cli-cursor

Cursor CLI Orchestrator - Cross-AI Task Delegation

CRITICAL — SELF-INVOCATION PROHIBITED

This skill dispatches to the Cursor CLI binary (cursor-agent). If the agent currently reading this skill is itself running inside cursor-agent (detection signals listed in §2), the skill MUST refuse to load and return the documented error message instead of generating any cursor-agent invocation.

A running CLI skill never dispatches itself. The cli-X skills are for cross-AI delegation only — never self-invocation.

Orchestrate Cursor's terminal coding agent (cursor-agent) for tasks that benefit from a second AI perspective, Cursor's native model (Composer), read-only plan/ask exploration, or its unique worktree/cloud-worker/plugin surfaces.

Core Principle: Use Cursor for what it does best. Delegate, validate, integrate. The calling AI stays the conductor.

1. WHEN TO USE

Activation Triggers

• Cross-AI Validation — code review second perspective, bug detection, an independent implementation attempt.
• Composer Dispatch — tasks that specifically want Cursor's own native model (composer-2.5/composer-2.5-fast), or the Grok (4.5/4.6) / GLM 5.2 tiers already on the enforced allowlist (see §3 Model Selection).
• Read-Only Exploration — --mode plan (read-only planning) or --mode ask (read-only Q&A) when a task wants Cursor's analysis without any file writes.
• Isolated Experimentation — Cursor's native git worktree isolation (-w/--worktree) for a change the operator wants tried in a disposable checkout, documented in references/cursor-tools.md as an opt-in escape hatch, not this packet's default dispatch shape.
• Specialized Generation — explicit Cursor requests naming Grok (4.5 or 4.6) or GLM 5.2 specifically (the only non-Composer models on the enforced allowlist).

When NOT to Use

• You ARE Cursor already. If your runtime is cursor-agent (detection signal: CURSOR_AGENT=1 set, CURSOR_CONVERSATION_ID set, cursor-agent in process ancestry, or an active-session probe under ~/.cursor/), this skill refuses to load. Self-invocation creates a circular dispatch loop and burns tokens for no value.
• Simple, quick tasks where CLI overhead is not worth it, or tasks needing an immediate response (auth/network round-trips may delay).
• Context already loaded and understood by the current agent.
• Interactive refinement requiring the full-screen TUI (use cursor-agent directly instead).
• Tasks where Cursor CLI is not installed, or the account is not authenticated (cursor-agent about reports "Not logged in").

2. SMART ROUTING

Prerequisite Detection

# Verify Cursor CLI is available before routing
command -v cursor-agent || echo "Not installed. Run: curl https://cursor.com/install -fsS | bash"

Self-Invocation Guard

def detect_self_invocation():
    """Returns a non-None signal when the orchestrator is already running inside cursor-agent."""
    # Layer 1: env var lookup — CURSOR_AGENT=1 is set unconditionally whenever the
    # current process runs under cursor-agent (confirmed live); CURSOR_CONVERSATION_ID
    # is the confirmed session-id marker (matches --output-format json's session_id).
    if os.environ.get('CURSOR_AGENT') == '1':
        return ('env', 'CURSOR_AGENT')
    if os.environ.get('CURSOR_CONVERSATION_ID'):
        return ('env', 'CURSOR_CONVERSATION_ID')
    # Layer 2: process ancestry — cursor-agent (the canonical binary, never the
    # bare "agent" alias, which would false-positive on unrelated commands) in
    # the parent tree.
    try:
        ancestry = subprocess.check_output(['ps', '-o', 'command=', '-p', str(os.getppid())]).decode()
        if '/cursor-agent' in ancestry or 'cursor-agent ' in ancestry:
            return ('ancestry', 'cursor-agent')
    except subprocess.SubprocessError:
        pass
    # Layer 3: best-effort session probe. No lock-file convention is documented
    # for Cursor CLI (unlike Codex's ~/.codex/state/<id>/lock) — this layer is
    # honestly weaker than the siblings' layer 3. Absence of a detected signal
    # here is NOT proof no session is active.
    state_dir = os.path.expanduser('~/.cursor')
    if os.path.isdir(state_dir) and os.environ.get('CURSOR_INVOKED_AS'):
        return ('state-probe', 'CURSOR_INVOKED_AS')
    return None

if detect_self_invocation():
    refuse(
        "Self-invocation refused: this agent is already running inside Cursor CLI. "
        "Use a sibling cli-* skill or a fresh shell session in a different runtime to dispatch a different model."
    )

Resource Loading Levels

Smart Router

Provider-specific dictionaries (used by the shared helper functions in system-spec-kit/references/cli/shared-smart-router.md):

INTENT_SIGNALS = {
    "GENERATION":        {"weight": 4, "keywords": ["generate", "create", "build", "write code", "cursor create"]},
    "REVIEW":            {"weight": 4, "keywords": ["review", "audit", "bug", "second opinion", "cross-validate"]},
    "COMPOSER":          {"weight": 4, "keywords": ["composer", "cursor's own model", "auto router", "cursor native model"]},
    "PLAN_ASK":          {"weight": 4, "keywords": ["plan mode", "ask mode", "read-only", "just explore", "don't write anything"]},
    "AGENT_DELEGATION":  {"weight": 4, "keywords": ["delegate", "agent", "background", "cursor agent", "subagent"]},
    "TEMPLATES":         {"weight": 3, "keywords": ["template", "prompt", "how to ask", "cursor prompt"]},
    "PATTERNS":          {"weight": 3, "keywords": ["pattern", "workflow", "orchestrate", "session", "resume", "continue"]},
    "HOOKS":             {"weight": 4, "keywords": ["hook", "hooks", "hooks.json", "startup context", "sessionstart", "beforesubmitprompt"]},
    "UNIQUE_SURFACES":   {"weight": 3, "keywords": ["worktree", "cloud worker", "plugin marketplace", "shared config", "editor config"]},
    # WHY: DESIGN is an intent signal only. The durable sk-design-md-generator loading contract lives in the
    # always-fires Design Standards Loading rule and the dispatch manifest; RESOURCE_MAP stays
    # limited to same-skill markdown paths.
    "DESIGN":            {"weight": 4, "keywords": ["sk-design-md-generator", "extract design system", "generate design.md", "style reference", "design tokens", "css extraction", "tokens.json"]},
}

RESOURCE_MAP = {
    "GENERATION":        ["references/cli-reference.md", "assets/prompt-templates.md"],
    "REVIEW":            ["references/integration-patterns.md", "references/agent-delegation.md"],
    "COMPOSER":          ["references/cli-reference.md", "assets/prompt-quality-card.md"],
    "PLAN_ASK":          ["references/cli-reference.md", "references/agent-delegation.md"],
    "AGENT_DELEGATION":  ["references/agent-delegation.md", "references/integration-patterns.md"],
    "TEMPLATES":         ["assets/prompt-templates.md", "references/cli-reference.md"],
    "PATTERNS":          ["references/integration-patterns.md", "references/cli-reference.md"],
    "HOOKS":             ["references/hook-contract.md", "references/shared-editor-config.md"],
    "UNIQUE_SURFACES":   ["references/cursor-tools.md", "references/shared-editor-config.md"],
}

LOADING_LEVELS = {
    "ALWAYS": ["references/cli-reference.md", "assets/prompt-quality-card.md"],
    "ON_DEMAND_KEYWORDS": ["full reference", "all templates", "deep dive", "worktree", "cloud worker", "plugin marketplace", "hook contract", "shared config"],
    "ON_DEMAND": ["references/cursor-tools.md", "references/shared-editor-config.md", "assets/prompt-templates.md"],
}

UNKNOWN_FALLBACK_CHECKLIST = [
    "Is the user asking about Cursor CLI specifically?",
    "Does the task benefit from a second AI perspective or Cursor's own Composer model?",
    "Does the task want read-only plan/ask exploration instead of a write-capable dispatch?",
    "Would Cursor's worktree isolation or cloud worker genuinely help (opt-in, not default)?",
]

Call sequence (using shared helpers from shared-smart-router.md):

1. discover_markdown_resources() — recursively enumerate current .md files under existing references/ and assets/ folders at routing time.
2. _guard_in_skill() + load_if_available() — sandbox paths to this skill, reject non-markdown loads, skip missing files, and suppress duplicates.
3. score_intents(task) and select_intents(scores, ambiguity_delta=1.0) — preserve provider-specific weighted intent scoring and top-2 ambiguity handling.
4. get_routing_key(task, intents) — derive the provider routing key from task/provider context, then fall back to cursor-agent.
5. ALWAYS-load LOADING_LEVELS["ALWAYS"], then return UNKNOWN_FALLBACK with UNKNOWN_FALLBACK_CHECKLIST when max score is 0.
6. CONDITIONAL-load RESOURCE_MAP[intent], ON_DEMAND-load keyword matches, and return a notice when no provider-specific knowledge base is available beyond always-load resources.

The route_cursor_resources(task) function body lives in shared-smart-router.md — substitute <PROVIDER> = cursor.

3. HOW IT WORKS

Prerequisites

Install with curl https://cursor.com/install -fsS | bash (Windows: irm 'https://cursor.com/install?win32=true' | iex). cli-cursor authenticates through Cursor account OAuth — run cursor-agent login and complete the browser flow (NO_OPEN_BROWSER disables browser opening for headless hosts). Headless/CI auth uses CURSOR_API_KEY or --api-key. Full install, auth, flag, hook, and unique-surface tables live in the ALWAYS-loaded cli-reference.md — this section keeps only the routing decisions and dispatch-critical gotchas.

Execution Ownership

This packet owns user-facing routing, the command -v cursor-agent availability probe, prompt construction, and the self-invocation guard. Actual process construction and execution delegate to the already-shipped deep-loop runtime at ../../system-deep-loop/runtime/scripts/fanout-run.cjs, using executor kind cli-cursor.

The runtime is the single Cursor execution adapter. Do not add a packet-local wrapper, command builder, or spawn path. Direct cursor-agent -p snippets below are operator reference and manual-testing examples; orchestrated dispatches use the shared runtime.

Provider Auth Pre-Flight

MANDATORY before any first dispatch in a session. cursor-agent -p without a valid session exits 0 even on an auth failure — the exit code is never a reliable availability signal. Run this check once per session, cache the result, and re-run it only if a dispatch's output text (not its exit code) shows an authentication error.

# One-shot pre-flight: capture account auth status for routing
CURSOR_ABOUT=$(cursor-agent about 2>&1)
echo "$CURSOR_ABOUT" | grep -qi "not logged in" && CURSOR_AUTH_OK=0 || CURSOR_AUTH_OK=1

Decision tree (apply in order — first match wins):

User prompt template — not logged in:

Cursor CLI is not authenticated on this machine. cli-cursor uses Cursor account OAuth
(or CURSOR_API_KEY for headless auth). Run `cursor-agent login`
(browser flow), then confirm when login finishes — the skill will retry the original dispatch.

Error-recovery contract. If a dispatch's output text shows an auth error after pre-flight passed (session expired or revoked), invalidate the cache, re-run the pre-flight, and re-check before retrying. Never substitute a model the user didn't approve.

Default Invocation (Skill Default)

Default model + approval: composer-2.5 (Cursor's own native model) · auto-review approval (Smart Auto). auto (Cursor's own router) is NOT used — it can silently resolve to a model outside the enforced allowlist below, which defeats the point of enforcing one.

cursor-agent -p "<prompt>" \
  --output-format text \
  --model composer-2.5 \
  --auto-review \
  --sandbox enabled

User override (honor explicit user phrasing verbatim, but ONLY within the enforced allowlist — see Model Selection below):

Honor whichever dimensions the user names (approval level, mode). Model stays on composer-2.5 unless the user explicitly names a different ALLOWED model.

Model Selection — Enforced Allowlist

cli-cursor dispatch is scoped to exactly 21 ids — never dispatch a model outside the allowlist (including auto), and never substitute the closest-sounding allowed model without telling the user. Default composer-2.5; pick a Gemini 3.7 Flash High / GLM 5.2 / GPT-5.6 Luna Max / Grok (4.5 or 4.6) tier only when the task or user explicitly names that family. Cursor has no --reasoning-effort flag and rejects parameterized model brackets outright — effort tiers must be selected through exact enumerated ids.

The enforced allowlist (21 ids) and the per-task rationale table live inline in references/cli-reference.md §5 and references/providers-and-models.md §2. Enforced at the runtime layer (CURSOR_SUPPORTED_MODELS in executor-config.ts, checked by fanout-run.cjs and dispatch-model.cjs before any command is constructed) and at this skill layer. If a task seems to need a model outside the allowlist, escalate to the user rather than fabricating a substitute or falling back to auto.

Cursor Agent Delegation

The calling AI is the conductor; Cursor's own skill system at ~/.cursor/skills-cursor/ (observed live: automate, babysit, canvas, create-hook, create-rule, create-skill, create-subagent, loop, migrate-to-skills, sdk, shell, split-to-prs, statusline, update-cli-config, update-cursor-settings) confirms Cursor supports subagents natively, but these are Cursor-editor-side conventions, not a -p <profile> flag like Codex's. Full delegation contract and the --mode plan/--mode ask/default-agent execution-mode roster: agent-delegation.md.

Repository Rules, Hook Delivery, and Parity Boundaries

Cursor CLI reads project .cursor/rules/*.md, root AGENTS.md, root CLAUDE.md, and legacy .cursorrules automatically. This repo now uses .cursor/rules/skill-routing.md as a compact, always-on pointer to the relevant top-level .opencode/skills/*/SKILL.md packets. It is static session context, not a replacement for dynamic per-turn classification.

The beforeSubmitPrompt adapter is designed to deliver a dynamic skill-advisor-equivalent brief, but delivery is confirmed dormant under the installed Cursor CLI build. Its source marks the status as registered but unconfirmed and records the shared-advisor delegation in user-prompt-submit.ts and user-prompt-submit.ts. A live marker re-probe against cursor-agent 2026.07.23-e383d2b confirmed that beforeSubmitPrompt did not fire. The static rules file therefore complements a missing dynamic brief; it does not claim to provide per-turn advisor output. The hook registration and adapter remain unchanged.

Custom Subagents (CORRECTION -- earlier claim was wrong)

An earlier pass recorded that "cursor-agent --help has no custom-agent-loading concept." That was wrong, and the error is instructive: agent profiles are discovered by file convention, never by a CLI flag, so grepping --help for a flag proves nothing about whether the concept exists.

Cursor CLI loads custom subagents from two places, confirmed live:

Cursor's own bundled create-subagent skill documents the format: name + description frontmatter (both required), markdown body as the system prompt. A live cursor-agent --force -p roster probe lists all 13 repo agents (ai-council, code, context, debug, deep-alignment, deep-improvement, deep-research, deep-review, design, markdown, orchestrate, prompt-improver, review) alongside Cursor's own built-ins, with no duplicate entries when both sources define the same name.

Each .cursor/agents/<name>.md is a symlink to the canonical .claude/agents/<name>.md, matching the discovery-mirror precedent already used for .claude/hooks/ and .codex/hooks/. One source of truth, so a mirror can never drift from the agent it mirrors.

Note: ~/.cursor/agents/ (user-level) is documented by Cursor but a live probe found the CLI did not load a profile placed there; only the project-level paths above are verified working. Dispatch remains subject to this repo's own preToolUse spec-gate, which is correct behavior.

cursor-agent --help has no native command-roster listing subcommand. This repository still exposes a .cursor/commands/ parity surface: the synchronizer derives it from eligible .opencode/commands/ files, excludes goal-opencode.md, and adds the runtime-native goal-cursor.md; validate membership with command-scope.cjs and validate behavior with a real slash-command invocation.

Dispatch-Critical Gotchas

The full flag glossary, hook contract, shared-config surface, and troubleshooting table are in the ALWAYS-loaded cli-reference.md. Gotchas that silently break a dispatch and must be honored at routing time:

• The exit code is never an availability signal. cursor-agent -p without auth exits 0 and prints an error to stdout/stderr instead. Every guard and pre-flight in this packet checks output text, never exit code.
• The canonical binary is cursor-agent, never the bare agent alias. agent is a symlink to the same binary; using it in a process-ancestry match risks colliding with an unrelated agent command.
• Cursor shares its entire config surface with the Cursor editor (.cursor//~/.cursor/: mcp.json, hooks.json, rules/, cli-config.json). A dispatched cursor-agent silently inherits the operator's shared hooks/MCP/rules unless a workspace/config-isolation flag is used — see shared-editor-config.md.
• No model[effort=...] bracket support. Unlike some sibling CLIs' parameterized model syntax, cursor-agent --model 'cursor-grok-4.6[effort=high]' is rejected outright ("Cannot use this model") — effort tiers must be selected via an exact enumerated id (cursor-grok-4.6-high), never a bracket.
• --auto-review/--force are the write-capable escalation, not --sandbox. --sandbox enabled|disabled toggles the OS-level sandbox; the approval decision (whether unattended actions run without a human) is --auto-review (Smart Auto) or --force/--yolo (Run Everything) — omitting both leaves Cursor's own prompt-and-block default in place, which cannot proceed unattended.
• Project-scoped MCP servers show not loaded (needs approval) until trusted. For a non-interactive dispatch that needs MCP tools, add --approve-mcps to the cursor-agent -p command (auto-approves configured MCP servers for that run); for a persistent operator grant use cursor-agent mcp enable <server> (a trust mutation the automation itself must never run).

4. RULES

✅ ALWAYS

1. Verify Cursor CLI is installed before first invocation (command -v cursor-agent).

2. Delegate orchestrated execution to ../../system-deep-loop/runtime/scripts/fanout-run.cjs with executor kind cli-cursor; never build a second adapter in this packet.

3. Use --mode plan or --mode ask for read-only exploration/analysis/research; use the default agent mode with --auto-review or --force for code generation/file modification.

4. Validate Cursor-generated code (XSS, injection, eval, syntax checks via node --check, tsc --noEmit, etc.) before applying.

5. Capture stderr (2>&1) so errors surface; check output TEXT for auth/availability failures, never the exit code (always 0).

6. Redirect cursor-agent stdin from /dev/null when dispatching in a while read loop, mirroring the family-wide convention: cursor-agent -p "$PROMPT" > "$LOG" 2>&1 </dev/null &. Live-verified: a real cursor-agent -p ... </dev/null dispatch completes normally with no hang.

7. Specify model and approval mode explicitly — never rely on caller environment. Default: --model composer-2.5 --auto-review --sandbox enabled. Honor user overrides verbatim, but ONLY within the enforced allowlist (§3 Model Selection) — never auto, never a model outside the 21 allowed ids.

8. Route to --mode plan/--mode ask/default agent per the task type (see Section 3 routing table).

9. Pass the spec folder to the delegated agent in the prompt: if the calling AI has an active Gate-3 spec folder, include Spec folder: <path> (pre-approved, skip Gate 3). If none, ASK the user before delegating — the delegated agent cannot answer Gate 3 in --force/non-interactive mode.

10. Prompt construction & model-craft (cli- family precedence).* Compose every dispatch prompt via the 3-tier rule canonical in ../../sk-prompt/sk-prompt-models/assets/cli-prompt-quality-card.md:

11. Fast path (default). Build from the local assets/prompt-quality-card.md, which delegates the framework table + CLEAR check to the canonical card.

12. Model override (mandatory for a profiled model). If the target model has a profile at ../../sk-prompt/sk-prompt-models/references/models/<id>.md, that profile OVERRIDES the cross-model default.

13. Deep path (escalation). Dispatch @prompt-improver via the Task tool (never load full sk-prompt inline) when any canonical Tier 3 trigger applies — the trigger list lives in ../../sk-prompt/sk-prompt-models/assets/cli-prompt-quality-card.md under "Tier 3 — Deep path"; do not re-enumerate it here.

14. Never inject user-level voice/personalization content into AI-orchestrated Cursor delegations. Cursor CLI reads rules from .cursor/rules, AGENTS.md, CLAUDE.md, and legacy .cursorrules automatically. When an AI delegates via cursor-agent -p, the calling AI's own voice rules govern the response — keep delegations focused on task/model/approval/(spec-folder pre-approval).

15. Code Standards Loading (surface-aware contract) — When dispatching for code review or code generation, instruct the dispatched session to: (1) load sk-code; (2) let sk-code emit a surface tag matching the detected stack from markers and target files; (3) load the selected surface resources and run its verification commands; (4) add code-review only for formal findings-first review output. Fallback: if the surface cannot be determined confidently, ask for the runtime surface and verification command set. NEVER hardcode obsolete sibling code skills in dispatch prompts.

16. Design Standards Loading (measured-reference contract) — When dispatching for design or UI work, instruct the dispatched session to: (1) load sk-design-md-generator; (2) extract a measured Style Reference DESIGN.md (named color tokens, type scale, components, Quick-Start CSS/Tailwind) from the live source before building UI; (3) build against those measured tokens and run the extraction's validate step to confirm hex/section fidelity. Fallback: if there is no live source to measure, ask for the reference URL or the exact tokens to build against. NEVER treat mcp-figma or sk-design-md-generator as a taste, visual-direction, or critique authority — the extraction measures real CSS, it does not judge design.

17. Pass the design reference manifest to the dispatched session — when dispatching design or UI work, inline a DESIGN_DISPATCH_MANIFEST v1 block in the prompt (the child cannot resolve skill paths, so the manifest travels in the payload, not by reference): styleReferenceExtracted true, the live source that was measured, the measured design tokens / type scale / components the child must build against, loadedFiles, and proofDemandBack. If the manifest cannot be assembled — no Style Reference extracted, or no live source to measure — ASK before launching the child rather than starting a silent design dispatch. The child returns the demanded proof; the parent reconciles it on the return path.

18. Single-dispatch discipline (operator-gated, session-scoped) — Default: launch ONE cli-* dispatch at a time across the cli-* family. Wait for the dispatched agent's work to return, verify outputs exist, then SIGKILL only the dispatch THIS skill started: capture its PID at launch (cursor-agent -p ... & CURSOR_PID=$!) and kill that captured PID directly plus its own orphan children (kill -9 "$CURSOR_PID" 2>/dev/null; pkill -9 -P "$CURSOR_PID" 2>/dev/null). Never use a blanket pkill -9 -f "cursor-agent" pattern — that matches and kills EVERY running cursor-agent process on the machine, including the operator's unrelated Cursor sessions (and the operator's own Cursor editor, since the CLI shares the binary family). Only launch the next dispatch (this skill OR a sibling) after the prior one is dead and RSS has dropped. Within a deep-flow session (deep-review / deep-research): the operator authorizes the whole multi-iteration session at start — iterations chain back-to-back with kill-between as the safety mechanism, NOT a per-iteration confirmation prompt. Exception (cross-skill parallel): when the operator explicitly authorizes N parallel dispatches, run N concurrently — but still SIGKILL each by its own captured PID as its work returns.

19. Set AI_SESSION_CHILD=1 in the dispatched child's env when sessions may be launched through the per-session worktree wrapper (.opencode/bin/worktree-session.sh). A dispatched cursor-agent -p is an orchestrated sub-session, not a new top-level session, so it must SHARE the parent's worktree rather than allocate its own (distinct from Cursor's own native -w worktree flag, which this packet's default dispatch never passes). Pattern: AI_SESSION_CHILD=1 cursor-agent -p ... </dev/null. Harmless when the wrapper is not in use. See .opencode/bin/README.md → "Worktree session isolation".

20. Agent-persona injection (attach identity, not just the task). Every dispatch composes {resolved agent persona + task prompt} — never a bare task. Resolve the persona from the ACTIVE runtime's agent directory per AGENTS.md §7 (.cursor/agents/<name>.md, .claude/agents/<name>.md, .opencode/agents/<name>.md, etc. — never hardcode one runtime), and map each subtask to the RIGHT agent (code→code, review→review, design→design, research→deep-research, docs→markdown), not one default. Cursor has a native file-convention subagent surface: .cursor/agents/*.md + .claude/agents/*.md mirror all 13 canonical agents — dispatch by naming the resolved subagent, which satisfies the rule. On a bare cursor-agent -p that names no subagent, INLINE the persona block into the payload using the same in-payload pattern as the DESIGN_DISPATCH_MANIFEST (Rule 14) — the child cannot resolve agent paths by reference. A persona-less dispatch runs the leaf as a generic assistant, silently dropping the agent's tool-scope, verification gates, and output contract. Canonical contract: ../../sk-prompt/sk-prompt-models/assets/cli-prompt-quality-card.md "Persona Injection"; native precedent: orchestrate.md "Agent Loading Protocol". Rare exceptions (native surface used, focused summary for a small-context model, pure-mechanical command) are declared at the dispatch site.

⛔ NEVER

1. Pass -w/--worktree or dispatch cursor-agent worker from orchestrated fan-out without explicit user approval — these are opt-in escape hatches documented in references/cursor-tools.md, not this packet's default dispatch behavior.
2. Trust Cursor output blindly for security-sensitive code, send sensitive data (API keys, passwords, credentials) in prompts, or hammer the CLI with rapid sequential calls.
3. Use Cursor for tasks where context is already loaded — direct action by the calling AI is faster.
4. Assume Cursor output is correct without verification — cross-reference codebase and project standards.
5. Build or maintain a packet-local Cursor execution adapter; the deep-loop runtime is the execution authority.
6. Treat a 0 exit code as proof of a successful, authenticated dispatch — always inspect output text.
7. Dispatch a --model value outside the enforced 21-id allowlist (§3 Model Selection) — including auto — or silently substitute the closest-sounding allowed model instead of asking the user. Enforced at the runtime layer (CURSOR_SUPPORTED_MODELS/isCursorModelAllowed in executor-config.ts; a hard-rejecting check in fanout-run.cjs's buildCursorLineageCommand and dispatch-model.cjs's cli-cursor case) — this rule states the same constraint for any advisory/manual dispatch the runtime layer cannot see.

⚠️ ESCALATE IF

1. Cursor CLI is not installed and user has not acknowledged (provide curl https://cursor.com/install -fsS | bash).
2. cursor-agent about reports "Not logged in" and the user has not acknowledged running cursor-agent login.
3. Cursor output conflicts with existing code patterns (present both perspectives; user decides).
4. Task requests cursor-agent worker (cloud execution) or native worktree isolation (-w) as part of an orchestrated fan-out (describe the scope-creep risk; get explicit user approval).

Memory Handback Protocol

When the calling AI needs to preserve session context from a Cursor CLI delegation, run the canonical 7-step procedure (extract MEMORY_HANDBACK section → build structured JSON → scrub secrets → invoke generate-context.js via --stdin/--json/temp-file → memory_index_scan). Full procedure and caveats: system-spec-kit/references/cli/memory-handback.md. Cursor-specific Memory Epilogue template: assets/prompt-templates.md.

printf '%s' "$JSON_PAYLOAD" | node .opencode/skills/system-spec-kit/scripts/dist/memory/generate-context.js --stdin [spec-folder]

5. REFERENCES

Core References

• cli-reference.md - Complete CLI subcommands, flags, auth, and troubleshooting reference
• integration-patterns.md - Cross-AI orchestration patterns and workflows
• cursor-tools.md - Cursor-unique surfaces: native worktree (-w), cloud worker, plugin marketplace, MCP
• hook-contract.md - Cursor's shared hooks.json contract (events, schema, discovery, envelope)
• shared-editor-config.md - The shared .cursor//~/.cursor/ editor-config surface and dispatch-isolation implications
• agent-delegation.md - Cursor execution-mode roster (default/plan/ask) and delegation patterns

Templates and Assets

• prompt-templates.md - Copy-paste ready prompt templates for common tasks
• prompt-quality-card.md - Fast-path prompt framework + CLEAR check (ALWAYS loaded)

Manual Testing

• manual-testing-playbook.md - 19 CU-NNN scenarios across 9 categories; the canonical no-mocking PASS/FAIL/SKIP validation gate for this skill

Shared (cli-* family)

• shared-smart-router.md - Helper-function bodies for the smart router.
• memory-handback.md - Canonical 7-step Memory Handback procedure.

External

• Cursor CLI docs - Official documentation
• Cursor account - Cursor account (auth for cli-cursor)

Reference Loading Notes

• Load only references needed for current intent; Smart Routing (Section 2) is the single routing authority.
• cli-reference.md is ALWAYS loaded as baseline.

6. SUCCESS CRITERIA

Task Completion

• Cursor CLI invoked with correct subcommand, flags, model, and approval mode.
• Output captured, validated, and integrated appropriately; no security vulnerabilities introduced.
• Auth failures handled gracefully (text-based detection, never exit-code-based).
• Appropriate execution mode matched to task type (--mode ask/--mode plan for read-only, default agent for generation).
• Orchestrated execution delegated to the shared deep-loop runtime without a packet-local adapter.

Skill Quality

• All 8 sections present with proper anchor comments.
• Smart routing covers all intent signals with UNKNOWN_FALLBACK.
• Reference files provide deep-dive content without duplication.

7. INTEGRATION POINTS

Framework Integration

This skill operates within the behavioral framework defined in AGENTS.md.

Key integrations:

• Gate 2: Skill routing via skill_advisor.py
• Tool Routing: Per AGENTS.md Section 6 decision tree
• Memory: Context preserved via Spec Kit Memory MCP
• Execution: Shared deep-loop runtime (../../system-deep-loop/runtime/scripts/fanout-run.cjs)

Tool roles: Bash dispatches the CLI; Read/Glob/Grep validate output.

8. REFERENCES AND RELATED RESOURCES

The router discovers reference, asset, and script docs dynamically (Section 5 is the authored index). Start with the ALWAYS-loaded references/cli-reference.md and assets/prompt-quality-card.md, then load task-specific resources per Smart Routing.

Related skills: cli-codex for OpenAI-backed delegation, cli-claude-code for extended reasoning, cli-opencode for full OpenCode runtime dispatch, sk-code for code-quality contracts, mcp-code-mode for external MCP work, and system-spec-kit for packet handback.