Session Bootstrap

Session Bootstrap

Infrastructure skill that provides cloud environment setup and coordinator lifecycle hooks.

Architecture

Cloud environments are ephemeral. Two scripts handle setup at different lifecycle points:

On a resumed session, bootstrap-cloud.sh takes <1 second — it only checks that venvs, openspec, and skills still exist. If something was deleted mid-session, it repairs it.

Claude Code Web

• Setup Script: Paste setup-cloud.sh into Environment Settings > Setup Script
• SessionStart hook: Wired in .claude/settings.json (committed to repo)
• Pre-installed: Python 3.x, uv, pip, npm, pnpm, docker, git, PostgreSQL 16, Redis 7.0

Codex

• Setup Script: Configure in environment settings (cached up to 12h)
• Maintenance Script: Use bootstrap-cloud.sh as the maintenance script for resume
• Pre-installed: Common languages and tools via codex-universal image

Shipped Scripts

Wiring for Target Repos

1. Cloud Setup Script (Environment Settings UI)

The Setup Script field is a text area in the cloud UI (not committed to git). A single snippet must serve two repo layouts:

• Canonical layout (this repo): the script ships at <skill-base-dir>/scripts/setup-cloud.sh. The runtime mirrors .claude/skills/ and .agents/skills/ are gitignored, so a fresh clone does NOT contain them — setup-cloud.sh rebuilds them with the source-repository installer.
• Mirror layout (consumer repos): there is no canonical skills/; the script is committed at <repo>/.claude/skills/... and <repo>/.agents/skills/....

Paste this one snippet into Environment Settings > Setup Script — it works for both layouts and both harnesses (Claude Code web and Codex):

# Find session-bootstrap's setup-cloud.sh across BOTH repo layouts:
#   canonical source checkout: session-bootstrap/scripts/setup-cloud.sh
#   mirror     <repo>/.claude/skills/... or <repo>/.agents/skills/...   (mirrors committed)
# At Setup-Script time pwd is often the PARENT of the clone, so search downward.
matches="$(find "$(pwd)" -maxdepth 8 \
  -path '*/session-bootstrap/scripts/setup-cloud.sh' \
  -not -path '*/.git-worktrees/*' -not -path '*/node_modules/*' -print)"

# Reduce every match to its repo root, then dedupe: one repo may expose the
# script in 1-3 layouts, but they must all resolve to a SINGLE root.
roots="$(printf '%s\n' "$matches" | sed '/^$/d' | while IFS= read -r m; do
  r="${m%/skills/session-bootstrap/scripts/setup-cloud.sh}"   # source-contribution-only layout normalization
  r="${r%/.claude}"; r="${r%/.agents}"                        # strip mirror dir, if any
  printf '%s\n' "$r"
done | sort -u)"

nroots="$(printf '%s\n' "$roots" | sed '/^$/d' | wc -l | tr -d '[:space:]')"
if [ "$nroots" -ne 1 ]; then
  printf 'setup-cloud.sh: expected exactly 1 repo, found %s:\n' "$nroots" >&2
  printf '%s\n' "$matches" >&2
  exit 1
fi

# One repo — prefer canonical skills/, then the Claude mirror, then the Codex mirror.
for cand in skills .claude/skills .agents/skills; do
  candidate="$roots/$cand/session-bootstrap/scripts/setup-cloud.sh"
  if [ -f "$candidate" ]; then
    echo "setup-cloud.sh: using $candidate"
    exec bash "$candidate"
  fi
done
echo "setup-cloud.sh: no runnable copy under $roots" >&2
exit 1

For Codex's Maintenance Script field (runs on resume), paste the same snippet with every setup-cloud.sh replaced by bootstrap-cloud.sh.

Why search-then-resolve instead of a literal path. CLAUDE_PROJECT_DIR isn't set yet at Setup-Script time (Claude Code injects it later, for hooks), and on Claude Code web $(pwd) is the parent of the clone — typically /home/user, while the repo lives at /home/user/<reponame>/. The older bash "$(pwd)/.claude/skills/session-bootstrap/scripts/setup-cloud.sh" therefore resolves to /home/user/.claude/... and fails "file not found" — and it hard-codes the mirror layout, which the canonical repo no longer ships.

The snippet instead searches downward, collapses every hit to its repo root, and:

• Fails loudly when the matches span more than one repo root (sibling repos cloned under /home/user, each carrying session-bootstrap — common on accounts that clone multiple projects). Failing fast is safer than -print -quit, which would silently bootstrap whichever repo find visited first — filesystem-order dependent and repo-wrong more often than you'd think.
• Within the one repo, prefers canonical skills/, then .claude/skills/, then .agents/skills/ — so the canonical repo runs its source-of-truth copy (and rebuilds the gitignored mirrors via install.sh), while a mirror-layout consumer repo, which has no skills/, falls through to its committed .claude/skills/ copy. The dedupe-by-root step keeps the canonical repo's local-dev checkout (where all three layouts coexist after install.sh) from tripping the multi-repo guard.

Once a copy is chosen, setup-cloud.sh derives its own PROJECT_DIR from BASH_SOURCE[0] (git-root walk), so uv sync / npm install / install.sh run in the right directory regardless of which layout was matched.

The wrapper intentionally avoids mapfile and process substitution: some cloud runners invoke the field via /bin/sh, and some images still ship older Bash where mapfile is unavailable. The find + sort -u + wc -l variant is portable across both. Keep each -path argument on a single line when copying into the cloud UI; a wrapped line that injects whitespace inside session-bootstrap/scripts/... breaks the match.

2. .claude/settings.json — Hooks

Hook path depends on layout. The template below uses .claude/skills/…, which is correct for mirror-layout consumer repos (mirrors are committed and tracked, so the hook scripts survive a resume). In a canonical-layout repo where .claude/skills/ is gitignored (like this one), point the session-bootstrap hooks at canonical skills/session-bootstrap/scripts/… instead. Rationale: the runtime mirror can be wiped on an ephemeral resume, and bootstrap-cloud.sh is the hook that rebuilds it (verify_skills()) — if that hook lived only in the wiped mirror it could never run to repair itself. Canonical skills/ is tracked, so it always survives. (The langfuse Stop hook already follows this pattern.) $CLAUDE_PROJECT_DIR is set for hooks, so no find is needed here — unlike the Setup Script in §1, which runs before it is injected.

{
  "hooks": {
    "SessionStart": [{
      "matcher": "",
      "hooks": [
        { "type": "command", "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/skills/session-bootstrap/scripts/bootstrap-cloud.sh", "timeout": 30 },
        { "type": "command", "command": "python3 \"$CLAUDE_PROJECT_DIR\"/.claude/skills/session-bootstrap/scripts/hooks/print_coordinator_env.py" },
        { "type": "command", "command": "python3 \"$CLAUDE_PROJECT_DIR\"/.claude/skills/session-bootstrap/scripts/hooks/register_agent.py" }
      ]
    }],
    "Stop": [{
      "matcher": "",
      "hooks": [
        { "type": "command", "command": "python3 \"$CLAUDE_PROJECT_DIR\"/.claude/skills/session-bootstrap/scripts/hooks/report_status.py" }
      ]
    }],
    "SubagentStop": [{
      "matcher": "",
      "hooks": [
        { "type": "command", "command": "python3 \"$CLAUDE_PROJECT_DIR\"/.claude/skills/session-bootstrap/scripts/hooks/report_status.py --subagent" }
      ]
    }],
    "SessionEnd": [{
      "matcher": "",
      "hooks": [
        { "type": "command", "command": "python3 \"$CLAUDE_PROJECT_DIR\"/.claude/skills/session-bootstrap/scripts/hooks/deregister_agent.py" }
      ]
    }]
  }
}

3. Environment Variables (cloud UI)

COORDINATION_API_URL=https://coord.yourdomain.com
COORDINATION_API_KEY=<your-api-key>

Environment Variables

All hook scripts are stdlib-only (no third-party dependencies) and never block sessions (all exceptions swallowed, always exit 0).

Canonical Source

The hook scripts here are the complete runtime copies distributed by install.sh; they do not import or invoke coordinator source-repository scripts.