cursor-caller
Skill0 starsChanged 3 months agoApache-2.0View on GitHub
Call Cursor CLI (binary `agent`) through the canonical RevHarness wrapper. Use when the task is light/medium and Codex-grade reasoning is overkill, or when the user explicitly asks for Cursor. Triggers include "cursor で" / "cursor agent で" / "ask モードで調べて" / "yolo で全自動". Three roles align to Cursor official mode terminology: `ask` (true read-only via --mode ask), `agent` (default write-capable agent mode), `yolo` (agent + --force command auto-approval). Cross-family delegation invariant: always invoke through `scripts/cursor-wrapper.sh`; never call `agent` directly. Safety note: per Cursor official docs, `agent -p` (default mode) has write/shell access; only `--mode ask` is a guaranteed read-only path.
In this repository78 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .agents/skills/go-skills-knowledge-pack/AGENTS.mdAGENTS.md
- .agents/skills/rust-skills-knowledge-pack/AGENTS.mdAGENTS.md
- .agents/skills/typescript-skills-knowledge-pack/AGENTS.mdAGENTS.md
- .claude/skills/go-skills-knowledge-pack/AGENTS.mdAGENTS.md
- .claude/skills/rust-skills-knowledge-pack/AGENTS.mdAGENTS.md
- .claude/skills/typescript-skills-knowledge-pack/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .cursor/rules/00-general.mdcCursor rule
- .cursor/rules/revharness-critical.mdcCursor rule
- .cursor/rules/revharness-detailed.mdcCursor rule
- auto-orchestratorSkill
- baseline-protectionSkill
- client-distribution-readinessSkill
- cloudflare-deploy-guardSkill
- codex-app-server-guardSkill
- codex-app-server-product-integrationSkill
- codex-callerSkill
- context-compactorSkill
- delegated-development-disciplineSkill
- deploy-command-disciplineSkill
- design-principlesSkill
- development-constitutionSkill
- development-junk-cleanupSkill
- go-skills-knowledge-packSkill
- harness-official-docs-updateSkill
- naming-normalization-guardSkill
- orchestrator-bootstrapSkill
- payload-cms-deploy-guardSkill
- production-function-implementerSkill
- research-handoffSkill
- rev-harness-lifecycleSkill
- review-workflowSkill
- root-cause-first-disciplineSkill
- rust-skills-knowledge-packSkill
- scope-guardSkill
- self-growth-proposal-triageSkill
- shared-client-wrapper-safetySkill
- slice-designerSkill
- staff-code-reviewerSkill
- structured-mentorSkill
- supabase-deploy-guardSkill
- system-plannerSkill
- typescript-skills-knowledge-packSkill
- auto-orchestratorSkill
- baseline-protectionSkill
- client-distribution-readinessSkill
- cloudflare-deploy-guardSkill
- codex-app-server-guardSkill
- codex-app-server-product-integrationSkill
- codex-callerSkill
- context-compactorSkill
- cursor-callerSkill
- delegated-development-disciplineSkill
- deploy-command-disciplineSkill
- design-principlesSkill
- development-constitutionSkill
- development-junk-cleanupSkill
- go-skills-knowledge-packSkill
- harness-official-docs-updateSkill
- The first 60 shown.
The file
Files
Cursor Caller
This skill is a selection hint (lens), not a workflow owner. Cursor is RevHarness's third orchestration target alongside Claude Code and Codex.
Safety model (read this first)
Per Cursor's official docs (https://cursor.com/docs/cli/reference/parameters):
agent -p(default agent mode) has access to all tools, including write and shell. It is not proposal-only.--force/--yoloauto-approve commands. They are NOT a file-write gate.--mode askis Cursor's documented true read-only mode.
The wrapper exposes 3 roles that align with these semantics. The naming intentionally avoids words like "composer" or "proposal-only" because those would misrepresent what the underlying CLI does.
Roles
| Role | wrapper passes to agent | Effective behavior |
ask (default) | -p --output-format text --mode ask | True read-only. No file writes. No shell execution. |
agent | -p --output-format text | Default agent mode. Write + shell tool access. Cursor still prompts per command (no --force). |
yolo | -p --output-format text --force | Write + shell + command auto-approval. Maximum automation, minimum safety. CI / mechanical execution only. |
ask is default because it is the only role that gives a hard read-only guarantee. If you want writes, opt in explicitly.
When to pick Cursor over Codex
| Scenario | Pick Cursor | Pick Codex |
| Repo Q&A / "explain this file" / search-style | ask | (research) |
| Single-function rewrite < 50 LOC, no security concerns, no deterministic-check requirement | agent | — |
| Single-function rewrite where Reviewer LGTM / deterministic checks are needed | — | coder + production-function-implementer |
| Mechanical bulk transform (CI-driven, well-scoped) | yolo | — |
| Production function with input validation, error handling, tests | — | high-coder + production-function-implementer |
| Security-sensitive surface | — | high-coder |
| Code review with verdict | — | reviewer + staff-code-reviewer |
Rule of thumb: if you want a hard read-only guarantee, use ask. If you want Codex-grade contract compliance (envelope schemas, deterministic verification), use Codex.
How to invoke
Canonical entrypoint: scripts/cursor-wrapper.sh. Direct agent invocation is forbidden under RevHarness governance.
Read-only (default)
cat prompt.md | ./scripts/cursor-wrapper.sh --role ask --stdin > answer.mdThe wrapper appends --mode ask to the agent argv; Cursor will not modify files.
Standard agent (writes allowed)
cat prompt.md | ./scripts/cursor-wrapper.sh --role agent --stdin > output.mdNo --mode flag is passed. Cursor runs in default agent mode and will prompt before potentially destructive commands.
Full automation (command auto-approval)
cat prompt.md | ./scripts/cursor-wrapper.sh --role yolo --stdin > output.md--force is passed. Commands run without per-command approval. Use only in CI or mechanical workflows where you accept the risk.
Dry-run (0-cost validation)
./scripts/cursor-wrapper.sh --role ask --dry-runResolves role, emits the delegation metric, but does not invoke agent.
Delegation metric (JSONL on stderr)
REV_HARNESS_DELEGATION_METRIC {"schema_version":1,"vendor":"cursor","wrapper_role":"cursor-ask",...,"cursor_mode":"ask","cursor_force_flag":""}Cursor-specific fields:
vendor:"cursor"(distinguishes from Codex/Claude metrics)cursor_mode:"ask"for read-only role, empty otherwisecursor_force_flag:"--force"foryolorole, empty otherwise
Important: cursor_force_flag records command approval behavior (auto-approve), not file-write permission. File writes are gated by cursor_mode == "ask" only.
Fail-closed conditions
agentbinary not found at$HOME/.local/bin/agent,/usr/local/bin/agent, orPATH.- Unknown role (
ask | agent | yoloonly). --rolespecified more than once (role escape rejected).--stdinomitted on non-dry-run invocation.- Vendoring guard trip from non-canonical root.
What this skill is NOT for
- Heavy implementation with RevHarness role contracts → Codex coder/high-coder/reviewer
- Reviewer LGTM gate →
scripts/codex-wrapper.sh --role reviewer(Codex 固定) - Anything requiring specialty manifests (production-function-implementer, staff-code-reviewer, etc.) — Cursor wrapper does not implement the specialty system in round 2.
Pointers
- Wrapper source:
scripts/cursor-wrapper.sh - Tests:
test/unit/test-cursor-wrapper.sh(24 passed, includes argv assertions + signal trap + vendoring guard regression) - Fake fixture:
test/fixtures/fake-cursor/agent - Cursor official docs (always primary source of truth):
- Auth: handled by the user out-of-band (
agent logininteractive).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

