Applied review findings from the simplify pass:
1. Extract fromBunStream(name, get) and forkStderrDrain(stream, into)
helpers — 4 duplicated Stream.fromReadableStream call sites collapse
to one factory, and the identical stderr drain across serve/acp is
now a single helper. Error messages now include the underlying cause
instead of swallowing it.
2. acp.send awaits proc.stdin.write's backpressure promise. The bare
Effect.sync was discarding the Promise<number> form, which can
reorder ndjson lines under pipe-buffer-full conditions and corrupt
framing.
3. acp.close drops the try/catch around proc.stdin.end() — idempotent
in Bun, the bare catch only masked future regressions.
4. Effect.ignore on stream drains is now Effect.ignore({ log: true })
so a real protocol or decode error surfaces in test debug output
instead of disappearing silently.
5. Help-snapshots replaces the manual failures[] accumulator + continue
with Effect.partition. Same behavior, no mutable state, declarative.
324/324 CLI tests stay green; typecheck clean.
One test file. Spawns `opencode <cmd> --help` for every documented
command + key subcommand (35 in total) in parallel under concurrency:8,
snapshots the stderr output (yargs writes --help to stderr, not stdout).
Snapshots are normalized — the tmpdir prefix that bleeds through
`acp --cwd`'s default is rewritten to `<HOME>` so test runs in
different sandboxes stay stable. macOS `/private` realpath form and
the unresolved `os.tmpdir()` form both covered.
Pinned snapshots catch flag removals, renames, reordering, and exit-code
regressions across the entire user-facing CLI surface in one place.
Diff in the .snap file is the surface-change report.
Excluded: `opencode completion --help` is a yargs built-in that emits
top-level help and exits 1; not a real opencode command.
~6s wall-clock thanks to parallel spawns.
Kit Langton