feat(harness): add verified Cursor worker adapter (Cursor CLI + Grok 4.5)

[Amplify-Logic]

Intent

Verify Cursor CLI (the agent binary, 2026.07.16-899851b) with Cursor Grok 4.5 as a fully supervised Firstmate WORKER harness adapter, then implement the adapter only where every required behavior was empirically proven. The captain authorized this so Cursor/Grok can become the primary non-Codex volume pool; Codex is nearly exhausted, so no Codex model or Pi/Codex helper was used.

Deliberate decisions a reviewer reading only the diff would not know:

1. WORKER ONLY, intentionally. Primary-orchestrator support (session-start nudge, turn-end guard, PreToolUse seatbelt, watcher supervision protocol) was NOT verified and is deliberately NOT claimed. This mirrors the existing Kimi primary-only boundary in reverse. Kimi deliberately stays primary-only; this change does not add Kimi worker support.

2. Two defects in SHARED composer monitoring were found during verification and fixed in the shared owners rather than worked around per-harness. (a) Cursor draws the terminal cursor as a reverse-video cell (SGR 7) over the idle placeholder's first character; reverse video is neither dim/faint nor dark-foreground, so fm_composer_strip_ghost kept it and an idle composer reduced to a lone 'A' and classified as 'pending', which would defer every away-mode escalation forever. The obvious fix (an FM_COMPOSER_IDLE_RE override) does NOT work, because the idle regex was only matched against ghost-stripped content; it is now also matched against the plain row, which is styling-independent. (b) tmux's #{cursor_y} does not point at Cursor's composer (Cursor parks the cursor in its bottom status area, observed 5 rows off), so a composer holding real unsubmitted text classified as 'empty' - a FALSE-EMPTY, the dangerous direction, because the away-mode injector picks injection targets by emptiness and would type over pending input. Fixed with a structural composer-row scan mirroring herdr's existing approach, which is why herdr was never affected. Both are regression-covered against verbatim captured pane rows.

3. Cursor has NO effort flag - reasoning effort is a SUFFIX on the model id - so the effort axis folds into the model via cursor_model_with_effort instead of emitting a flag. xhigh/max cap at high per the harness-adapters effort-fallback rule rather than being dropped silently; an already-tiered or bracket-parameterized model id is never retiered; -fast variants are a deliberate separate cost/speed choice and are never selected implicitly.

4. The launch template deliberately omits -w/--worktree. Cursor worktree allocation is opt-in, and Firstmate already owns the isolated task worktree, so passing it would create a second one. --workspace pins the root instead. Verified: ~/.cursor/worktrees gained no entry across every launch.

5. Busy signature is the footer hint 'ctrl+c to stop', and the spinner VERB is deliberately NOT matched: it flips 'Working' to 'Running' mid-turn, so matching the verb reads a tool-executing pane as idle and trips premature stale detection.

6. CURSOR_AGENT=1 detection is deliberately ordered BEFORE the existing CLAUDECODE check, because Cursor does not clear an inherited CLAUDECODE=1; without that ordering a cursor worker spawned from a claude-hosted firstmate misreports as claude and gets steered with claude's interrupt/exit/resume vocabulary.

7. tmux liveness: Cursor's wrapper execs node so pane_current_command is 'node'. The existing code deliberately declined argv introspection for pi because a bare node is unattributable. Cursor is different - the versioned cursor-agent bundle path survives exec -a in argv - so argv inspection is applied narrowly to the node COMM only. Any other bare node still returns 'unknown', never 'dead', preserving the rule that a wrong 'dead' must never respawn over a live agent. pi's behavior is unchanged.

8. Backends were reviewed per the compatibility rule rather than assumed: tmux verified; herdr composer-safe by construction (structural scan + shared idle default) with its liveness path unverified for cursor; zellij/orca/cmux not exercised with cursor at all. Only tmux is claimed as verified.

9. Documented but not fixed: Cursor executes claude-format hooks (.claude/settings*.json, mapping Stop to stop), so cursor must never be launched from the firstmate primary checkout; and --model mutates the account-global default in ~/.cursor/cli-config.json.

10. Known incomplete acceptance criterion, deliberate: a live end-to-end fm-spawn dispatch was NOT run. fm-spawn calls 'treehouse get', which allocates a real pooled worktree and writes live fleet state outside the task worktree - something this crewmate must not cause. Every underlying behavior was verified by raw launches, and the integrated path (launch with folded model id, turn-end hook firing, composer empty, liveness alive) was confirmed together on a real pane. This is recorded as limitation 2 in docs/cursor-harness.md and left as firstmate's gate.

config/crew-dispatch.json is captain-private and was deliberately NOT edited; the proven profile values are reported in docs/cursor-harness.md for firstmate to apply after merge.

What Changed

• Added a Cursor CLI worker harness adapter (bin/fm-spawn.sh, bin/fm-harness.sh, backend scripts): launch template with --workspace pinning (no -w worktree), effort folded into the model id via cursor_model_with_effort (xhigh/max cap at high), CURSOR_AGENT=1 detection ordered before the CLAUDECODE check, busy detection on the ctrl+c to stop footer hint, and narrow argv-based liveness inspection for Cursor's node wrapper (other bare node panes still report unknown, never dead). Worker-only by design — primary-orchestrator support is not claimed.
• Fixed two shared composer-monitoring defects surfaced during verification: the idle regex is now also matched against the plain (non-ghost-stripped) row so Cursor's reverse-video cursor cell no longer classifies an idle composer as pending, and composer-row detection uses a structural scan (mirroring herdr) instead of trusting tmux #{cursor_y}, eliminating a false-empty that could let the away-mode injector type over pending input. The structural → scan is scoped to identified cursor panes per review feedback. Both fixes are regression-covered against captured pane rows in tests/fm-cursor-adapter.test.sh.
• Documented the adapter in docs/cursor-harness.md (including known limitations: Cursor executes claude-format hooks, --model mutates the account-global default, and the live end-to-end fm-spawn dispatch is deliberately left as firstmate's gate), plus updates to configuration, architecture, tmux-backend docs, and the harness-adapters skill. Only the tmux backend is claimed as verified.

Note: the pipeline's Test stage reported a failure attributed on re-check to a watcher-lock timeout flake; no code changes resulted.

Risk Assessment

✅ Low: The confirmed cross-harness composer-hijack regression from round 1 is now fixed with positive cursor-pane identification that fails safe toward the pre-existing cursor_y behavior, is regression-covered for both the classification and fm-send consequences, and the rest of the branch was already verified intent-conformant with the one remaining tradeoff (generic busy phrase) explicitly accepted by the user.

Testing

Completed 1 recorded test check.

• Outcome: ⚠️ 1 error across 2 runs (50m22s)

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

🔧 **Review** - 3 issues found → auto-fixed ✅

• ⚠️ bin/fm-tmux-lib.sh:127 - fm_tmux_cursor_composer_row is applied unconditionally to every tmux pane: fm_tmux_composer_state has no harness scoping, so on a non-cursor pane (claude/codex/grok/pi) any visible line matching '^[[:space:]]*→ ' — e.g. an arrow-prefixed bullet in model output or tool output — redirects composer classification from #{cursor_y} to that output line. The redirected row then reads as pending/unknown, which (a) defers away-mode escalations indefinitely (the same false-pending wedge class this change fixes for cursor) and (b) makes fm_tmux_submit_enter_core retry Enter and return a false 'pending' swallow verdict to fm-send, which can cause duplicated steers. The cited herdr precedent is narrower: it only promotes bordered rows and ❯/› prompt-glyph rows, which don't appear in ordinary output text. Consider scoping the scan by harness (thread an expected-label like the other backends) or constraining the → row to sit directly above cursor's status rows. Flagged ask-user because the unconditional design appears deliberate ('Harnesses that draw no "→ " composer are untouched' assumes → never appears in other harnesses' output).
• ℹ️ bin/fm-watch.sh:121 - The new busy alternation 'ctrl+c to stop' is a generic phrase common in dev-server/tool output ("Press CTRL+C to stop"); fm_pane_is_busy and fm-watch match it case-insensitively over the last 6 non-blank lines, so tool output lingering near the pane bottom after a turn ends can read an idle pane as busy and delay stale/turn-end handling. Same accepted-risk class as the existing 'Working...' alternation; noting the tradeoff only.
• ℹ️ bin/fm-harness.sh:3 - The usage comment at the top of bin/fm-harness.sh still enumerates the own-harness outputs as claude|codex|opencode|pi|grok|kimi|unknown, but detect_own now also returns 'cursor'. Mechanical comment-accuracy fix.

🔧 Fix: scope cursor composer-row scan to identified cursor panes
✅ Re-checked - no issues remain.

⚠️ **Test** - 1 error

• 🚨 tests failed with exit code 1
• command -v tmux >/dev/null || { echo "tmux is required for e2e tests" >&2; exit 1; }; tmux -V; rc=0; for t in tests/*.test.sh; do echo "== $t =="; bash "$t" || rc=1; done; exit "$rc"

🔧 Fix: confirm watcher-lock timeout flake; tests pass, no changes
1 error still open:

• 🚨 tests failed with exit code 1
• command -v tmux >/dev/null || { echo "tmux is required for e2e tests" >&2; exit 1; }; tmux -V; rc=0; for t in tests/*.test.sh; do echo "== $t =="; bash "$t" || rc=1; done; exit "$rc"

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.