A monitor tool (plus monitor_list/monitor_stop) lets the model watch
a long-running background command's output for a pattern, or until it
exits, and get notified back into the session as new messages -- a
standing watch, not just a one-shot check. Analogous to Claude Code's
own Monitor tool. Gated behind OPENCODE_EXPERIMENTAL_MONITOR, matching
the existing experimentalBackgroundSubagents convention.

Ported from upstream anomalyco/opencode PR anomalyco#33806 "feat: add Monitor
tool" (1,398 lines, closed by the automated PR-cleanup bot, never
reviewed) -- built directly on this fork's existing BackgroundJob
primitive rather than modifying it. Reused nearly word-for-word: the
re-arm/replace parameter shape, session-scoped list/stop logic, the
untrusted-output fence with an unpredictable id so watched output
can't forge its own closing tag, control-character stripping, the
never-grant-"always"-for-a-glob-command permission rule, the
detached-process-plus-parent-death-watchdog design, the flood guard,
and most of the tool description and test scenarios. Rewritten rather
than ported: avoided anomalyco#33806's own change to core BackgroundJob (a
WakeScope to dodge a self-re-arm deadlock) by forking notifications
into the tool's own scope instead, the same pattern task.ts already
uses; its hand-written reader/debounce became an Effect Stream
pipeline; its base64 command-wrapper isn't needed since this repo's
bash tool already runs commands through plain ChildProcess.make;
permission goes through the existing bash permission rather than a
new one, so existing bash rules govern monitors too.

Also considered anomalyco#44191 (cron scheduler, closed by the same bot the
same week, praised by a human reviewer first) and the
anomalyco#47166/anomalyco#47193/anomalyco#47284 durable-heartbeat resubmits (each closed within
hours by a different, 2-hour PR-template-compliance bot, never
reviewed) -- neither ported; see known gaps.

Found and fixed a real, generalizable bug along the way: a command
that writes output continuously with no pause (e.g. `yes`) hung
teardown and starved timers, because the stdout read pipeline never
yielded control back to the scheduler while data kept arriving.
Fixed with a single Stream.tap(() => Effect.yieldNow), documented
inline -- the same issue could affect other handle.stdout readers
facing a non-pausing producer.

Known gaps (deliberately scoped): not durable, in-memory only; no
scheduler (only a shell poll-loop inside the watched command); opencode
run's non-interactive mode exits once idle, killing active monitors;
no UI (model-only via monitor_list); stdout only, 4000-char line cap,
5000-line flood guard; parent-death watchdog is POSIX-only; no
external/webhook delivery.

New packages/opencode/src/tool/monitor.ts (~360 lines) + monitor.txt.
New packages/opencode/test/tool/monitor.test.ts (13 tests: self-cancel
re-arm without deadlock, concurrency, list/stop, byte-flood guard,
multi-line commands) -- passes 3 runs in a row. Verified against this
fork's current dev (not just the agent's own stale worktree base):
bun typecheck clean; full packages/opencode suite 3600 pass, only the
4 pre-existing, confirmed-unrelated cf-ai-gateway failures.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>