preflight

name	preflight
description	Run pre-publish checks on a blog post — frontmatter, content, feed-page exposure, and cover image checks per platform. Use before publishing a post under YYYY-MM-DD-slug/.

Preflight

Run pre-publish checks on a post. The skill is a thin orchestrator around a deterministic stage and a VLM verification stage. It owns no checking logic itself. Cover expectations are self-contained here: X article cover is 5:2 full-bleed with readable title + kicker + rhetoric; WeChat 头条 cover is 2.35:1 centered-visual with the rhetoric fully inside the center 1:1 square (the share card).

What it checks

1. Content and metadata (deterministic, .pi/bin/preflight): frontmatter fields, non-empty body, no TODO, local image refs resolve, cover files exist with the right aspect ratios (X 5:2, WeChat 头条 2.35:1).
2. Feed-page exposure (deterministic, same script): title length, subtitle presence, and an approximate 300-character excerpt sanity check that mirrors the gatsby feed/PostExcerpt query.
3. Cover checks (SAM 2 + VLM), per platform spec:
   • WeChat 头条 (cover-image-wmp.png): the center 1:1 square is the share card (third-party and in-app share both crop to it). SAM finds candidate squares; the VLM confirms the center square is a complete, self-contained visual (metaphor fully inside, uncropped, no text bleeding in). The center square is the deciding candidate.
   • X article (cover-image-x-article.png): full-bleed, no center-square rule. The VLM confirms it is a coherent full-bleed banner with readable title + subtitle + metaphor. (Run only when the user asks.)

Responsibilities

• Resolve the target post directory.
• Run .pi/bin/preflight <post-dir> and read the deterministic report. If it fails, present failures and stop.
• Run the cover checks above via the SAM script (.pi/scripts/cover-crop/cover-crop-safety-check.py) and the cover-crop-safety-verifier subagent.
• Present a single consolidated report.

Limits

MUST NOT:

• Duplicate the deterministic checks or the SAM logic here. Keep them in .pi/bin/preflight and .pi/scripts/cover-crop/cover-crop-safety-check.py.
• Run the VLM verification in the main agent. It belongs in the cover-crop-safety-verifier subagent for context hygiene.
• Skip the deterministic stage before the cover stage.
• Edit the post unless the user explicitly asks.

Workflow

1. Resolve the target post directory.

2. Run the deterministic stage:

   python3 .pi/bin/preflight <post-dir>
   

   Exit code 1 means a hard failure. Present failures and stop unless the user says to continue.

3. WeChat 头条 center-crop check. Run the SAM candidate script:

   .pi/scripts/cover-crop/.venv/bin/python .pi/scripts/cover-crop/cover-crop-safety-check.py <post-dir>/cover-image-wmp.png --json
   

   If the venv is missing, tell the user to run zsh .pi/scripts/cover-crop/setup.zsh and stop.

4. Spawn the cover-crop-safety-verifier subagent with the candidates. In the spawning prompt, state that the center square is the deciding candidate (third-party platforms crop to it) and must pass; left/right are optional in-app share exports. Ask for PASS center or NONE with one short clause naming what is cut or missing.

5. X article cover check (when the user asks). Spawn the cover-crop-safety-verifier subagent with cover-image-x-article.png and ask whether it is a coherent full-bleed banner with readable title + subtitle + metaphor. Do not apply the center-square rule to X.

6. Present a consolidated report: the deterministic result, the center-crop verdict for the 头条, and (if run) the X verdict. When the 头条 center verdict is NONE, say the cover has no complete center crop and suggest regenerating via the cover-image-platforms skill (WeChat 头条: centered-visual banner, share = center crop).

Response Handling

MUST:

• Present the deterministic failures and the VLM verdicts in plain language.
• Name the post directory and each cover file checked.
• When the 头条 center verdict is NONE, name the regeneration step: regenerate with the cover-image skill.

MUST NOT:

• Paste the raw SAM JSON unless the user asks.
• Re-attach the cropped images to the main conversation.
• Re-run the VLM judgment in the main agent.