Claude Code on the Web: Why Your .env Vars Don't Reach the Setup Script (and How SessionStart Hook Fixes It)

TL;DR

• Environment variables you put in the .env panel of Claude Code on the web (Cloud Sandbox) do not reach the setup script.
• They only reach the shell inside the running Claude Code session.
• So git clone "https://x-access-token:${GH_TOKEN}@..." inside the setup script fails — GH_TOKEN is empty at that point.
• Move the clone into a SessionStart hook and it just works, because by then $GH_TOKEN is populated.

This is a write-up of how I narrowed down a behavior that the official docs don't spell out clearly.

Background

What I wanted to do

I've been collecting custom slash commands, skills, and a personal CLAUDE.md under ~/.claude/ locally. I wanted the same setup available inside Claude Code on the web. Everything is stored in a private repo called miyashita337/agent-base.

According to the docs, cloud sessions do not carry over user-level settings like ~/.claude/CLAUDE.md — only what's committed to the repo is available. So my plan was: on session start, clone agent-base and symlink its contents into ~/.claude/.

First attempt (failed)

I put a GitHub PAT into the .env panel of the Cloud Sandbox settings and tried to clone from the setup script.

# setup script
#!/bin/bash
set -e
git clone "https://x-access-token:${GH_TOKEN}@github.com/miyashita337/agent-base.git" "$HOME/agent-base"

.env panel:

GH_TOKEN=github_pat_...   # ~90 chars, real value
GIT_AUTHOR_NAME=miyashita337
...

Result: fatal: could not read Username.

Narrowing it down

Symptom 1: my echo output never showed up

For debugging I added echo "GH_TOKEN length: ${#GH_TOKEN}" in the setup script. But the setup-script UI only shows the last few lines of output, so anything mid-script gets silently dropped.

Workaround: dump everything to a log file

I rewrote the script to redirect diagnostics into /tmp/env-diag.log and then cat it from inside the session.

#!/bin/bash
set -e

LOG=/tmp/env-diag.log

{
  echo "===== ENV DIAGNOSTICS ====="
  echo "GH_TOKEN length: ${#GH_TOKEN}"
  echo "GIT_AUTHOR_NAME: [${GIT_AUTHOR_NAME}]"
  echo "TZ: [${TZ}]"
  echo "LANG: [${LANG}]"
  echo ""
  echo "--- All env vars (names only) ---"
  env | cut -d= -f1 | sort
} | tee "$LOG"

Then I started a new session and asked Claude Code to cat /tmp/env-diag.log.

Symptom 2: the cache trap

On some runs I didn't even see the "setup script executed" log line. Per the docs, the setup script runs only on first creation — after that, a filesystem snapshot is reused. Editing the .env panel alone does not invalidate the snapshot.

What does invalidate it:

• Changing the setup script body
• Changing the allowed-domains list
• ~7 days of age

So I added a throwaway comment line like # cache-bust 2026-04-19-01 at the top of the setup script. Semantically it's a no-op, but it's enough to force a rerun on the next session.

The result: every env var was empty

===== ENV DIAGNOSTICS =====
GH_TOKEN length: 0
GIT_AUTHOR_NAME: []
TZ: []
LANG: []
...
Custom vars found: 0 / expected 8

It wasn't just GH_TOKEN — nothing from the .env panel was reaching the setup script.

The clincher: the session shell has them

Inside the running Claude Code session, I checked the same variables directly from the shell:

$ echo "GH=${#GH_TOKEN}, TZ=[$TZ], LANG=[$LANG]"
GH=93, TZ=[Asia/Tokyo], LANG=[ja_JP.UTF-8]

All present. So:

Root cause and fix

Root cause

The .env panel is injected only into the Claude Code session shell, not into the setup script. It's not a bug — it's just not documented clearly. The docs show GH_TOKEN as an example .env value, but that's aimed at the gh CLI picking it up inside the session, not at setup-script usage.

Fix: move clone logic into a SessionStart hook

Drop the clone from the setup script. Put it in a SessionStart hook defined in the repo's .claude/settings.json. The hook runs after Claude Code has started, so $GH_TOKEN is in scope.

.claude/settings.json

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "startup|resume",
        "hooks": [
          {
            "type": "command",
            "command": "\"$CLAUDE_PROJECT_DIR\"/scripts/setup-agent-base.sh"
          }
        ]
      }
    ]
  }
}

scripts/setup-agent-base.sh

#!/bin/bash
set -e

AGENT_BASE_DIR="$HOME/agent-base"

# Skip locally
[ "$CLAUDE_CODE_REMOTE" != "true" ] && exit 0

# Clone (GH_TOKEN is available at this point)
if [ ! -d "$AGENT_BASE_DIR" ]; then
  if [ -z "${GH_TOKEN:-}" ]; then
    echo "setup-agent-base: GH_TOKEN is not set" >&2
    exit 1
  fi
  # Pass the token via extraHeader so it doesn't end up in .git/config
  git -c http.extraHeader="Authorization: Bearer ${GH_TOKEN}" \
      clone "https://github.com/miyashita337/agent-base.git" "$AGENT_BASE_DIR"
  git -C "$AGENT_BASE_DIR" config --unset-all http.extraHeader 2>/dev/null || true
fi

# Symlink into ~/.claude/ (idempotent)
mkdir -p "$HOME/.claude"
for dir in commands skills agents hooks; do
  src="$AGENT_BASE_DIR/$dir"
  dst="$HOME/.claude/$dir"
  if [ -d "$src" ]; then
    # If the destination is a real directory, back it up first
    # (ln -sf into an existing dir creates a nested symlink inside it)
    if [ -d "$dst" ] && [ ! -L "$dst" ]; then
      mv "$dst" "${dst}.bak.$(date +%s)"
    fi
    ln -sfn "$src" "$dst"
  fi
done

if [ -f "$AGENT_BASE_DIR/CLAUDE.md" ]; then
  ln -sf "$AGENT_BASE_DIR/CLAUDE.md" "$HOME/.claude/CLAUDE.md"
fi

exit 0

A few deliberate choices:

1. CLAUDE_CODE_REMOTE guard — early-exit outside of cloud sessions so local dev isn't affected.
2. Idempotent — skip clone if the dir exists, but always refresh the symlinks. Partial-failure recovery just works.
3. No token in the URL — use http.extraHeader so .git/config never contains a plaintext token.
4. ln -sfn not ln -sf — if the destination is already a real directory, ln -sf nests the symlink inside it (e.g. ~/.claude/commands/commands). Backing it up first and using -n (--no-dereference) forces a clean replacement.

What about the setup script?

Leave it empty. You can delete the diagnostics too.

Verification

Fresh session, ls -la ~/.claude/:

CLAUDE.md -> /home/user/agent-base/CLAUDE.md
commands  -> /home/user/agent-base/commands
skills    -> /home/user/agent-base/skills
agents    -> /home/user/agent-base/agents
hooks     -> /home/user/agent-base/hooks

Typing / shows all the custom slash commands from agent-base (/capture, /pdca, /inv, ...) and they execute without issue.

Gotchas summary

Closing

The "setup script can't see .env vars" behavior is inferable if you read the docs carefully — the gh CLI example hints at "this is for in-session auto-pickup" — but it's never stated plainly. Easy to misread as "you can use these in the setup script too."

Hope this saves someone else the afternoon I lost.

References

• Use Claude Code on the web — Claude Code Docs
• Hooks configuration — Claude Code Docs
• Original Japanese post: https://zenn.dev/harieshokunin/articles/b1064354319ce2