feature: cursor-agent `monitor`, and pid name for cursor-agent

[apstndb]

Summary

Builtin cursor monitor is not useful unless the pid walk can see cursor-agent.

1. _agmsg_agent_binaries does not list cursor-agent, so cursor falls through to claude codex gemini. Watchers are not liveness-gated. Manifest already has detect_proc.
2. Builtin cursor is delivery_modes=turn off. Cursor Agent CLI has no Monitor tool; idle wake is Shell notify_on_output. Offer monitor (BETA is fine). Keep monitor=no on spawn.

Needs the #477 - fold (sibling #855 / #857) because the launch command uses "${CURSOR_CONVERSATION_ID:--}". That is a runtime dependency, not a git stack: #858 stays a draft until #857 and #860 have landed.

The grok-build pid mapping is independently mergeable on #859 / #860.

Pid name

. scripts/lib/resolve-project.sh
_agmsg_agent_binaries cursor        # claude codex gemini  (wrong)

Fix, before the * arm:

cursor)      echo "cursor-agent" ;;

The matcher already accepts "$bin"|"$bin"-*. Follow-up (not this issue): read detect_proc from type.conf for plugins (#631).

Shared basename agent

Do not add agent to _agmsg_agent_binaries or to detect_proc for cursor or grok-build.

Homebrew grok-build installs both grok and agent. The Cursor CLI installer also ships agent as an alias of cursor-agent. The basename is not type-distinctive. detect_proc globs are required to be disjoint (detect-cli-type.sh); a shared name would attach the wrong pid (#93).

Env detection already distinguishes the CLIs (CURSOR_AGENT / CURSOR_CONVERSATION_ID vs GROK_SESSION_ID). Pid walk still needs distinctive comm/argv0 (cursor-agent, grok) to find the enclosing process. Missing the alias is intentional. Disambiguating via cmdline or install path is a follow-up, not this issue.

Also do not list cursor (the desktop app code shim).

#860 comments this omission on _agmsg_agent_binaries.

Cursor monitor

• type.conf: delivery_modes=turn monitor off. Optional detect=CURSOR_AGENT CURSOR_CONVERSATION_ID.
• _delivery.sh: keep the existing turn .mdc + check-inbox.sh rule. Add a monitor rule that launches watch.sh once via Shell + notify_on_output. Status must distinguish turn vs monitor. Short AGMSG-DIRECTIVE on set monitor.
• template.md: offer monitor; Ensure-monitor; actas/drop 4th arg; mode accepts monitor, rejects both.

Do not invent a Monitor tool. Do not use /loop sleep as delivery. Do not change grok caller templates.

Maintainer merge order

1. fix(watch): fold sentinel '-' into the empty session-id path #857 and fix(resolve): map grok-build to grok in the pid walk #860 — independently mergeable, either order
2. Rebase feat(cursor): offer monitor delivery and name cursor-agent #858 onto main (picks up the fold, the grok arm, and the agent comment) and then merge. Keep feat(cursor): offer monitor delivery and name cursor-agent #858 a draft until that rebase; do not mark it ready earlier.

Related

• fix(watch): fold sentinel '-' into the empty session-id path #857 — watch.sh fold of sentinel - (needed at runtime by cursor monitor)
• bug: grok-build pid walk falls through to claude/codex/gemini #859 / fix(resolve): map grok-build to grok in the pid walk #860 — grok-build → grok pid name, plus the agent omission comment
• fix(resolve): detect a type's agent process from its manifest, not a fixed list #631 — read detect_proc from the manifest

[apstndb]

Implementation: #858 (draft, stacked on #857)

[apstndb]

The grok-build pid arm was split to #859 / #860. This issue is now cursor monitor + cursor-agent only (#858).

[apstndb]

I tried the proposed Cursor monitor (Shell notify_on_output + watch.sh) on current Cursor CLI and Cursor app. watch.sh itself is fine: matching lines print in both.
Cursor IDE starts a new agent turn from those lines while idle (including after ~26 minutes); Cursor CLI does not — the session stays at the prompt until I type. So this looks like a Cursor-host wakeup gap after turn-end, not something agmsg can fix in-process.
I did not find a public Cursor thread for CLI + mid-stream notify_on_output + idle-at-prompt; closest app-side reports are forum 162544 and forum 166335.
I’d keep #858 a draft BETA and not claim Cursor CLI idle-wake as working; app looks usable.