morning-recap

morning-recap

Recap merged PRs in a GitHub repo over the last N hours, skip ones the user authored or has already touched (any review state / comment / @-mention), and rank the rest by review priority.

The script filters server-side using GitHub's -involves:USER search qualifier — covers author, assignee, commenter, reviewer, and @-mentioned. Excluded PRs never come back over the wire.

Setup (one-time)

This skill is repo-agnostic but needs to know which repo to recap. Pick one:

Option A — set a default in your shell (recommended if you only watch one repo):

# ~/.zshrc or ~/.bashrc
export MORNING_RECAP_REPO="owner/repo"

Option B — pass it inline each time:

MORNING_RECAP_REPO=owner/repo /morning-recap 12

Option C — edit the default in run.sh (change the REPO default near the top).

You also need:

• gh CLI installed and authenticated (gh auth status)
• jq installed

The exclude-user defaults to your authenticated gh user; override with EXCLUDE_USER=someoneelse if needed.

Inputs

• Optional positional argument: lookback in hours (integer). Default 12.

Steps

1. Run the helper script. It does the GitHub queries, filters PRs you've touched, trims bodies, and returns compact JSON. Do not call gh directly — the script consolidates everything in one call.

   bash "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/skills}/morning-recap/run.sh" "${HOURS:-12}" 2>&1
   

   If the skill is installed under ~/.claude/skills/morning-recap/, the absolute path works too. When loaded as a plugin, $CLAUDE_PLUGIN_ROOT points at the plugin's skill dir.

   Stderr (one line) reports the window, the filters applied, and how many PRs came back. Stdout is a JSON array of the remaining PRs:

   [
     {
       "number": 11652,
       "title": "...",
       "author": "yhpark",
       "mergedAt": "...",
       "url": "...",
       "additions": 19,
       "deletions": 4,
       "changedFiles": 4,
       "files": ["..."],
       "body": "trimmed, comments stripped, capped at 500 chars"
     }
   ]
   

   Env overrides: MORNING_RECAP_REPO=... (or legacy REPO=...) and EXCLUDE_USER=....

2. Recap section — print a chronological markdown list:

   ## Merged in the last <N>h (<count> PRs — excludes ones you authored or already touched)
   
   - [#1234](url) **Title** — @author — one-line summary
   - …
   

   • One-line summary derived from title + body. Be concrete (what changed), don't just rephrase the title.
   • If body is empty/boilerplate after trimming, fall back to title alone.
   • Never invent details that aren't in the title or body.
   • If the JSON array is empty: print No unreviewed PRs merged in the last <N>h. and stop.
   • If the API returned exactly 100, mention the window may be truncated.

3. Review priority section — classify each PR High / Medium / Low using these heuristics. Lean on judgment; rules are signals, not a checklist.

   High signals (any one is usually enough):

   • Touches shared state stores (atoms, providers, query client config)
   • Touches navigation root or tab navigators (*Navigator.tsx, *ScreenStack*)
   • Touches secure storage, MMKV/persistence, or react-query cache layer
   • Performance-critical primitives: list virtualization, animation hooks, suspense boundaries, hooks consumed app-wide
   • Large diff (>500 LOC) or >15 files changed
   • Title/body mentions "refactor", "rendering cascade", "perf", "CPU", "stringify", "suspense", or removes a widely-used hook

   Medium signals:

   • Feature-area UI changes that touch a shared component used in >1 screen
   • New animation logic, gesture handlers, or useEffect with non-trivial deps in a hot path
   • Inconsistent styling primitives (raw View/Text/StyleSheet in a Tamagui/NativeWind-styled area)
   • Diff 100–500 LOC

   Low signals:

   • Localized cosmetic / alignment / spacing fixes
   • Single-file change under 100 LOC with no shared-state or list/animation files
   • Title is fix(...): center align …, fix padding, etc.

   Adapt the heuristics to the target repo — these defaults assume a React Native app. For a backend or library repo, substitute the relevant hotspots (migrations, public API surface, auth, schema, etc.).

4. Print priorities sorted High → Medium → Low (drop the Low list if it's long, summarize):

   ## Review priority
   
   ### High
   - [#NNNN] **Title** — concrete reason it's worth a look (1 line). Files: `path/a.ts`, `path/b.tsx` (+N more)
   
   ### Medium
   - [#NNNN] **Title** — reason
   
   ### Low (skim)
   - [#NNNN], [#NNNN] — single-file cosmetic fixes
   

   The "why" line should name a concrete risk rather than restate the title. Cap file list at 3 paths + "(+N more)".

5. End with a one-line recommendation: "Top N to actually open: #X, #Y, #Z."

Notes

• The script consolidates all gh calls into one — don't loop gh pr view per PR. If you need extra detail on a single PR, fetch only that one.
• Don't fetch full diffs — file paths + diff size + (trimmed) body is enough for a triage call.
• This is triage, not review — don't try to render verdicts on correctness.
• If the script errors with gh auth issues, surface verbatim and stop.
• Heuristics are guidance: a small diff that swaps the root provider is High; a 600-line copy edit is Low. Use judgment.

Optional: scheduled daily recap

cron-runner.sh is a launchd-friendly wrapper that runs the recap headlessly, summarizes via claude -p, writes markdown to ~/Documents/Claude/morning-recap/YYYY-MM-DD.md, and fires a macOS notification.

It's personal-setup-flavored (hardcoded paths, macOS-only osascript, assumes the claude CLI is installed). Treat it as a starting point — copy it somewhere editable and adjust SKILL_DIR, OUT_DIR, MORNING_RECAP_REPO, and the schedule for your machine.

Example launchd plist snippet:

<key>ProgramArguments</key>
<array>
  <string>/bin/bash</string>
  <string>/path/to/morning-recap/cron-runner.sh</string>
</array>
<key>StartCalendarInterval</key>
<dict>
  <key>Hour</key><integer>7</integer>
  <key>Minute</key><integer>30</integer>
</dict>
<key>EnvironmentVariables</key>
<dict>
  <key>MORNING_RECAP_REPO</key><string>owner/repo</string>
</dict>