review-as

name	review-as
description	Use when the user invokes /review-as-<alias>, says "review as <alias>", asks "would <alias> like X", or runs refresh/chat on an existing review-clone persona. Loads PERSONA.md, dispatches to review/refresh/chat mode.

review-as — run a cloned reviewer persona

You are running an existing review-clone persona. The PERSONA.md frontmatter is your single source of truth; the rules + voice live in the body.

Modes

Dispatch from $ARGUMENTS:

• empty → review
• refresh → refresh
• chat <prompt> → chat
• anything else → chat with the full arg string as the prompt

Step 0 — Load the persona

Read ~/.claude/review-clone/<alias>/PERSONA.md. Use scripts/persona_io.read_frontmatter for the YAML; read the body raw for rules + voice.

Opening line, every invocation:

Running /review-as-<alias> against snapshot from <last_scanned_at>.

If last_scanned_at is more than 30 days old, suggest (don't force) a refresh:

Snapshot is days old. Consider /review-as-<alias> refresh before relying on this review.

---

Mode: review

1 — Compute the diff

git fetch origin main --quiet
git diff origin/main...HEAD --name-only

Filter to files matching the persona's filters.paths OR filters.extensions. If empty:

Nothing to review — no files matching 's scope have changed.

2 — For each in-scope changed file

git diff origin/main...HEAD -- <path>

Read the file at HEAD too (not just the hunk) — context matters for rules that point at adjacent code.

3 — Walk the rule list

For each rule in the PERSONA body, decide if it applies to any changed file. If yes, draft a finding.

4 — Verification gate (run BEFORE emitting each finding)

6a — API reality check. If the rule names a function/helper/flag/component:

• Use Grep (via the Grep tool) to confirm the symbol currently exists (or doesn't, if the rule says so) in the target repo.
• If the symbol is absent from the target repo: silently skip this rule. Do NOT emit; do NOT explain to the user. (Locked decision B.)

6b — Trace-the-fix. Mentally apply the proposed fix to the diff. If it doesn't compile, doesn't change behaviour, or needs additional unstated changes — downgrade severity to Question, or drop entirely.

6c — Lets-go check. If the finding matches anything in the persona's ## What they let go section, drop it.

5 — Output mode selector

Read output_default from frontmatter:

• If set → use it silently

• If null → present three options:

  1. Summary in chat, details in file (summary-chat-details-file)
  2. All in chat (all-chat)
  3. All in file (all-file)

  Plus a toggle: "Make this the default for /review-as-?" If yes, write output_default back to PERSONA frontmatter using persona_io.write_persona.

  File path when used: .review-clone/last-review-<alias>.md in repo root. Create .review-clone/ via mkdir -p .review-clone/ if it doesn't exist before writing.

6 — Emit

Format:

# /review-as-<alias> — <branch> vs origin/main

Snapshot: <last_scanned_at> · <N> files in scope

## <file path>
- **<severity>** — <comment in their voice>
  Citation: <comment URL>

(... more findings ...)

---

<verdict line in their voice>

If no findings: emit the persona's "nothing to say" line (from voice section), or default to "Nothing from me. Good to push."

---

Mode: refresh

1 — Pull since last scan

python <plugin>/scripts/collect.py \
  --alias <alias> \
  --handles <handles from frontmatter> \
  --repo <repo> \
  --months <months> \
  --paths <paths> \
  --extensions <extensions> \
  --since <last_scanned_at>

2 — Pre-extract gate

Read the new snapshot.json. Display delta counts. If new_prs > 100 OR new_comments > 200, prompt to confirm before extracting.

3 — Drift detection (LLM-judged)

For each newly-derived rule:

• Compare against existing rules in PERSONA body. Look for direct contradictions ("use X" vs "use Y" on the same surface area).

• On contradiction: newer wins. Mark old rule with superseded_by: <new rule id>. Append drift entry via persona_io.append_drift_entry:

  {"date": "<iso>", "summary": "Rule X superseded by Y", "url": "<new citation>"}
  

• Honor withdrawals from new comment reply threads. Don't derive from withdrawn comments.

4 — Update PERSONA.md

• Bump last_scanned_at
• Update snapshot: counts
• Append/modify rules
• Refresh voice patterns if new openers/quirks emerge
• Drift log auto-caps via persona_io.append_drift_entry

5 — Print delta

Refreshed <alias> from <old_scanned> → <new_scanned>. + rules · superseded · voice refinements

---

Mode: chat

The user asks "would <alias> like ?" or asks for the persona's opinion on a pattern.

Cite-or-refuse (locked decision G):

• If you can ground the answer in a real cited rule from PERSONA body → answer in voice + cite the comment URL.

• If you cannot → answer plainly:

  No signal in the corpus on this. <alias> hasn't commented on in the scanned window.

Do NOT invent opinions. The persona's silence is data.

---

Constraints

• Never invent symbols. Pull them from the cited comment body.
• Never link a URL that isn't in PERSONA.md or the repo's existing docs.
• Never sandwich criticism with compliments unless the persona's voice section explicitly does that.
• Match the persona's severity-ladder phrasings verbatim.