/catchup

Answer "where did we leave off?" — a fast, read-only recap of a project's recent activity and current open state. Pulls recent commits, CI/deploy health, open PRs (with mergeability), open issues, local git state, and session memory into one scannable summary, ending with a single suggested next step.

Use this at the start of a session to re-orient. It is the recap layer: /recon explains what a repo is, /start-working decides what to do next, and /catchup reports what just happened and what's open right now. It never edits, commits, or merges.

Arguments

• $ARGUMENTS - Optional configuration. Space-separated tokens:
  • First positional: repo slug override (owner/name). Defaults to the origin remote, then a configured fallback, then inference from the directory name.
  • scope=all|issues|prs|recent — Which sections to include (default: all).
  • since=N — Lookback window in days for commits and CI runs (default: 7).
  • limit=N — Max rows per list (default: 20).

Examples:

/catchup
/catchup scope=prs
/catchup owner/repo since=3
/catchup scope=issues limit=40

Instructions

0. Resolve Target and Mode (gate)

Determine whether there's a local checkout, and which GitHub repo to query. Everything downstream branches on MODE.

# Is the current directory a git work tree?
if git rev-parse --is-inside-work-tree >/dev/null 2>&1; then
  MODE="local+github"
  R=$(git remote get-url origin 2>/dev/null | sed -E 's#(git@github.com:|https://github.com/)##; s#\.git$##')
else
  MODE="github-only"   # e.g. a control directory with .claude/ but no clone
  R=""
fi

# Fall back to a configured slug, then to the first positional arg.
R="${ARG_REPO:-$R}"
R="${R:-blamechris/chroxy}"

If R is still empty (no remote, no fallback, no arg), try gh repo list --limit 50 and match the directory name; if still ambiguous, ask the user for the slug rather than guessing.

State the resolved R and MODE to the user in one line before continuing. In github-only mode, the Local git state section is skipped (print "— skipped (no local clone)") — this graceful degradation is the point of the gate, never error out on a missing clone.

Parse $ARGUMENTS for scope (default all), since (default 7), and limit (default 20).

1. Recent Activity — commits + CI/deploy

Run these queries in parallel where possible.

# Recent commits on the default branch (works without a local clone)
gh api "repos/$R/commits" \
  --jq '.[] | "\(.commit.author.date[0:10])  \(.sha[0:7])  \(.commit.message | split("\n")[0])"' \
  | head -"$LIMIT"

# Recently merged PRs for "what shipped" context
gh pr list -R "$R" --state merged --limit 8 \
  --json number,title,mergedAt -q '.[] | "#\(.number)  \(.mergedAt[0:10])  \(.title)"'

# Latest CI run health
gh run list -R "$R" --limit 8

# Deploy health on the default branch. Customize the workflow name; if the name is
# wrong or the repo has no deploy workflow, this simply returns nothing — treat an
# empty result as "no deploy info" and skip the line, never as an error (graceful
# degradation, consistent with the rest of the skill).
gh run list -R "$R" --workflow="Release" --limit 4

Summarize: what shipped recently, whether the latest default-branch run is green, and whether the most recent deploy succeeded. Call out any failing run on the default branch explicitly — that's the "something's on fire" signal.

2. Open PRs — with mergeability

gh pr list -R "$R" --state open --limit "$LIMIT"

For each open PR, pull the merge/CI rollup so stale or conflicting PRs are obvious. Loop over the already-capped list from above so $PR is defined and the number of gh pr view calls stays bounded by $LIMIT (no unbounded N+1 on repos with many open PRs):

for PR in $(gh pr list -R "$R" --state open --limit "$LIMIT" --json number -q '.[].number'); do
  gh pr view "$PR" -R "$R" \
    --json number,title,mergeable,mergeStateStatus,statusCheckRollup \
    -q '{num:("#"+(.number|tostring)), title:.title, mergeable:.mergeable, state:.mergeStateStatus,
         checks:([.statusCheckRollup[]?|.conclusion]|group_by(.)|map("\(.[0]):\(length)")|join(", "))}'
done

Flag each PR using mergeStateStatus + statusCheckRollup:

Signal	Meaning
BEHIND	Branch is behind the base — needs a rebase/update before it can merge
CONFLICTING / DIRTY	Has merge conflicts — needs manual resolution
FAILURE in checks	CI is red — may resolve after a rebase, or needs a real fix
CLEAN + all SUCCESS	Ready to merge

Note when several stale PRs predate a recent toolchain/dependency change on the default branch — their failures may simply be a missing rebase.

3. Open Issues — grouped

gh issue list -R "$R" --state open --limit "$LIMIT" \
  --json number,title,labels,updatedAt \
  -q '.[] | "#\(.number)  [\([.labels[].name]|join(", "))]  \(.title)"'

Group issues by the repo's label taxonomy so the list reads as themes, not a flat dump:

This repo has no consistent label taxonomy beyond enhancement and from-review — group by updated date instead of label theme, and flag any from-review issues (deferred PR feedback) at the top.

Surface the count, the most recently updated items, and any bug/critical/blocked issues at the top. Do not read every issue body — this is a recap, not triage (/start-working does the deep prioritization).

4. Local Git State (local+github mode only)

Skip this entire section in github-only mode (print "— skipped (no local clone)").

git status -sb                       # current branch + ahead/behind + dirty files
git stash list                       # forgotten work-in-progress
git log --oneline @{u}..HEAD 2>/dev/null   # unpushed local commits, if upstream set

Report the current branch, ahead/behind vs upstream, count of uncommitted/staged files, and any stashes — these are the "you were mid-something" signals.

5. Context Layer — memory + repo-memory (best-effort)

Both sources are optional; skip silently if absent.

• Auto-memory: if a project/auto memory exists (e.g. MEMORY.md or a memory/ dir), surface any project/feedback notes relevant to current work — these often record "why we left off here."
• repo-memory MCP: if the repo-memory server is enabled, call get_changed_files (and get_project_map for orientation) to show recently-touched areas. If the server is unavailable, skip without comment.

6. Report + Suggested Next Step

Print one scannable summary. Use tables for PRs and issues. Keep it tight.

## Where you left off — {R}

**Mode:** {local+github | github-only}   **Window:** last {since}d

### Recent activity
{1-3 lines: what shipped, latest CI/deploy state — green or red}

### Open PRs ({N})
| PR | What | State |
|----|------|-------|
| #573 | androidx bump | BEHIND, CI failing |
| #557 | AGP 9.1.0 | CONFLICTING |

### Open issues ({N})
{grouped by label theme, 1 line per theme; flag any bug/blocked at top}

### Local state
{branch · ahead/behind · N uncommitted · M stashed   — or "skipped (no local clone)"}

### Suggested next step
{ONE concrete action, e.g. "Rebase the 3 BEHIND Dependabot PRs against the new toolchain — likely cheap green wins; handle the conflicting AGP PR separately."}

End on exactly one suggested next step. Do not write any file.

Critical Rules

1. Read-only — /catchup never edits files, creates issues/PRs, commits, or merges. It does not merge or rebase anything; it only reports. The user decides what to act on.
2. Graceful degradation — every source is optional. No local clone → github-only mode (skip local git state). No gh auth, no deploy workflow, no memory, no repo-memory → skip that section and note it in the summary. Never fail the whole run on one missing source.
3. Recap, not triage — this is a fast orientation (target < 90 seconds). Don't read every issue body or trace code. For prioritized work selection use /start-working; for repo orientation use /recon.
4. One next step — end with a single concrete suggestion, not a menu. The user wanted "where did we leave off," not a backlog.
5. No attribution — never add any AI/agent attribution to output.