The installer (internal/installer/) wraps each backend in a streaming exec.Cmd and drives them in sequence. Three behaviours are worth calling out because they fix real bugs we hit.

PATH augmentation between steps

installer.Run prepends known dev-bin dirs (~/.local/bin, mise shims, brew dirs) to PATH at start, restores on exit, and re-augments before + after every step.

Without this, a mise install in step N landed in ~/.local/bin but step N+1's sh -c didn't see it, cascading mise: not found / npm: not found / npx: not found exits across the rest of the queue.

Available() pre-check + failedBackends

Non-bootstrap steps call i.Available() first; if false, emit a skipped — <backend> not on PATH line and mark the backend failed. Subsequent same-backend steps short-circuit with the same skip message instead of repeating exit-127 noise.

ANSI escape stripping

scanLines strips \x1b[...] escapes before forwarding output to the TUI log tail. Some installers (npx skills add, certain brew formulas) emit cursor-move + clear-line sequences inline with stdout; forwarding them raw scrambles the bottom of the frame.

Plan dedup

When a tool's name matches a bootstrap-able backend (e.g. user explicitly selects mise AND a tool with Source: "mise"), the auto-bootstrap is skipped — installing mise twice in one run was a real bug.

Live harness re-probe at skill-install time

skillsInstaller re-probes which AI harnesses are on PATH every time, instead of using the cached list from TUI startup. Reason: when a user installs codex + skills in the same run, codex isn't on PATH at detect time and skills only land in ~/.claude/skills/. Live re-probe picks up codex installed earlier in the run.

PostInstall failures = warnings

The main install (skill stub / brew install / etc) already succeeded; flagging the whole step for a Chrome-on-arm64 mismatch in agent-browser install was misleading. Surface a (main install still succeeded) line and continue.

Shell rc augmentation

internal/installer/shellrc.go::EnsureShellPath() writes a fenced # lfg-managed PATH block to every detected rc (~/.bashrc, ~/.zshrc, fish config), idempotent per file. Multi-shell so a future chsh doesn't lose the augmentation.

Always targets the interactive rc (.bashrc, not .bash_profile) so the matching exec bash reload command actually picks up the new PATH — login-shell bash on Linux sources .bash_profile, not .bashrc.

Shell detection priority:

1. $SHELL env var
2. /proc/<ppid>/comm parent process name (covers docker run -it bash where $SHELL stays unset)
3. Existing-file sniff
4. Last resort: linux → .bashrc, darwin → .zshrc