Human Summary

Human-readable progress report

You are the main agent. This skill produces a human-facing math report — for the operator, or the mathematician who posed the problem. The audience is a mathematician fluent in standard English math terminology who knows nothing about how the work was produced.

You do not author the prose yourself and you do not read the fact files. The report is written by an isolated report-writer codex behind the human-summary MCP tool, which is fed ONLY the verbatim problem statement and a scrubbed, id-free bundle of the project's verified results. That isolation is the structural guarantee that no internal identifier (fact_id, author, predecessors, …) or system/orchestration vocabulary can reach the report — the author's window never contains any of it. Your job is to call the tool, then render and deliver the PDF.

Step 1 — generate the clean report.md (the tool does the writing)

Call the MCP tool (server human-summary):

summary_write(project="<project>")

It assembles the writer prompt + PROBLEM.md + a scrubbed fact bundle (statement / proof / intuition bodies only — all frontmatter stripped, no fact ids, no author names, no machinery), drives an isolated codex, writes the result to <project>/report/report.md, and runs a leak check on the output. It returns a small dict:

{report_md_path, status, returncode, leak_findings, stderr_tail}

Honesty gate — do not proceed unless status == "ok":

• status == "ok" means: codex exited 0, produced a non-empty report, AND the leak check found zero hits (leak_findings == []). Only then does report.md exist as a clean artifact.
• status != "ok" (error / timeout / leak): no clean report.md is written. On a leak, the offending output is quarantined at report.leaky.md and leak_findings names what leaked — report this to the operator, do NOT render or deliver it, and do NOT hand-fix and pass it off as clean. If codex failed, surface stderr_tail.

You never read the fact graph and never write the report prose; the tool owns both. If the operator asks for a different language/register, that is a property of the writer prompt (agents/skills/human-summary/REPORT_WRITER_PROMPT.md, operator-editable) — the register rule (narrative in the operator's language, all standard math terminology in English) and the five-section structure are locked there, not here.

Step 2 — render the PDF and deliver

Once you have a clean report.md, render it to a self-contained PDF:

bash "${CLAUDE_SKILL_DIR}/render_pdf.sh" <report.md> <out.pdf> "Title"

This server-renders markdown + KaTeX into self-contained HTML and prints it to PDF via headless Chrome — so the math + fonts are handled without any LaTeX engine. Deliver the PDF path to the operator — never paste raw $...$/\boxed{} into chat; it shows as tex garbage and is unreadable.

Step 3 — backstop self-check (documented, kept as defence-in-depth)

The tool's leak check is the primary guard, and the scrub makes a leak structurally impossible. As a belt-and-braces backstop before you deliver, you may still grep the rendered source:

grep -E '[0-9a-f]{16}' <report.md>    # must return nothing (no fact_id / hash prefix)

If this (or the tool's leak_findings) ever fires, treat the report as compromised: do not deliver it, and report the finding.