fix(hetzner): retry Claude native installer with backoff + clear error

[madarco]

Problem

The rebuilt Hetzner base snapshot shipped with no claude, so boxes from it had no agent — the in-box tmux session died instantly and attach crash-looped on no server running on /tmp/tmux-1000/default.

Root cause: the bake's curl -fsSL https://claude.ai/install.sh | bash -s stable got an intermittent HTTP 403 from Cloudflare (which throttles cloud-datacenter egress IPs, Hetzner among them, under load). The bare curl | bash masked it — curl -f exits non-zero on the 403, but the pipeline's status is bash's 0, so set -e never fired and the step "succeeded" while baking a claude-less snapshot.

Fix

• A retry_backoff bash helper retries the native installer 3× with 60s then 240s backoff (~5 min), keeps set -o pipefail, and folds command -v claude into the retried command so a "succeeded but absent" result also retries. If all attempts fail, the bake aborts (exit 71) rather than ship a broken snapshot. Applied to hetzner / vercel / e2b bake scripts + the docker Dockerfile.box (sh/dash plain-loop variant).
• No npm fallback (an earlier attempt used one): npm install -g @anthropic-ai/claude-code lacks native-only features and lands at /usr/bin/claude, mismatching the host-seeded installMethod=native and tripping Claude Code's "missing or broken" doctor warning.
• prepareHetzner now special-cases exit 71 with an actionable error (the generic one showed empty stderr because the install runs bash -x ... 2>&1 | tee, merging stderr into stdout).

Known gap (deferred, documented)

The 403 can outlast the ~5-min retry window, so prepare can still fail and need a manual re-run — this PR does not fully fix that case. The validated reliable fix (host-proxies the native binary download — claude install has no offline mode, but the downloaded binary placed directly at ~/.local/bin/claude runs offline) is written up in docs/hertzner_backlog.md as a follow-up.

Verification

• retry_backoff unit-tested (sleep stubbed) in bash + dash: succeeds attempt 1, recovers attempt 3 with 60s/240s waits, fails after 3.
• bash -n clean on all bake scripts; hetzner/vercel/e2b package tests pass (148 total).
• New exit 71 message bundled into the CLI dist.

Action required after merge

Rebake: agentbox prepare --provider hetzner --force (the fix only applies to new bakes).

https://claude.ai/code/session_019m5WHxP4vmsoXaHUhQdY9e

---

Note

Low Risk
Changes are confined to image/prepare install scripts and Hetzner prepare error messaging; no runtime auth or data-path changes. Residual risk is prepare still failing if CDN blocks persist beyond the retry window.

Overview
Fixes claude-less base snapshots when Cloudflare intermittently 403s the native installer from datacenter egress (e.g. Hetzner). A bare curl | bash could report success while claude never landed, breaking in-box tmux/attach.

Bake scripts (install-box.sh, provision.sh, build-template.sh, Dockerfile.box) now use a retry_backoff helper (3 tries, 60s / 240s waits), pipefail, and a command -v claude check; all failures abort with exit 71 instead of shipping a broken image. No npm fallback — native ~/.local/bin/claude stays aligned with host installMethod=native.

prepareHetzner maps exit 71 to an actionable “transient Cloudflare 403, re-run prepare --force” message (generic errors showed empty stderr because install output is tee’d to stdout).

docs/hertzner_backlog.md records the issue as resolved and notes the deferred host-proxied binary download if 403 outlasts ~5 minutes. Rebake required after merge for Hetzner bases.

^{Reviewed by Cursor Bugbot for commit af41081. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agentbox-web	Ready	Preview, Comment	Jun 27, 2026 12:30pm

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit af41081. Configure here.}

[cursor]

Exit 71 log path wrong

Low Severity

The new exit 71 error tells users the full install trace lives at /var/log/agentbox/install.log on a box from the resulting snapshot. A failed prepare never creates a snapshot and the temp VPS is deleted in the failure cleanup, so that path is not reachable.

 

^{Reviewed by Cursor Bugbot for commit af41081. Configure here.}