires a live FirstMate supervisor session (the 大副) that this test phase has no credential or environment to start, and the tutorial's own rule marks the FirstMate checkout read-only / no state-changing commands - there is no tool or authority available here to stand up that session.; exercise-2-verify-pr2-and-pr3-status-with-real-firstmate: The exercise itself asks the reader's own FirstMate session to look up and narrate this state; no such live agent session is available to this test phase. The underlying facts were independently verified via gh-axi, which confirms the corrected claims are accurate, but that is a manual fact-check of the document, not a live run of the exercise as an end user would experience it.; guard-no-captain-private-data-quoted: Confirming the absence of private-data quoting is a document-content check, not something with a live product surface to drive - there is no application behavior to exercise here, only prose to read.; guard-no-false-ch3-merged-claim: Same as above - this is a factual-accuracy read of static tutorial prose, not a live product interaction.
• Live validation: ⚠️ no-surface - 0 of 5 scenarios driven live against the product

Scenario	Result	Live	Evidence
reader-follows-readme-to-ch5	⏸️ untested	no	This repo has no runtime application - chapters are static Markdown read by a human. There is no product process to launch or drive; the only executable check is the docs linter, which already ran as…
exercise-1-yolo-merge-scenarios-with-real-firstmate	⏸️ untested	no	Exercising this requires a live FirstMate supervisor session (the 大副) that this test phase has no credential or environment to start, and the tutorial's own rule marks the FirstMate checkout read-only…
exercise-2-verify-pr2-and-pr3-status-with-real-firstmate	⏸️ untested	no	The exercise itself asks the reader's own FirstMate session to look up and narrate this state; no such live agent session is available to this test phase. The underlying facts were independently verif…
guard-no-captain-private-data-quoted	⏸️ untested	no	Confirming the absence of private-data quoting is a document-content check, not something with a live product surface to drive - there is no application behavior to exercise here, only prose to read.
guard-no-false-ch3-merged-claim	⏸️ untested	no	Same as above - this is a factual-accuracy read of static tutorial prose, not a live product interaction.

• python3 scripts/check-docs.py
• python3 scripts/check-docs.py
• gh-axi pr view 3
• gh-axi pr view 4
• manual re-read of chapters/05-pr-merge-authority.md diff against the three round-1 fix decisions

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

[jasontopia]

Superseded by #14 (same content, re-validated on a fresh branch after a quality-check tool issue on this branch).