agent-setup

name	agent-setup
description	Initialize or update setup for pi-side-agents (running asynchronous agents spawned/controlled from main session)

Parallel Agent Setup

Set up the pi-side-agents lifecycle scripts for this project.

Initial first time setup (no .pi/side-agent* yet): Work through two phases: interview, then file creation. Finish by backing up this agent-setup SKILL.md to .pi/side-agents/agent-start~/ for easy future upgrades.

Update setup: Chat with the user about what needs changing, use file examples below as reference for comparison.

Upgrade setup to new pi-side-agents version: Diff this file with the current backup .pi/side-agents/agent-start~/ and discuss the changes to apply (or not) with the user. When upgrade is finished, update the backup to the current version.

Important (local-only files): everything this skill creates is under .pi/ and is runtime/local configuration. Do not git add, git add -f, or commit these files. They should remain untracked. If they were accidentally committed, remove them from git tracking while keeping them locally.

Phase 1: Interview

Ask the user the following questions. You may ask them all at once or one at a time — use your judgment based on how they engage.

1. Main branch name – What is the primary integration branch? (default: main)

2. Bootstrap steps – Does each agent worktree need custom setup before work begins? For example: npm install, copying .env files, running migrations. If yes, what commands specifically?

3. Finish policy – When an agent finishes, should it:

   • Rebase locally, then fast-forward into the main branch in the parent checkout (default), or
   • Open a pull request instead?

4. Overwrite existing files – If .pi/side-agent-start.sh or similar already exist, overwrite them? (default: no — skip existing files)

Before asking the questions, autonomously try to answer questions 1, 2 and 4, and propose project-specific answers to the user.

Collect all answers before proceeding to Phase 2.

---

Phase 2: Create Setup Files

Determine the git repo root:

GIT_ROOT=$(git rev-parse --show-toplevel)

Create the three files below. For each file, check whether it already exists before writing — if it exists and the user said not to overwrite, skip it and tell the user. Otherwise write (or overwrite) it.

---

File 1: $GIT_ROOT/.pi/side-agent-start.sh

Write this file and make it executable (chmod +x).

Use MAIN_BRANCH set to whatever the user specified in question 1 (or main by default). The start script validates the branch but does not reset HEAD — the TypeScript extension already sets the worktree to the parent's HEAD. Do not force-update the local MAIN_BRANCH ref (e.g. git branch -f) because that branch is often checked out in the parent worktree.

Default content — substitute MAIN_BRANCH_VALUE with the actual branch name, then append bootstrap steps based on question 2:

#!/usr/bin/env bash
set -euo pipefail

PARENT_ROOT="${1:-}"
WORKTREE="${2:-$(pwd)}"
AGENT_ID="${3:-unknown}"
MAIN_BRANCH="MAIN_BRANCH_VALUE"

BRANCH="$(git -C "$WORKTREE" rev-parse --abbrev-ref HEAD 2>/dev/null || true)"
if [[ "$BRANCH" == "HEAD" ]]; then
  BRANCH=""
fi
if [[ -z "$BRANCH" ]]; then
  echo "[side-agent-start] Could not determine current branch in $WORKTREE."
  exit 1
fi

echo "[side-agent-start] agent=$AGENT_ID branch=$BRANCH main=$MAIN_BRANCH"

if [[ "$BRANCH" == "$MAIN_BRANCH" ]]; then
  echo "[side-agent-start] ERROR: child worktree is on $MAIN_BRANCH; expected a dedicated agent branch."
  exit 1
fi

# The worktree is already set to the parent's HEAD by the TypeScript extension.
# Just verify it's on the right branch.
echo "[side-agent-start] Worktree based on parent HEAD ($(git -C "$WORKTREE" rev-parse --short HEAD))."

• If the user wants no custom bootstrap: append the optional hook block:

  # Optional project bootstrap hook — create .pi/side-agent-bootstrap.sh to use.
  if [[ -x "$WORKTREE/.pi/side-agent-bootstrap.sh" ]]; then
    "$WORKTREE/.pi/side-agent-bootstrap.sh"
  fi
  

• If the user gave specific bootstrap commands: append them directly instead, e.g.:

  # Project bootstrap
  cd "$WORKTREE"
  npm install
  cp .env.example .env 2>/dev/null || true
  

---

File 2: $GIT_ROOT/.pi/side-agent-finish.sh

Write this file and make it executable (chmod +x).

Use MAIN_BRANCH set to whatever the user specified (or main by default).

For local rebase policy (default), use this content — substituting MAIN_BRANCH_VALUE with the actual branch name:

#!/usr/bin/env bash
set -euo pipefail

PARENT_ROOT="${PI_SIDE_PARENT_REPO:-${1:-}}"
AGENT_ID="${PI_SIDE_AGENT_ID:-${2:-unknown}}"
MAIN_BRANCH="MAIN_BRANCH_VALUE"
BRANCH="$(git rev-parse --abbrev-ref HEAD 2>/dev/null || true)"
if [[ "$BRANCH" == "HEAD" ]]; then
  BRANCH=""
fi

if [[ -z "$PARENT_ROOT" ]]; then
  echo "[side-agent-finish] Missing parent checkout path."
  echo "Usage: PI_SIDE_PARENT_REPO=/path/to/parent .pi/side-agent-finish.sh"
  exit 1
fi

if [[ -z "$BRANCH" ]]; then
  echo "[side-agent-finish] Could not determine current branch."
  exit 1
fi

LOCK_DIR="$PARENT_ROOT/.pi/side-agents"
LOCK_FILE="$LOCK_DIR/merge.lock"
mkdir -p "$LOCK_DIR"

MERGE_LOCK_TIMEOUT=120

iso_now() {
  date -u +"%Y-%m-%dT%H:%M:%SZ"
}

acquire_lock() {
  local payload started elapsed
  payload="{\"agentId\":\"$AGENT_ID\",\"pid\":$$,\"acquiredAt\":\"$(iso_now)\"}"
  started=$(date +%s)
  while true; do
    if ( set -o noclobber; printf '%s\n' "$payload" > "$LOCK_FILE" ) 2>/dev/null; then
      return 0
    fi
    elapsed=$(( $(date +%s) - started ))

    # Check if the lock holder is still alive (stale lock after crash/reboot)
    if [[ -f "$LOCK_FILE" ]]; then
      local holder_pid
      holder_pid="$(grep -o '"pid":[0-9]*' "$LOCK_FILE" 2>/dev/null | head -n 1 | grep -o '[0-9]*' || true)"
      if [[ -n "$holder_pid" ]] && ! kill -0 "$holder_pid" 2>/dev/null; then
        echo "[side-agent-finish] Removing stale merge lock (pid $holder_pid no longer running)."
        rm -f "$LOCK_FILE"
        continue
      fi
    fi

    if [[ "$elapsed" -ge "$MERGE_LOCK_TIMEOUT" ]]; then
      echo "[side-agent-finish] Timed out after ${MERGE_LOCK_TIMEOUT}s waiting for merge lock."
      echo "[side-agent-finish] Stale lock? Inspect: $LOCK_FILE"
      exit 3
    fi
    echo "[side-agent-finish] Waiting for merge lock... (${elapsed}s / ${MERGE_LOCK_TIMEOUT}s)"
    sleep 1
  done
}

release_lock() {
  rm -f "$LOCK_FILE" || true
}

trap 'release_lock' EXIT

while true; do
  echo "[side-agent-finish] Reconciling child branch: git rebase $MAIN_BRANCH"
  if ! git rebase "$MAIN_BRANCH"; then
    echo "[side-agent-finish] Conflict while rebasing $BRANCH onto $MAIN_BRANCH."
    echo "Resolve conflicts (git status / git rebase --continue), then rerun .pi/side-agent-finish.sh"
    exit 2
  fi

  acquire_lock

  set +e
  (
    cd "$PARENT_ROOT" || exit 1
    git checkout "$MAIN_BRANCH" >/dev/null 2>&1 || exit 1
    git merge --ff-only "$BRANCH"
  )
  merge_status=$?
  set -e

  release_lock

  if [[ "$merge_status" -eq 0 ]]; then
    echo "[side-agent-finish] Success: fast-forwarded $MAIN_BRANCH to include $BRANCH in parent checkout."
    rm -f "$(pwd)/.pi/active.lock" || true
    exit 0
  fi

  echo "[side-agent-finish] Parent fast-forward failed (likely $MAIN_BRANCH moved)."
  echo "[side-agent-finish] Retrying rebase reconcile loop..."

  sleep 1
done

For PR policy: write a finish script that pushes the branch and opens a PR via the gh CLI:

#!/usr/bin/env bash
set -euo pipefail

AGENT_ID="${PI_SIDE_AGENT_ID:-${1:-unknown}}"
MAIN_BRANCH="MAIN_BRANCH_VALUE"
BRANCH="$(git rev-parse --abbrev-ref HEAD 2>/dev/null || true)"
if [[ "$BRANCH" == "HEAD" ]]; then
  BRANCH=""
fi

echo "[side-agent-finish] Pushing $BRANCH..."
git push -u origin "$BRANCH"

echo "[side-agent-finish] Opening pull request against $MAIN_BRANCH..."
gh pr create --base "$MAIN_BRANCH" --head "$BRANCH" --fill

---

File 3: $GIT_ROOT/.pi/side-agent-skills/finish/SKILL.md

This is a skill for the child agent (not this session) that tells it how to finalize its work.

For local rebase policy, write:

---
name: finish
description: Rebase the branch with current work onto upstream and fast-forward it after explicit user sign-off (e.g. "LGTM")
---

# Parallel-agent finish workflow

When the user explicitly approves the work (e.g. says "LGTM", "ship it", "merge it"):

1. **Confirm** the finish action with the user before doing anything.

2. Run the finish script:

```bash
PI_SIDE_PARENT_REPO="$PI_SIDE_PARENT_REPO" .pi/side-agent-finish.sh

3. If the finish script exits with code 2 (conflict rebasing child branch onto MAIN_BRANCH_VALUE):

   • Stay in this worktree
   • Resolve conflicts (git status, then git rebase --continue)
   • Re-run the finish script after the rebase completes

4. If the parent-side fast-forward fails because MAIN_BRANCH_VALUE moved ahead:

   • The finish script retries the rebase reconcile loop automatically
   • Parent-side integration is a bit sensitive operation as it can make big mess; solve simple issues yourself, but escalate to the user with major issues (such as dirty parent worktree)

5. After success: report the landed commit(s). Suggest /quit if no further work is needed.

**For PR policy**, write a simpler finish skill:

```markdown
---
name: finish
description: Open a PR for the branch with current work to upstream after explicit user sign-off (e.g. "LGTM")
---

# Parallel-agent finish workflow

When the user explicitly approves the work (e.g. says "LGTM", "ship it"):

1. **Confirm** with user before pushing.

2. Run the finish script to push the branch and open a PR automatically:

```bash
.pi/side-agent-finish.sh

3. Suggest /quit if no further work is needed.

---

## Phase 3: Report

Tell the user which files were created, updated, or skipped, including backup/reference files, and how to proceed.

Explicitly remind the user that `.pi/side-agent-*` files are local runtime setup and should stay untracked (not committed to git).


- Backup reference for future diffs
- Start an agent: `/agent <task description>`
- Watch status: statusline shows active agents, ...@<number> is the tmux window to switch to; `/agents` lists all
- Ask you to set up and manage a flock of multiple side agents on your own to solve a task (you have the tools)