Configure Web Session

/configure:web-session

Check and configure a SessionStart hook that installs missing tools when Claude Code runs on the web.

When to Use This Skill

Context

• Install script: !find . -name 'install_pkgs.sh' -path '*/scripts/*'
• Settings hooks: !find . -maxdepth 3 -name 'settings.json' -path '*/.claude/*'
• Pre-commit config: !find . -maxdepth 1 -name '.pre-commit-config.yaml'
• Justfile: !find . -maxdepth 1 \( -name 'justfile' -o -name 'Justfile' \)
• Has helm charts: !find . -maxdepth 3 -name 'Chart.yaml' -print -quit
• Has terraform: !find . -maxdepth 3 \( -name '*.tf' -o -type d -name 'terraform' \) -print -quit

Parameters

Parse from $ARGUMENTS:

• --check-only: Report current state without creating or modifying files
• --fix: Apply all changes automatically without prompting
• --tools <list>: Comma-separated list of tool names to install (overrides auto-detection)
  • Supported: helm, terraform, tflint, actionlint, helm-docs, gitleaks, just, pre-commit

Execution

Execute this web-session dependency setup:

Step 1: Detect required tools

Auto-detect which tools are needed from project signals:

Step 2: Check existing configuration

1. Read .claude/settings.json if it exists
2. Look for a SessionStart hook that references install_pkgs.sh
3. Check scripts/install_pkgs.sh for completeness — verify each detected tool has an install block

Report current status:

Step 2b: Detect spec drift in an already-onboarded repo

An install_pkgs.sh that merely exists reads as compliant even when the canonical spec has moved on — the gap is invisible until a manual audit (see issue #1670). When the repo is already onboarded, compare the existing files against the current spec and report each item as PRESENT / DRIFT / ABSENT. Treat any DRIFT or ABSENT as a re-apply trigger, not a no-op:

Report drift as a positive signal:

If --check-only is set, stop here and print the status + drift report. Otherwise, re-apply the drifted items in Steps 3-5 (update pins, add Renovate annotations, wire path-bootstrap.sh first, replace latest lookups with pinned URLs) so the repo returns to spec.

Step 3: Build tool inventory

For each tool that needs to be installed, pin versions to match .pre-commit-config.yaml rev values where applicable. For tools not in pre-commit, use latest stable. Use this reference:

All download sources are compatible with the "Limited" network allowlist (github.com, releases.hashicorp.com, raw.githubusercontent.com, pypi.org).

Keep the pins fresh, not hand-maintained. Annotate each <TOOL>_VERSION="x.y.z" line with a # renovate: datasource=... depName=... comment and add a matching customManager to renovate.json so the pins are auto-updated rather than rotting (see .claude/rules/version-pinning.md). Where a tool's version is also pinned in .pre-commit-config.yaml (e.g. gitleaks), enable Renovate's pre-commit manager and group the dep so both bump in lockstep.

Step 4: Create or update scripts/install_pkgs.sh

Create scripts/install_pkgs.sh with:

1. Remote guard — exit immediately if not in a remote session:

   if [ "${CLAUDE_CODE_REMOTE:-}" != "true" ]; then
     exit 0
   fi
   

2. Idempotency guard per tool — use command -v <tool> before downloading:

   if ! command -v helm >/dev/null 2>&1; then
     # install helm
   fi
   

3. Install to ~/.local/bin — writable without sudo regardless of whether the session runs as root. Ensure this directory is on the PATH that agent subshells inherit: add it to a path-bootstrap.sh SessionStart hook (see this repo's scripts/path-bootstrap.sh) or append it to $CLAUDE_ENV_FILE. A bare export PATH=... inside the install script does not persist to later tool calls (see .claude/rules/sandbox-guidance.md).

4. Temp directory cleanup — use a temp dir per download, remove it after:

   tmp_dir=$(mktemp -d)
   # ... download and extract ...
   rm -rf "$tmp_dir"
   

5. unzip bootstrap — terraform and tflint ship as .zip; install unzip via apt if absent.

6. One install block per tool in this order: pre-commit, helm, terraform, tflint, actionlint, helm-docs, gitleaks, just.

Make the script executable: chmod +x scripts/install_pkgs.sh

Step 5: Update .claude/settings.json

1. Read existing .claude/settings.json (or start from {} if absent)
2. Add or merge the SessionStart hook:

"hooks": {
  "SessionStart": [
    {
      "matcher": "",
      "hooks": [
        {
          "type": "command",
          "command": "bash \"$CLAUDE_PROJECT_DIR/scripts/install_pkgs.sh\""
        }
      ]
    }
  ]
}

3. Preserve all existing permissions and other keys — do not overwrite them.

Step 6: Verify and summarise

Print a final summary:

Web session configuration complete
===================================
scripts/install_pkgs.sh  [CREATED/UPDATED]
.claude/settings.json    [CREATED/UPDATED]

Tools configured: helm, terraform, tflint, actionlint, gitleaks, just, pre-commit

Next steps:
1. Commit both files: git add scripts/install_pkgs.sh .claude/settings.json
2. Smoke-test locally: CLAUDE_CODE_REMOTE=true bash scripts/install_pkgs.sh
3. Run again to verify idempotency: CLAUDE_CODE_REMOTE=true bash scripts/install_pkgs.sh
4. Start a remote session on claude.ai/code and confirm tools are available

Step 7: Re-audit onboarded repos after a spec change (portfolio sweep)

When the canonical spec itself changes (new pinned tool, a wired-first path-bootstrap.sh, a download-source fix), every previously-onboarded repo silently falls out of spec — install_pkgs.sh still "exists", so nothing flags the drift (issue #1670). Make drift a positive signal: re-audit the whole portfolio rather than waiting for a manual cross-repo check.

Run /configure:web-session --check-only in each repo that already has scripts/install_pkgs.sh and collect the Step 2b drift reports. A thin sweep helper over the onboarded repos turns the silent non-event into an explicit PRESENT/DRIFT/ABSENT list — find the onboarded repos, then re-audit each:

# Discover onboarded repos under a portfolio root (each has scripts/install_pkgs.sh)
find . -maxdepth 3 -path '*/scripts/install_pkgs.sh' -print | while read -r script; do
  repo_dir=$(dirname "$(dirname "$script")")
  echo "=== ${repo_dir} ==="
  # Re-audit against the current spec (Step 2b drift checks)
  grep -q 'renovate:' "$script" && echo "renovate-pins: PRESENT" || echo "renovate-pins: DRIFT"
  find "${repo_dir}/scripts" -maxdepth 1 -name 'path-bootstrap.sh' -print -quit | grep -q . \
    && echo "path-bootstrap: PRESENT" || echo "path-bootstrap: ABSENT"
  grep -q 'api.github.com' "$script" && echo "allowlist: USES api.github.com (DRIFT)" || echo "allowlist: OK"
done

For each repo that reports DRIFT/ABSENT, run the full /configure:web-session (no --check-only) so Steps 3-5 re-apply the current spec, then open one PR per repo. Surface the deltas as a table so the sweep result is reviewable at a glance.

Agentic Optimizations