survey

name	survey
description	Portfolio archaeology: scan ALL local branches for forgotten/neglected work and rank by 'could Claude do something useful on this right now?'
allowed-tools	Bash, Agent

Survey — Forgotten Branch Archaeology

You are scanning every local branch to find work that was important but has been neglected. The goal is not recency (that's /work-context) and not a single branch deep-dive (that's /autopilot). This is: what have I forgotten, and can Claude make progress on it right now?

Step 1 — Enumerate all local branches

Run these in parallel:

git for-each-ref --format='%(refname:short)|%(committerdate:unix)|%(committerdate:relative)|%(subject)' refs/heads/

git worktree list --porcelain

gh pr list --state open --json headRefName,number,title,state 2>/dev/null || true

Step 2 — Filter and score candidates

For each branch, compute:

• Days since last commit (primary staleness signal)
• Has open PR? (importance signal — someone cares)
• Diverged commits from main (git rev-list --count main...<branch>) — invested work
• Branch name contains issue ID? (e.g. proj-123) — signals tracked work

Exclude immediately:

• main, master, develop, or your current branch
• Branches with 0 diverged commits from main and no open PR (nothing was done)
• Branches merged into main (check: git branch --merged main)
• Staging/deploy branches (names matching staging-*, deploy-*, release-*)

Score each remaining branch (higher = more interesting):

• +3 if has open PR
• +2 if diverged > 5 commits
• +2 if branch name contains a ticket/issue ID
• +1 per 7 days stale (capped at +5)
• −2 if diverged < 2 commits (may be abandoned stub)

Select the top 5–8 by score for deeper analysis.

Step 3 — Spawn one Explore subagent per candidate

For each candidate branch, launch an Agent (subagent_type: Explore) in parallel. Each agent receives:

"Investigate this branch: <branch-name>. Run:

• git log main...<branch> --oneline — what's been committed
• git diff main...<branch> --stat — scope of changes
• gh pr list --head <branch> --state all --json number,title,body,state,url — PR info
• grep for TODO/FIXME in files changed vs main
• If a worktree exists for this branch, check its CLAUDE.md and memory/ directory

Report (under 150 words): what's being built, what's done, what's obviously next, any blockers that would stop autonomous work, and a rough effort estimate if Claude ran autopilot on it now."

Do NOT spawn agents for branches you've already excluded.

Step 4 — Synthesize into ranked output

After agents return, organize into sections:

## Forgotten Work — Branch Survey

<N branches scanned> | <M candidates> | top <K> shown

### Ready for autopilot (Claude can proceed now)
1. **branch-name** — <X days stale> | <N commits ahead> | [PR #NNN]
   What: [1-2 sentence description]
   Next: [concrete next action]
   Runway: ~N hours before needing input

### Needs one decision first
2. **branch-name** — [the specific question that unblocks it]

### Has open PR but stalled
3. **branch-name** — [what's blocking merge]

### Low signal / probably dead
4. **branch-name** — [why it didn't make the cut; safe to delete?]

### Observations
- [Patterns: lots of stale PRs, common blocker type, cluster of related branches, etc.]

---
Pick a branch and run /autopilot to go deep on it.

Guidance on classification

• Ready for autopilot: Claude can write code/tests/docs on this right now without a decision from you. Clear next step, no blocking questions.
• Needs one decision first: The next action is clear but requires a choice only you can make (scope, naming, approach). Name the decision explicitly — don't just say "needs direction."
• Has open PR but stalled: Work is done but something is blocking merge — review, CI, a comment to address. Name it.
• Low signal / probably dead: Fewer than 2 commits, no PR, stale > 30 days. Flag as safe-to-delete candidate.

Keep the whole response tight. The user wants to pick a target and get moving, not read a report.