Feature: Team-scoped /retro mode for multi-repo orgs

[ethangui]

Summary

Add a /retro team mode that generates a user-agnostic, team-focused engineering retrospective across all repos in a directory. The current /retro and /retro global modes are personal — they orient around "you" (the person running the command). A team mode would orient around the org's collective output, designed to be shared in a team Slack channel or Monday standup.

Rationale

The gap today

/retro global does 90% of the work already — it discovers repos, gathers git data, computes streaks, context switching, and per-contributor breakdowns. But it's framed as a personal tool:

• The shareable personal card leads the output and only shows the current user's stats
• The narrative uses "your" framing throughout ("Your contributions", "Your Week")
• The contributor leaderboard puts "You (name)" first
• Per-project breakdowns include a "Your contributions" sub-section

For a team lead running retros for a Monday standup, this framing is wrong. They want:

1. Equal treatment of all contributors — no "you" vs "teammates" distinction
2. Team-level metrics first — total commits, LOC, streak, context switching across the org
3. Per-person analysis that reads like a manager's 1:1 prep — what each person shipped, praise, growth areas
4. Week-over-week comparison — "are we shipping more or less than last week?" is the docs: add README and CLAUDE.md #1 question
5. Scoped to a folder — not all AI session repos globally, just the repos that belong to this team/org

Why this belongs in gstack (not a custom skill)

I built a standalone custom skill to solve this for my org (~15 repos, ~16 contributors across two timezones). It works, but:

• It duplicates most of /retro's logic (repo discovery, git data gathering, streak calculation, hourly distribution, contributor aggregation)
• It misses improvements when /retro gets better
• The data gathering patterns are identical — the only difference is the framing and output structure
• Any team with a multi-repo setup would benefit from the same thing

What I learned building it

1. Multi-timezone coverage is a real pattern. My team spans US and Asia timezones. The hourly histogram shows near-24h shipping coverage when normalized to one timezone. /retro doesn't call this out — a team mode should detect and highlight timezone distribution automatically.

2. Context switching per day is the key team-health metric. How many repos had commits on a given day reveals whether the team is focused or fragmented. /retro global computes this but buries it. In team mode it should be prominent.

3. Week-over-week deltas are the docs: add README and CLAUDE.md #1 thing a manager wants. Not absolute numbers — deltas. "Commits up 15%, test ratio down 3pp, one new contributor." The snapshot/compare infrastructure already exists in /retro — team mode just needs its own snapshot namespace.

4. The personal card should be optional, not the lead. In team mode, the summary table and contributor leaderboard should lead. No personal card at all.

Proposed Design

Invocation

/retro team           — team retro, last 7 days, all repos in cwd
/retro team 14d       — team retro, last 14 days
/retro team ~/org/    — team retro, repos in specified directory

Behavior changes vs /retro global

Aspect	/retro global	/retro team
Framing	Personal ("you")	Team (no "you")
Lead section	Personal shareable card	Summary table + leaderboard
Repo discovery	AI session history	All git repos in directory
Contributor treatment	Current user first	All equal, sorted by commits
Per-project "Your contributions"	Yes	No — just top contributors
Week-over-week comparison	Optional	Prominent (designed for weekly cadence)
Snapshot location	~/.gstack/retros/	.context/team-retros/ (project-local)
Timezone analysis	Not called out	Auto-detect multi-TZ coverage

Shared infrastructure

Most of the implementation should be shared with /retro global:

• Repo discovery (generalized to "find git repos in directory")
• Git data gathering (commits, LOC, authors, timestamps, streaks)
• Contributor aggregation across repos
• Hourly distribution and context switching computation
• Snapshot save/load/compare

The only new code is the output template and the directory-scoped discovery.

Key implementation notes from my custom skill

• Discovers repos via ls -d <dir>/*/.git
• Detects default branch per repo (origin/main vs origin/master)
• Uses full paths for shell utils (/usr/bin/grep, /usr/bin/awk) — bare names fail in Claude Code's sandbox eval context
• Avoids IFS='|' read in for loops (breaks in the sandbox)
• Saves snapshots to .context/team-retros/ for week-over-week comparison
• Skips repos with 0 commits in the window silently
• For ~15 repos, the full data gather takes ~30-60s with parallel fetches

Happy to share the full standalone skill implementation if helpful.