apply-review

Apply Review Command

Consume the feedback JSON written by /bpe:review and apply the user's annotations to the reviewed markdown artifact. The user has already reviewed in the browser and clicked Save; this command surfaces their decisions and applies the changes after explicit confirmation.

Step 1: Find the Feedback File

If the user passed a feedback path as an argument, use that file. This is the escape hatch for replaying an older review or for the case where the user ran two reviews in parallel and the wrong one is "most recent."

Otherwise, find the most recent feedback JSON in the temp directory (/bpe:review generates the path with mktemp, which honors $TMPDIR):

ls -t "${TMPDIR:-/tmp}"/bpe-review-*-feedback.json 2>/dev/null | head -1

If none exists, tell the user to run /bpe:review first and stop.

If the chosen file's saved_at timestamp is more than 24 hours old, surface that to the user and ask whether they meant to apply this older feedback or run /bpe:review again.

Step 2: Load and Summarize

Read the feedback JSON. Expected shape:

{
  "sections": [
    {
      "id": "section-N",
      "heading": "<heading text>",
      "decision": "ship" | "update" | "redirect" | "reject" | "unset",
      "comment": "<free-text feedback>"
    }
  ],
  "global_comment": "<free-text>",
  "artifact_path": "<absolute path to the reviewed markdown file>",
  "saved_at": "<ISO timestamp>"
}

Each entry is one decision unit; the review is intentionally fine-grained, so expect many entries. Display a summary that front-loads the destructive decisions, since those are what the user most needs to verify before saying yes in Step 3:

• One-line counts header: N reject, N redirect, N update, N unset, N ship.
• Then group the non-ship entries under sub-headings, in this exact order (most-destructive first):
  1. Absolutely reject and delete (reject): list each unit's heading + comment. These are about to be removed; surface them first so the user can catch a mistake before authorizing.
  2. Completely off, do this instead (redirect): list each unit's heading + comment. These are full rewrites driven by the comment.
  3. Close, but update (update): list each unit's heading + comment. These are tweaks.
  4. No decision recorded (unset): list any units that have a comment (those'll be treated as update) and any that don't (those'll be left alone).
• The global comment, if non-empty.
• The artifact path that will be modified.