fix(bootstrap): detect installed AI assistant before non-interactive skip check

[evansenter]

Summary

Fixes ordering in prompt_ai_install in bootstrap.sh. Previously, the non-interactive check ([[ ! -t 0 ]]) was evaluated before checking whether claude or agy was already installed (command -v). As a result, running ./bootstrap.sh -f in non-interactive environments (such as agent runners, CI, or subshells) defaulted INSTALL_AI to false and skipped AI directory and MCP server configuration even on machines where Claude or Antigravity was installed.

Changes

• In prompt_ai_install, check command -v claude || command -v agy before checking [[ ! -t 0 ]].
• Added unit test test_prompt_ai_install_detects_installed_assistant in tests/test-bootstrap.sh verifying that an installed assistant is detected even when stdin is not a tty.

Verification

• make check passes (129/129 hook tests, 54/54 bootstrap tests, shellcheck, syntax checks).
• Verified running ./bootstrap.sh -f in non-interactive subshell now configures AI setup.

[claude]

[Suggestion] Worth noting the blast radius this reorder opens up on the --pull/-p path, since it goes beyond the symlink/MCP configuration the PR description motivates.

With the installed-assistant check now ahead of the tty check, a non-interactive ./bootstrap.sh -f -p on a machine that has claude or agy also takes the INSTALL_AI == true branches at bootstrap.sh:573 (SteamOS Tailscale binary download), bootstrap.sh:658 (curl -fsSL https://tailscale.com/install.sh | sh on Debian/Ubuntu), bootstrap.sh:888 (pip3 install --upgrade piper-tts), and the Brewfile.ai bundle at bootstrap.sh:1026. Those are network-and-sudo side effects that previously could not happen without a human at the terminal.

This is not a defect — the Debian branch already runs sudo apt-get unconditionally a few lines earlier, so sudo-in-non-interactive is a pre-existing property of -p, and sudo without a tty errors rather than hanging. But if the intent was narrower ("configure AI dirs and MCP servers non-interactively") and not "install AI packages non-interactively", a tighter shape would be to keep INSTALL_AI=false for the package phase and gate only sync_dotfiles on assistant presence. Purely your call on which semantics you want.

[claude]

[Suggestion] Two small robustness gaps in the new test:

1. PATH="$stub:$PATH" prepends the stub rather than replacing PATH. On any machine where a real claude or agy is already on PATH — i.e. the maintainer machine, which is where make check usually runs — the command -v inside prompt_ai_install succeeds regardless of the stub, so the test would pass even if stub creation silently broke. Using PATH="$stub:/usr/bin:/bin" (or asserting the stub is what gets resolved) makes the test prove what it claims.

2. Only the positive branch is covered. The reorder must also preserve "nothing installed + non-interactive → INSTALL_AI=false", which is the behavior the original ordering existed to provide. A sibling case with an empty stub dir and a scrubbed PATH asserting false would pin both halves, and would catch a future change that drops the tty check entirely.

Minor style note: the neighboring tests extract functions through _load_<fn> helpers (e.g. _load_symlink_agy_configs immediately below). Inlining the sed/eval here works fine, but a helper would match the local idiom.

[claude]

Prompt: evansenter/dotfiles/.../claude-review.md

Code Review — Round 1

Summary

The reorder in prompt_ai_install is correct and minimal: the "already an AI machine" fast path now runs before the non-interactive skip, so ./bootstrap.sh -f under an agent runner or subshell no longer drops Claude/agy directory symlinking and MCP registration on machines that clearly want it. The caching guard stays first, so repeat calls across the ~6 call sites still short-circuit, and the interactive prompt path is untouched. The new test is a genuine regression test — under the old ordering with stdin redirected from /dev/null it would observe false and fail.

Previously Addressed (Filtered)

None — no prior automated review rounds on this PR.

Verdict

APPROVE - No blocking findings. Two non-blocking suggestions posted inline: the widened scope of the non-interactive -p package path, and test isolation plus negative-branch coverage.

---

Automated review by Claude Code