pr-buddy（公关伙伴）

pr-buddy

When to use

Use when reviewing a PR where you want to build a real mental model — not rubber-stamp it.

Right for: unfamiliar codebase areas, large or architectural changes, any time you sense you're about to approve without truly understanding. Skip for trivial PRs (typos, dependency bumps, config-only) where the diff is self-evident.

See references/cognitive-debt.md for background.

Procedure

Skill intro (one-time greeting)

Render the pixel-art dog + greeting line + divider exactly once, as the first visible output of the skill. See references/terminal-colors.md § "Skill intro greeting" for the exact block.

Do not re-render the mascot at any later point — not between stations, not at the Phase 2 handoff, not at the review complete block. The mascot is a one-time greeting.

Pre-phase (silent)

Fetch the PR metadata and diff silently — output nothing until Station 1:

gh pr view <PR> --json number,title,body,author,baseRefName,headRefName,files,additions,deletions
gh pr diff <PR>

Build an internal mental model: goal, entry point, 2–3 interesting decisions, most plausible failure mode.

Compute the difficulty tier from the metadata using the formula in references/terminal-colors.md § "Difficulty line" (score = additions + deletions + files_count * 30 → easy / medium / hard / epic). Store difficulty, total_changes, and files_count internally — they feed the Station 1 header and the final Review Depth score.

Station transition

At the opening of every Station 1–6 and at the Phase 2 handoff, emit the station divider as plain text. Do not emit the mascot — that happens once at skill intro only.

Immediately below every Station 1–6 divider, render the cognitive debt meter line — see references/terminal-colors.md § "Cognitive debt meter" for the per-station values. The Phase 2 handoff divider does not carry a meter; the meter was already cleared at the review complete block.

Inline tree — available any time

If the reviewer asks to see the files ("tree", "show me the files", "what's changed", etc.), render the change tree and resume from exactly where you were — do not advance the station.

Session state is fully preserved (station, tracker marks, discussion context). After rendering, say: *"Where were we — ready to go on?"*

Render the tree as plain text with ANSI colors. See references/terminal-colors.md for the exact format.

Phase 1 — Socratic Walkthrough (6 stations)

At each station: ask → wait for response → reveal your reading (1–3 sentences, no spoilers).

All questions use the Question UX format (dim separator + bold yellow ▶). See references/terminal-colors.md.

Station 1 — Goal · divider: ── station 1 · goal

Print the PR header before asking:

PR #<number> — <title>
Author: <author>  ·  <baseRef> ← <headRef>
difficulty: <tier>  ·  +<additions> / -<deletions>  ·  <N> files

<body>
─────────────────────────────────────────────────────

See references/terminal-colors.md § "Difficulty line" for the tier formula.

Ask: *"Based on the title and description, what problem do you think this PR is solving? What's the outcome the author was aiming for?"*

Reveal: confirm, correct, or sharpen. State the goal in one sentence.

Station 2 — Architecture · divider: ── station 2 · architecture

Auto-render the change tree immediately after the station divider + meter (no trigger needed).

Ask: *"Which file is the entry point? Walk me through your mental model — how do these changes fit together?"*

Reveal: draw the actual call/data flow from the diff. Note anything surprising.

Open the file walkthrough tracker (main files only — skip tests, generated, lock, trivial config). Re-print it whenever a file is marked done. See references/terminal-colors.md for the colored render format.

Station 3 — Decisions · divider: ── station 3 · decisions

Pick 2–3 notable choices from the diff:

Ask: *"In <file>, the author chose [X]. What would you have done differently, if anything? Why do you think they made this choice?"*

Reveal: your reading of the trade-off and its implications.

See references/walkthrough-principles.md for what makes a decision worth discussing.

Station 4 — Code Tour · divider: ── station 4 · code tour

This station always runs. Emit the Station 4 divider + meter, then fetch and render function signatures for every main file in the walkthrough tracker.

Fetching file content:

For each file in the tracker, fetch its content from the PR head:

gh api repos/{owner}/{repo}/contents/{path}?ref={headRefName}

Decode the base64 content field. Fetch one file at a time. Use owner, repo, and headRefName from the PR metadata already fetched in the pre-phase.

Extracting signatures by file extension:

Rules: Show only functions new or modified in the diff (include when unclear). First line verbatim, truncated at 120 chars + …. Skip unnamed callbacks, IIFEs, test-only helpers. If none found: (no signatures found).

Render format (plain text, no code fence):

Code tour
──────────────────────────────────────────────────────
---scripts
  |---- deploy.mjs
  |       · deployToEnvironment(env, options)
  |       · rollback(deploymentId)
  |---- validate.mjs
          · validateSchema(input)
---src
  |---- auth/middleware.ts
          · export async function authenticate(req: Request, res: Response): Promise<void>
          · export function requireRole(role: string): Middleware
README.md
  · ## Overview
  · ## Installation
──────────────────────────────────────────────────────
  N files · M signatures total

Bullet char: · (U+00B7). Spaces not tabs. No code fence.

After rendering, ask:

────────────────────────────────────────────────────
\033[94m▶  Does this give you a clear enough picture of the code structure?\033[0m

   [1]  yes — move on to Station 5 · Risk     (default)
   [2]  go deeper — walk me through the most interesting function

Affirmative / enter / "1" → proceed to Station 5.

[2] Deep dive · divider: ── station 4b · deep dive

Skill picks the single most interesting function: highest churn in the diff, or most central to the PR goal if churn is equal. Do not ask the reviewer — pick and announce:

*"The most significant change is in <file> — <function signature>. Here's what it does:"*

Narrate in 3–5 sentences: what it does (from signature + diff context, not body), what changed vs. before, why it matters to the PR goal. Do NOT show the function body. Do NOT paste code. Prose only.

After narration, proceed directly to Station 5 — no further question.

Station 5 — Risk · divider: ── station 5 · risk

Ask: *"If you had to name one thing that could go wrong in production — edge case, race condition, missing test, performance concern — what would it be?"*

Reveal: surface the risk you identified from the diff. Compare notes.

Station 6 — Ownership · divider: ── station 6 · ownership

Ask: *"Could you explain this PR to a teammate right now, confidently? What, if anything, are you still unsure about?"*

Do not proceed to Phase 2 until the reviewer confirms they feel ready. If uncertain, offer to revisit the relevant station or re-read a specific section of the diff.

Once confirmed, render the Review complete block: debt meter at 0% cleared plus the Review Depth score and flavor tags. See references/terminal-colors.md § "Review complete block" for the exact format and the score computation rules.

The Review Depth score is reviewer-facing only — never include it in the final PR comment or in any output bound for the PR author/team.

Phase 2 — Automated Quality Checks

2a — Detect toolkit availability

pr-review-toolkit is available when the host is Claude Code AND the toolkit is registered (i.e. you can dispatch pr-review-toolkit:code-reviewer, pr-review-toolkit:silent-failure-hunter, and pr-review-toolkit:pr-test-analyzer as sub-agents). Otherwise it is unavailable — for example on GitHub Copilot or in a Claude Code session where the toolkit was never installed.

This is a branch, not a block. Phase 2 always proceeds.

2b — Build Phase 1 context (two artifacts, internal — never shown)

Artifact 1 — Structured block:

PHASE_1_CONTEXT:
  goal:          <one sentence from Station 1>
  entry_point:   <file from Station 2>
  decisions:     [<decision + trade-off from Station 3>, ...]
  reviewer_risk: <risk the reviewer named in Station 5>
  skill_risk:    <risk Claude identified in Station 5>
  confidence:    <reviewer's Station 6 self-report>

Artifact 2 — Reviewer narrative: 3–5 prose sentences synthesizing the walkthrough (what the reviewer said, what surprised them, concerns voiced, what Claude observed, confidence level). This is the only input for ## [MANUAL REVIEW] in the final comment — agent findings stay out of that section.

See references/pr-review-toolkit-bridge.md for the narrative format and an example.

2c — Phase 2 menu · divider: ── phase 2 · automated checks

Ask: *"How would you like to proceed?"*

When pr-review-toolkit is available:

  [1]  full review    — code-reviewer · silent-failure-hunter · pr-test-analyzer  (toolkit, parallel)
  [2]  code only      — code-reviewer                                              (toolkit)
  [3]  inline review  — host model runs the same three prompts, no toolkit needed
  [4]  stop here      — finalize with walkthrough findings only

When pr-review-toolkit is unavailable (Copilot, or Claude Code without the toolkit):

  [1]  inline review  — host model runs the three review prompts against the diff
  [2]  stop here      — finalize with walkthrough findings only

Renumber the inline / stop-here options to [1] / [2] when the toolkit is unavailable so the menu is always contiguous.

Branch on selection:

Full review (toolkit) — run all three sub-agents in parallel, each receiving the diff and both Phase 1 artifacts. See references/pr-review-toolkit-bridge.md for the exact prompt templates, escalation rules, and framing patterns. Template: references/output-template-full.md

Code only (toolkit) — run pr-review-toolkit:code-reviewer only (same prompt). Final comment's ## [AUTOMATED] includes only those findings. Template: references/output-template-full.md

Inline review (host-driven) — the host model runs the three prompts itself as three sequential analysis passes against the diff, each prompt receiving both Phase 1 artifacts. This is a feature-parity fallback: same prompts, same merge rules, no toolkit needed. See references/pr-review-toolkit-bridge.md § "Inline mode" for details. Template: references/output-template-full.md

Stop here — skip review; go directly to Final Summary. Template: references/output-template-walkthrough.md

Before launching review (any non-stop option) announce:

"Now running automated checks — with the full walkthrough context injected."

Wait for all passes/agents to complete before synthesizing.

Final Summary

Compose the comment in two passes, in this order, to keep the sections cleanly separated:

1. ## [MANUAL REVIEW] first — using only Phase 1 artifacts (the structured block and the reviewer narrative). Do not reference, paraphrase, summarize, or anchor to any agent finding here. This is the reviewer's voice — 4–7 sentences as an informed co-reviewer.
2. ## [AUTOMATED] second — using only the sub-agent / inline-review output. Compress each finding to one colloquial line with a severity emoji (🔴 / 🟠 / 🟡). No verbatim agent output, no per-agent sub-headings.
3. # Summary last — Blocking / Requests / Nice to Have — drawing from both sections, written from the walkthrough lens.

Then ask: *"Want me to post this? I'll run gh pr comment <PR> --body '...'"*

Only post on explicit confirmation.

Output format