fix: fall back to PATH when the baked agtermctl path is gone

[bot-rogerthat]

Fixes #472.

The problem

AgentHooksInstaller.bakeAgtermctlPath writes Bundle.main's absolute agtermctl path into the two installed wrappers, and the wrappers call it as "${AGTERMCTL:-agtermctl}". That default only fires when AGTERMCTL is empty, so the PATH rung agterm-agent-status.sh documents as step 3 is unreachable in any wrapper that was baked — and nothing ever checks whether the baked path still exists.

A bundle that moves after the install therefore kills the hooks. Installing from the mounted DMG is the easy way in: Bundle.main is /Volumes/agterm/agterm.app then, and the baked path dies on eject. The wrapper suppresses stdout/stderr and exits 0 by design, so nothing surfaces — the sidebar glyphs just stop appearing. Mine had been dead for weeks.

The fix

Bake a block that tests the path before committing to it:

# >>> agterm agtermctl path (installer-baked) >>>
if [ -z "${AGTERMCTL:-}" ]; then
  AGTERMCTL='/Applications/agterm.app/Contents/MacOS/agtermctl'
  [ -x "$AGTERMCTL" ] || AGTERMCTL="$(command -v agtermctl || true)"
fi

Both the assignment and the test sit inside the unset branch, so an explicit AGTERMCTL override is never second-guessed and the documented order stays 1 > 2 > 3. A moved bundle now degrades to the CLI from Help ▸ Install Command Line Tool… instead of going silent. Re-running the installer still heals the path outright; this is what covers the window before anyone thinks to.

The transform moves into host-free AgentHooksInstall — the type the rules already put in charge of the marker and merge policy — leaving AgentHooksInstaller the file glue. That is what makes it testable at all: the app-side helpers were private in a target swift test cannot reach, which is why the bake/strip pair had no coverage. It also lets stripBakedBlock take its line count from the block being emitted rather than a literal skip = 1 that would have silently gone stale as the block grew from one line to four.

Tests

agtermCore, all new:

• AgentHooksInstallTests — block lands after the shebang, carries the -x fallback inside the unset branch, and a re-bake replaces the block rather than stacking it (a second bake over a /Volumes one equals a first bake with the new path).
• AgentStatusWrapperTests — bakes the shipped wrapper through the installer's own transform and runs it with three stub binaries that record which of them was invoked: the baked one wins when it exists, PATH wins when the baked bundle is gone, and an explicit AGTERMCTL wins over both.

The two fallback tests fail on the previous block (the run records nothing at all, which is the bug), and the other two pass either way, so they pin the behavior that must not change.

What I ran

• cd agtermCore && swift test — green, 2607 tests.
• swiftlint lint --strict — 0 violations in 344 files.
• bash -n over both wrappers with the block baked in, and a standalone swiftc -typecheck -swift-version 6 -strict-concurrency=complete of the changed AgentHooksInstaller.swift against the built agtermCore.

I could not run make build or make test-app: this machine has Command Line Tools only, no Xcode 26, so scripts/setup.sh, xcodegen and the app target are out of reach here. The app-target change is the deletion of three private helpers and one call into agtermCore, but it has not been compiled in the real project and no XCUITests were run.

No user-facing action or control command changes, so no skill or site/ update. site/docs.html says the scripts are installed "with the bundled agtermctl's absolute path baked in so they work even without the CLI on your PATH", which stays true.

[umputun]

lgtm