Address reader questions

Plan for /address-question: new comments in, one review page out, shown above the comments for Milton to check.

What the skill does (dashed bracket), what it reads, what it leaves for review.

Steps

Collect Download, keep unanswered questions.

sync-comments.py --download-only  →  comments.json
keep:  not in reviews.json / addressed.json, visible, a mathematical question
list:  edit requests ("remove this line") as edits for Milton
read:  the whole page at each anchor + every page, PDF, run it cites

Frame One self-contained problem per question.

setting     every object defined, copied from the page, with file paths
question    one quantified statement   ("prove or refute the step, and what it uses")
criterion   what a complete answer proves · what does not count
checklist   how a wrong answer fails
origin      the comment, quoted verbatim   (same question twice → one problem)

Solve /solve-maths with two agents.

round 1   A1 solves            (all problems in parallel)
round 2   A2 audits A1, fresh  (no opinion from the root)
record    briefs/ returns/ tree.json tree.md README.md verdict.md
status    "proved (in-run audit)"; /solve-maths-audit on request

Present One review page, /proof-breakdown per answer.

per question:  comment (name · time · passage link)
               Answer box: one-line verdict + proposition
               proof plan → phases → steps (≤ 5-word titles; Proof, Sources closed)
               proposed edit to the page · link to the run folder

Post Add the review row above the comments.

reviews.json  {date, title, href, status: "ready for review",
               answers: [{comment, anchor, short, oneline}]}
dashboard     newest review = highlighted row; "AI answer" chip per comment
after review  Milton edits pages → addressed.json; status "reviewed"

Reader preferences and update email

The user-preferences report collects the evidence from each reader’s visible comments. It describes requested explanation styles and clarifications, with tentative longer-term suggestions.

Each completed review is automatically published and emailed to Toan, unless a draft is requested. The workflow verifies the deployment and live page, sends the review link and answer summaries through Gmail, and saves a private receipt to prevent duplicate sends. Delivery runs at the end of the skill; mathematical page edits still await Milton’s approval.

Where things live

Table 1. what and where and public?
what	where	public?
downloaded comments	`~/collaborative-generation/local/comments/` (daily, 07:00)	no
solve runs	`solve-maths/reader-questions/<slug>/`	after push
review page	`collaboration/reviews/<date>.html`	after push
review index	`collaboration/reviews.json`	after push
the skill	`.agents/skills/address-question/` = `.claude/skills/address-question/`	repo

First run: 20 October 2026, Toan's four questions → review.

Online approval

In the dashboard, use Approve beside Open, or approve directly on a review page. Older reviews remain visible. No key is required: anyone visiting the site can save an approval or revision request.

Awaiting approval → Approved · awaiting application → Applying → Published.

The server saves a receipt for the exact review and original page versions. The dashboard and the workspace’s private queue read the same record; a changed target requires fresh approval. The daily download and each review/publishing session check this queue.

Approval does not immediately edit a mathematical page. At the next workspace publishing run, the approved edits are applied and checked; a successful live deployment produces the Published status and commit link. Request revision and application failures remain visible.