infocard-publish-parallel-batch-pattern

name	infocard-publish-parallel-batch-pattern
description	Light-batch publish 2–3 cards in parallel without worktree.
version	2.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["infocard","publish","batch","parallel","light-route"],"related_skills":["infocard-publish-sop"]}}

Parallel Light-Batch Infocard Publishing

When to use

• 2 or 3 self-contained cards, each with a different theme (e.g. hardblue + pixelstack)
• Each card has its own fixed facts (no shared research handoff)
• Total API budget fits in one round (max_concurrent_children=3)
• All cards belong to one user-facing batch release

No worktree is used. All cards written directly to primary repository docs/.

Anti-pattern

Do NOT use this when:

• Cards share heavy research (use the full infocard-publish-sop batch with content.json + main renderer)
• Each card needs deep cross-source verification (use the full SOP)
• Batch size is ≥4 (split into rounds of ≤3, wait between rounds)

Batch cardinality and artifact ownership

The user's requested number of independent cards is a release invariant. If the user says "发布 7 张卡", author and verify exactly 7 card identities; never collapse them into one collection card. A later "go/继续" means execute without another confirmation.

All author outputs must land in docs/ before build. The orchestrator writes each HTML directly into the primary repository. A subagent claiming "created" is not evidence until the orchestrator verifies the exact HTML and meta paths, byte sizes, and matching slug/path pair.

Before build, run a cardinality preflight:

1. Expected slugs = actual HTML files = actual sidecars = index candidates.
2. Every sidecar slug and path must match the dated HTML filename exactly when the repository convention requires date-prefixed identities.
3. Do not trust an HTTP 200 alone. Parse the public _index.yaml and assert every expected slug is present, with the expected date, updated, path, and title.
4. If one card is missing or indexed under a different slug, stop the closeout, normalize the sidecar, rebuild _index.yaml and index.html, then push and re-run the full per-card public check.

Workflow

Step 1 — Fetch and confirm

cd ~/qbox/opendir/project/infocard-pub
git fetch origin main

Step 2 — Dispatch parallel Author subagents

Each delegate_task call:

• goal: produce a complete candidate for ONE card, content Y; do not include a preselected visual theme
• context: canonical URL, source content, and the frozen .docs/<run-id>/<slug>/theme-decision.json path
• explicit output rule: "read and consume theme-decision.json.selected_theme exactly; if it is missing or invalid, return THEME_BLOCKED. Write only under .docs/<run-id>/<slug>/; do not write docs/, assets/, indexes, Git state, commit, or push."

Each subagent must end by reporting:

• slug
• byte size
• completion status (COMPLETE / PARTIAL)

Step 3 — One build, one verify, one commit, one push

npm run build
git add docs/<slug1>.html docs/<slug1>.html.meta.yaml \
        docs/<slug2>.html docs/<slug2>.html.meta.yaml \
        _index.yaml index.html
git commit -m "feat: publish <batch-title>"
git push origin main

Step 4 — Public verification (all cards)

sleep 60
curl -sI https://ccwq.github.io/infocard-pub/docs/<slug1>.html | head -1
curl -sI https://ccwq.github.io/infocard-pub/docs/<slug2>.html | head -1

Both must return HTTP/2 200.

Pitfalls

• meta.yaml trailing ---: write_file overwrites the whole file. If the existing file ends with --- and your new snippet starts with ---, the on-disk YAML becomes two documents. Re-read first; write without trailing ---. Detection: grep -c "^---$" docs/*.meta.yaml returns 0.
• Section padding-left on mobile: every theme's section-head needs @media (max-width:720px) { padding-left: 14px; }. Otherwise the number block touches the left edge.
• Table → case-card-list: 5+ column tables need mobile collapse to .case-card-list.
• Overview block mandatory for hardblue: between topbar and stats row, must include .overview-sentence (red, ≤50 chars) + .overview-body (100–400 chars).