PR #1239 — The Book Gets Co-Authors: Part 0 + Part I, v0.3-DRAFT

The Broader Goal at the End of the Previous PR

"The Supervised Machine" (docs/book/) existed at v0.2-DRAFT as a single-voice engineering memoir written by Nikhil, with Claude/Codex/Grok/Agy appearing only as reviewers and marginal commentators. That draft closed out a phase of internal use — it was a project artifact, not yet a book positioned for an outside reader. The multi-author repositioning brainstorm (docs/superpowers/specs/2026-08-29-book-multiauthor-repositioning-design.md) reframed the project: broaden from an engineering memoir into a book about human-AI collaboration, and make the four harnesses genuine co-authors with discernible voices and real disagreements in the text, not just editorial sign-off.

Strategic Shifts in This PR

None beyond what the design doc already committed to. This PR is the first execution slice of that design — Part 0 and Part I only (4 chapters), with Parts II–IV explicitly deferred to follow-on plans once this slice is reviewed. The chapter-ownership map from the design's Section 4 was followed exactly: Grok led the Reckoning chapter (Part 0) and the "power tools have kickback" chapter; Claude led the "cognitive tool, not magic" chapter; Agy led the "context & working memory" chapter — each dispatched through synlynk dispatch <agent> --force-agent --context-mode full rather than authored inline, so the chapter drafting itself went through the same telemetry/cost capture path as any other dispatched work.

What This PR Shipped

• Four new chapters merged into docs/book/the-supervised-machine-v0.3-DRAFT.html (renamed from v0.2-DRAFT), each written by its assigned co-author under the .who.<agent> / .who.human dialogue-box CSS convention already established in the manuscript.
• Marginalia insertions gated by the design's hard rule: real disagreement or a cited incident only, never manufactured color. Every marginalia citation in this slice traces to a real GitHub issue or PR (#1166, #880, #426, #1176, #1177, #479, #1197, #1205) — verified by the reviewing job directly against gh issue view / gh pr view, not taken on faith.
• A rebuilt PDF (the-supervised-machine-v0.3-DRAFT.pdf, 968,851 bytes, 8 pages) via headless Chrome, and a docs/book/README.md bump from v0.2-DRAFT to v0.3-DRAFT with an added Editorial History note pointing back at the design doc.
• Structural verification before merge: html.parser clean parse, <div>/</div> counts matching at 217/217, <chapter>/</chapter> counts matching at 25/25, and a git diff -M origin/main...HEAD confirming zero deletions in the existing Parts One–Four — the new chapters extend the manuscript, they don't touch what was already there.

The PR itself surfaced one process wrinkle worth recording: this project's PR Review Discipline requires a non-authoring agent to review and merge, and Claude (the PM session) had authored PR #1239 directly. .synlynk/policy.json's review task-allocation routes to claude with agy as fallback — since Claude was disqualified as author, the fallback correctly took over. Dispatching that review hit two infrastructure snags before it ran clean: the worktree was missing its .synlynk/github_apps/ symlink (a known gap — see worktree-github-apps-gap.md), and once that was fixed, the qa role's GitHub App token had actually expired because synlynk daemon (which auto-refreshes it) wasn't running. Starting the daemon and re-running synlynk identity init --role qa cleared both. Once dispatched, Agy ran synlynk pr check, verified the structural checks above, posted a formal COMMENT review (the sanctioned fallback for shared-GitHub-identity constraints — see PR #417), cleared synlynk policy check-merge --role qa, and squash-merged. The job's own status label came back as the ambiguous succeeded_gh_write_failed despite having genuinely succeeded — resolved by cross-checking gh pr view --json state,mergedAt,mergeCommit directly, which matched the job's self-report exactly, including the merge commit SHA. Another data point for this project's standing rule: never trust a dispatch status label alone.

Brainstorm Visuals Used

None — this slice followed a text-only design/plan pair (docs/superpowers/specs/2026-08-29-book-multiauthor-repositioning-design.md and docs/superpowers/plans/2026-08-29-book-multiauthor-part0-part1-slice.md); no visual companion session was run for this brainstorm.

What This Achieved on the Path to Autonomy

This is the first manuscript content that ever went through the full dispatch → verify → non-authoring-review → merge pipeline that the rest of synlynk's engineering work already uses. It proves the same discipline — don't trust job status alone, verify structurally, route review away from the author, clean up dispatch worktrees immediately on merge — holds for prose authorship as well as code. Eight nested dispatch worktrees (four chapter-drafting jobs, the review job, and three earlier attempts/retries) and their branches were swept immediately after merge per the Worktree Hygiene Protocol, keeping this from becoming one more entry in the kind of stale-worktree backlog that prompted that protocol in the first place.

Strategic Note: The Goal at the End of This PR

The next goalpost is Parts II–IV of the repositioning: Agy's Part II chapter plus Codex's named contributions within it, and the Part III infra-evidence and meta chapters (Grok and Claude respectively, per the design's ownership table), each needing its own plan once this Part 0/I slice has had a chance to be read. The "How to Read This Book" front-matter segment map from the design's Open Items is still outstanding and belongs in that next plan rather than being retrofitted here.