Support Cursor Agent (cursor-agent) as a harness adapter

[Quidge]

Goal

Add Cursor's cursor-agent CLI as a verified firstmate harness adapter, so crewmates and scouts can run on it — notably Composer 2.5 for fast, lower-guided work.

This issue is the tracking hub. Discrete problems and integration gaps are filed as separate issues under the harness:cursor-agent label and linked here (GitHub sub-issues). We are collecting issues from real usage before committing to full adapter development; an upstream contribution may follow once it is stable.

Status (2026-07-31)

• Capability verified end-to-end: cursor-agent -p --trust --model composer-2.5 "..." returns correct output (exit 0).
• Auth via cursor login (device flow; token lives in Cursor's own store, never in the repo — same pattern as gh auth). cursor-agent status reports login state.
• Not yet spawnable or supervisable by firstmate (no adapter wired in).

Observed behaviors relevant to adapter work

• Autonomy / trust: cursor-agent gates on workspace trust; a fresh directory needs --trust / -f / --yolo for non-interactive or autonomous runs.
• Model selection: --model composer-2.5 (also composer-2.5-fast); cursor-agent models lists all available models.
• Other flags: --worktree, --resume, --continue, -p/--print, --output-format text|json|stream-json, --trust.
• Install: curl https://cursor.com/install -fsS | bash installs to ~/.local/bin (symlinks agent, cursor-agent), versioned under ~/.local/share/cursor-agent/versions/, and auto-updates by default.

Integration surfaces (firstmate harness checklist)

• Launch command + autonomy flag — bin/fm-spawn.sh
• Detection env marker — bin/fm-harness.sh
• Busy-pane signature — bin/fm-watch.sh, bin/fm-tmux-lib.sh
• Empty-composer / ghost handling — bin/fm-composer-lib.sh
• tmux agent-liveness process classification — bin/backends/tmux.sh
• Exit / interrupt / resume facts
• Per-turn turn-end wake hook
• Docs: harness-adapters skill + verification evidence

Notes

• Vocabulary: cursor-agent is a harness (an agent CLI, like claude/codex/grok), not a runtime backend (tmux/herdr are backends).

[Quidge]

Possibilities to consider — tracked, not committed

Captured from discussion on 2026-07-31. These appear feasible; recording so we don't lose them. Not decided yet — the captain wants them tracked as options, not scheduled.

1. Tiered crew-dispatch by task difficulty (worker model)

Firstmate could route each task to a different cursor-agent model via config/crew-dispatch.json — firstmate's natural-language dispatch rules, evaluated quota-aware at intake:

• easy / mechanical / low-risk → cursor-agent --model composer-2.5 (fast, cheap)
• ambiguous / complex / high-blast-radius → cursor-agent --model cursor-grok-4.5-high

One harness (cursor-agent), model chosen per task via --model — both models appear in cursor-agent models. Depends on cursor-agent being a verified firstmate harness (this issue). This is the clean, first-class way to get difficulty-tiered worker models.

2. no-mistakes reviews on Composer via the bundled acpx ACP bridge

no-mistakes natively supports agent: cursor / acp:cursor, which runs cursor-agent acp through no-mistakes' own bundled acpx ACP client (see ~/.no-mistakes/config.yaml). If acpx is installed, no-mistakes review/fix agents could run on Composer (or another Cursor model).

• no-mistakes doctor currently reports cursor — not found (acpx) (acpx not installed on this box).
• Caveat: no-mistakes runs one configured agent per run — it does NOT take firstmate's per-task difficulty signal. So "Composer for easy reviews, Grok for hard reviews" is not native to no-mistakes per-review; that would need a per-run agent override (unverified). no-mistakes tiers the review agent per repo/global, not per task.

Net: worker-model tiering by difficulty is clean via crew-dispatch (option 1); review-model tiering inside no-mistakes is not native (option 2 caveat) — firstmate tiers the delivery path (direct-PR vs no-mistakes) rather than the review model per task.