Install an agent from the picker, on this Mac as well as on a box

[pallaoro]

Follow-up to #217. Asked as "why is the option not showing the Install button if I wanted to install it directly from Ateam?"

Why there was no button

The whole flow was SSH, so it could only ever mean "a box". installAgent(alias: string, agentId) (shared/host.ts:163) cannot express the local machine, and the implementation is ssh <box> "bash -lc '<install> && command -v <bin>'" (main/host.ts:597). The picker suppressed the button accordingly: canInstall = !avail && alias !== null && …. That shape came from #108, "install coding agents on a box", and left this Mac and the phone with no way in.

SSH was never a constraint here: installAgent already requires a connected backend (host.ts:594) and runs after the engine is up (host.ts:574-581).

What it is now

An engine RPC, the way editor:install already installs code-server on the task's own machine (dispatcher.ts, aggregate.ts). The engine doing the installing is the machine that has to end up with the binary, so one implementation serves this Mac, a box, and eventually the phone, which has never had an install path. Routed by the engine's projectId, the key that already lands a task on a given environment (unify.ts).

Installing keeps && command -v <bin> as its acceptance test, so a "done" cannot contradict the next launch, and reports the installer's own last lines when it fails, as installCodeServer does.

The half without which the button would lie

ensureLoginEnv resolves the PATH once at startup and returns on the first success (engine.ts:137, login-env.ts:165-168); the retry backoff exists only for a probe that failed. So the PATH this process spawns everything with is a snapshot of the machine as it was when the app opened, and a CLI installed afterwards is invisible until a restart, whoever installed it.

Seen for real: opencode's installer appends to ~/.zshrc, so a fresh interactive login shell finds it and the running app cannot.

refreshLoginPath re-resolves and re-adopts, rate-limited because an interactive login shell can take PROBE_TIMEOUT_MS to answer. Two call sites, and it needs both:

• the agent catalog, because a greyed-out row is unselectable, so no launch can ever be attempted to discover the agent has arrived;
• the launch guard before it refuses, because a loop tick never opens the picker that would otherwise have refreshed it.

Verification

• Typecheck clean. 413 pass / 0 fail, run both normally and with agent CLIs stripped from PATH (the CI condition).
• Nine new tests; the two guard tests proven non-vacuous by disabling the refresh and watching them fail.
• Live on a Mac whose PATH predates the install, no stubs: probe → absent, refreshLoginPath → changed, probe → present, and a real spawnAgentInTask("opencode") spawns and files the card running. No restart.
• installAgentCli exercised against a synthetic agent: resolves when the binary lands, rejects when command -v doesn't find it.

Three things the build corrected along the way: box installs do stream today, so dropping streaming was a regression rather than a limitation (followed installCodeServer instead: spinner on success, installer's last lines on failure); a test caught two clocks fighting in my own rate limiter; and dispatcher.test.ts didn't stub the refresh, so it ran a live login shell and mutated the test process's PATH, breaking three later tests that shell out to git.