feed-score

name	feed-score
description	Score and publish AI Feed candidates through a controlled runner. Agent only writes scored-results JSON.
metadata	{"version":"2.4.0"}

Feed Score Skill

Score AI Feed candidates and publish generated posts. This skill is a controlled human-in-the-loop runner: scripts own repository state, validation, generation, optional build, git commit, and push; the agent only performs scoring judgment and writes scored-results.json.

Use When

• Scheduled AI Feed score/publish cron fires.
• A user asks to score collected candidates.
• You need to finalize or clean up a half-finished score run.

Core Rule

Do not execute the old Markdown procedure step-by-step. Use feed_score_ctl.py.

Allowed:

• exec for the commands below.
• read the generated task JSON and references/scoring-rules.md.
• write exactly the scored-results JSON path specified by the task.

Forbidden:

• git add -A, git add ., or committing arbitrary files.
• Editing repository files directly other than the scored-results JSON handoff.
• Using message directly; cron delivery handles logs.
• Comparing current candidates against data/seen.json for duplicate decisions.

Commands

Prepare Task

python3 scripts/feed_score_ctl.py prepare

Prepare outputs a JSON task path. Read that task before scoring.

If prepare returns status=no_content, stop successfully.

Prepare limits each task to a small batch by default (FEED_SCORE_LIMIT=30). Score only the task candidates array. Remaining candidates stay in data/candidates.json for later runs.

Score Handoff

Write a top-level JSON object to the task scoredResultsPath:

{
  "evaluated": 3,
  "scoredAt": "2026-06-14T17:00:00+08:00",
  "results": []
}

Rules:

• Every task candidate must appear exactly once in results; do not include URLs outside the task batch.
• verdict is only publish or skip.
• score >= 7.0 should usually publish; lower scores must not publish.
• publish entries must include all fields required by references/scoring-rules.md.
• skip entries must include reason; duplicate skips must include duplicateOf.

Finalize

python3 scripts/feed_score_ctl.py finalize --push

Finalize validates scored results, generates posts, commits/pushes generated posts, clears processed candidates, and prints JSON summary.

By default finalize skips the full Astro build so cron stays short and does not leave generated artifacts dirty. Use FEED_SCORE_RUN_BUILD=1 only for manual strict verification; FEED_SCORE_BUILD_TIMEOUT caps that optional build.

Output Contract

Final reply should copy final from runner stdout when present. Otherwise use:

📋 评分完成 HH:MM — <message>; generated=<generated>; cleanup=<cleanupCommitted>; pushed=<publishPushed|cleanupPushed>

status=ok and status=no_content are successful cron outcomes. status=failed or shell non-zero is failure.

References

• references/scoring-rules.md: scoring dimensions, thresholds, output schema.
• scripts/feed_score_ctl.py: prepare/finalize state machine.
• scripts/generate-posts.py: deterministic Markdown generator.
• scripts/validate-score-results.py: schema and coverage validator.