Parallel 6-agent report improvement rounds
Staging root: cf-pages-nocap-agent-dev/
Trigger: operator says “run a 6-agent report round” (or names skill nocap-parallel-round).
Default: no deploy, no Modal GPUs, workspace codes on public pages; no secrets.
Canonical surfaces: report.html, preview-schema.html, muon-explained.html, companion .md → HTML via build_html_docs.py, assets/site.css.
Leave cyan index.html alone unless the round goal says “promote schema home”.
Standing rules
- English on all public pages. “Dutch style” = metaphor for voice: direct, concise, no adverbs, no rhetoric — let the data speak. Not the Dutch language.
- Schema design: Inter + IBM Plex Mono, white paper, magenta
--ours(assets/site.css). - Claim hedges: not an official NoCap win; matched stock ΔT% N/A; numbers from
NUMBERS_CANON.md/ sealed summaries only. - - Never print secrets.
Six lanes
| Lane | Mission | Writes (exclusive) | Reads |
|---|---|---|---|
| L1 Animation | Motion that teaches; prefers-reduced-motion | reviews/ANIM_OPP_*.md, anim-review-*.html only | report / preview / muon (read) |
| L2 Design / clarity | Typography, IA, skim path, shared chrome | assets/site.css, assets/header.snippet.html, DESIGN_AUDIT.md | all HTML (read); CSS/snippets only |
| L3 Results communication | Scoreboard, compute tiers, verify path | reviews/RESULTS_COMM_*.md | GPU_HOURS_*.md, NUMBERS_CANON.md, seals |
| L4 Insight | Failure → bank → seal; non-claims | reviews/INSIGHT_*.md (drafts under reviews/ only) | timeline, failure ledger, inventory (codes) |
| L5 Literature | Papers / books/ → knowledge mapped to our claim | reviews/LIT_*.md, reviews/MUON_* extensions | muon-explained.html, public URLs, books/** |
| L6 Fact-check + merge | Accept/reject; apply to canonical pages; checklist | report.html, preview-schema.html, muon-explained.html, companion .md, FACT_CHECK_CHECKLIST.md, build_html_docs.py | reviews/ + seals |
Hard rule: L1–L5 never edit canonical HTML (except L2 CSS/snippets). Only L6 merges after scoring against NUMBERS_CANON.md.
Round protocol
- Open — create
reviews/rounds/ROUND_YYYYMMDD.mdwith a one-line goal. - Launch — six background agents in one parent message (see
reviews/PROMPT_TEMPLATES.mdor skill). - L1–L5 — proposals with required header (below).
- L6 — merge pass after L1–L5 (or read-only overlap, then write).
- Close — accept/reject table in the round log; preview URLs; deploy only if operator asks.
Parent agent = dispatcher + L6 waiter, not a seventh editor.
Proposal header (required)
## Proposal
- **Lane:** L1|L2|L3|L4|L5
- **Claim:** one sentence
- **Files touched:** paths
- **Risk to accuracy:** low|med|high — why
- **Demo path:** file or http://127.0.0.1:8766/...
Conflict matrix
| Conflict | Resolution |
|---|---|
Two agents need report.html | L3/L4 write reviews; L6 applies sequentially |
| Animation vs design | L1 → anim-review-*; L2 → tokens; L6 ports motion once |
| Literature invents a speed % | L6 rejects unless sealed SUMMARY / NUMBERS_CANON |
| Compute / billing | Only if round goal includes it; L3 drafts; L6 merges from GPU_HOURS_*.md only |
Success criteria
- Exactly six lane agents launched; L6 leaves a merge-ready tree (no deploy by default).
- Zero new Modal profile names on public HTML.
- Local preview coherent under
assets/site.css. FACT_CHECK_CHECKLIST.mdupdated for numeric/copy changes.- Round log lists rejects as well as accepts.
Preview
cd cf-pages-nocap-agent-dev && python3 -m http.server 8766
- http://127.0.0.1:8766/report.html
- http://127.0.0.1:8766/preview-schema.html
- http://127.0.0.1:8766/muon-explained.html
Deploy (operator only):
# promote allowlisted public files to cf-pages-nocap-agent/
# exclude WORKSPACE_MAP.local.md, EMAIL_SNIPPET.md, *.local.md
cd ../cf-pages-nocap-agent && wrangler pages deploy . --project-name=nocap-agent --commit-dirty=true
Next round (after a closed log)
- Read the latest closed
reviews/rounds/ROUND_*.md. - Prefer deferred accepts (e.g. chart teach-beat port) before inventing new themes.
- Round id:
ROUND_YYYYMMDD_rN. - Invoke skill
nocap-parallel-round— parent launches L1–L5 then L6; no seventh editor.
Related
- Prompts:
reviews/PROMPT_TEMPLATES.md(local-only; not hosted) - Round logs:
reviews/rounds/(local-only; not hosted) - Books drop zone:
books/README.md(local-only; not deployed) - Skill:
nocap-parallel-round(~/.cursor/skills/and repo.cursor/skills/)