Skills for working across Claude Code and Cursor agent.

A fork of chaseai-yt/claudex-loop that swaps Codex for the Cursor agent CLI (agent / cursor-agent). One provider plans; the other independently reviews the plan and inspects the final code.

Skill	Use it for	Dependencies
claudor-route	A model recommendation or one scoped handoff	Self-contained; selected CLI needed only for delegation
claudor-loop	Requirements, plan review, implementation, and final inspection	Both CLIs and Python 3.10+
cursor-review	Explicit Cursor plan-review command	Shared claudor-loop skill
cursor-build	Explicit Cursor builder command	Shared claudor-loop skill

Claudor Loop

Start here	Requirements and plan	Plan review	Default build	Final inspection
Claude Code	Current Claude session	Cursor agent	Claude	Fresh Cursor agent session
Cursor	Current Cursor session	Claude	Cursor	Fresh Claude session

Choose either builder with builder=claude or builder=cursor. The inspector is always the other provider. If the coordinator takes over fixes, those edits need another independent inspection.

flowchart LR
    R[Host scouts code and context] --> Q[Settle material decisions]
    Q --> P[Write plan and acceptance checks]
    P --> V[Other provider reviews]
    V -- Revise --> A[Host arbitrates and updates]
    A --> V
    V -- Approved --> B[Authorized builder implements]
    B --> T[Run proof checks]
    T --> I[Other provider inspects final code]
    I --> H[Present diff and remaining findings]

Loading

PLAN.md records what to build; PLAN-REVIEW-LOG.md records findings, dispositions, models, proof and remaining uncertainty. Both paths are configurable. CLI diagnostics live in a unique directory outside the target checkout.

Cursor models

Cursor can run many models, including Anthropic ones, so claudor-loop guards the "independent" part:

• Default: composer-2.5 for every role. It is in Cursor's included usage pool (as are Grok 4.5–4.7); most other models draw paid usage. Live, Composer found the seeded defects in ~1 minute per round. Grok 4.7 High caught subtler bugs but took 5–12 minutes per round on a tiny fixture and up to 45 minutes on a real review (see VALIDATION.md).
• Configurable per role (see configuration), or per run with reviewer_model=, builder_model=, inspector_model=.
• Effort is part of the Cursor model id (grok-4.7-medium, grok-4.7-xhigh); *_effort applies only to Claude.
• Anthropic models are refused on the Cursor side, before launch and again if Cursor reports one: Claude reviewing Claude is not a second opinion.
• auto hides which model actually ran. It requires --allow-unverified-model, and results are recorded as independence: unverified.

Every result records the requested model, where it came from (argument, env, default) and the model Cursor reported running.

Install

1. Prerequisites

Requirement	Install	Check
Python 3.10+	system package manager	python3 --version
Claude Code	docs	claude --version, claude auth status
Cursor agent CLI	curl https://cursor.com/install -fsS | bash, then agent login	agent --version, agent status

The runner launches the CLI (cursor-agent or agent on PATH), not the Cursor IDE. No pip packages or API keys are needed at runtime.

2a. Claude Code (plugin)

From GitHub:

/plugin marketplace add okkindel/ClaudorLoop
/plugin install claudor-loop@claudor-loop

Or from a local checkout (for development):

/plugin marketplace add /path/to/claudor-loop
/plugin install claudor-loop@claudor-loop

Restart Claude Code, then use /claudor-loop:claudor-loop, /claudor-loop:claudor-route, /claudor-loop:cursor-review or /claudor-loop:cursor-build. Update with /plugin marketplace update claudor-loop.

2b. Cursor

Load the checkout as a plugin for one session:

agent --plugin-dir /path/to/claudor-loop

Or copy the skills so both hosts always see them:

# run from this repository
mkdir -p ~/.cursor/skills ~/.claude/skills
cp -R skills/. ~/.cursor/skills/
cp -R skills/. ~/.claude/skills/

Copy all skill directories together: the compatibility commands share the runtime inside claudor-loop. claudor-route can be installed on its own. Updates are git pull and re-copy.

3. Optional configuration

# Cursor model per role; ids from `agent models`. Unset roles fall back to
# CLAUDOR_CURSOR_MODEL, then composer-2.5.
export CLAUDOR_CURSOR_REVIEW_MODEL=composer-2.5    # plan-review rounds
export CLAUDOR_CURSOR_BUILD_MODEL=composer-2.5     # Cursor as builder
export CLAUDOR_CURSOR_INSPECT_MODEL=grok-4.7-high  # final code inspection (slow, thorough)
export CLAUDOR_CURSOR_MODEL=grok-4.7-medium        # shared fallback for every role

Instead of a shell profile you can put them in ~/.claude/settings.json, so every Claude Code session picks them up:

{ "env": { "CLAUDOR_CURSOR_INSPECT_MODEL": "grok-4.7-high" } }

A per-run argument (inspector_model=...) beats the role variable, which beats CLAUDOR_CURSOR_MODEL. Each result records the model used and where it came from (model_source).

When Cursor is the builder, it runs without --force, so shell commands such as tests run only if allowed in ~/.cursor/cli-config.json, e.g. "permissions": {"allow": ["Shell(npm)", "Shell(python3)"]}. Otherwise the builder reports them as refused and the host runs the proof itself.

Cursor reviews refuse to start while Cursor MCP servers are configured (agent mcp list), because the reviewer's shell/write deny rules do not cover MCP tools. Remove them or pass --allow-cursor-mcp knowingly.

4. Smoke test

python3 skills/claudor-loop/scripts/runner.py roles --host claude

prints the role table (Claude plans and builds, Cursor reviews and inspects) without calling a model.

Examples

claudor this feature — plan and implement it
claudor this plan, mode=review, plan=docs/migration.md, rounds=3
claudor this feature, builder=cursor, builder_model=grok-4.7-high
claudor this feature, reviewer_model=grok-4.7-xhigh

Controls

Argument	Default	Purpose
mode	full	review starts from an existing plan
plan / PLAN_FILE	PLAN.md	Plan path, carried through every phase
log / LOG_FILE	PLAN-REVIEW-LOG.md	Append-only decision log
builder	current host	claude or cursor
reviewer_model, builder_model, inspector_model	Cursor: $CLAUDOR_CURSOR_<ROLE>_MODEL, then $CLAUDOR_CURSOR_MODEL, then composer-2.5; Claude: CLI configuration	Per-role model override
reviewer_effort, builder_effort, inspector_effort	Claude CLI configuration	Claude only; Cursor effort is in the model id
rounds / MAX_ROUNDS	5	Completed plan-review round cap
MAX_FIX_ROUNDS	2	Build-fix attempt cap
MAX_INSPECTION_ROUNDS	2	Initial inspection plus one reinspection
research	proportionate to task	none, web, or explicitly authorized deep
inspect	on	off is an explicit, logged opt-out
PROOF_CMD	from plan/repo	Agreed command that verifies the deliverable

What an approval means

The runner validates a successful CLI turn and a structured review; an empty output or a session-start event cannot count as approval. The approval records the plan's path and SHA256; changing the plan invalidates it. Inspections record the pre-build commit and a fingerprint of the inspected changes, including staged and untracked files. Later code changes need another inspection.

A clean structured result does not prove the model is right. Zero findings is valid; a large number of findings is not a quality score. BLOCKED, execution failures and exhausted round budgets are surfaced rather than converted to approval.

Cursor reviews run in read-only ask mode with a private Cursor config that denies all shell commands and writes (the read-only modes alone let some commands through), and refuse to start while Cursor MCP servers are configured unless --allow-cursor-mcp is passed. Claude reviews expose only file reading/search, with customizations disabled and no MCP tools. Cursor builders never get --force: shell commands outside your Cursor allowlist are refused and reported. Delegated builds require a clean checkout. Details: runtime reference.

Development and verification

python -m pip install -r requirements-dev.txt
python scripts/validate.py
python -m unittest discover -s tests -v

Tests use disposable Git repositories and fake CLI processes, without model quota. Live CLI checks are recorded in VALIDATION.md.

Credits

• Claudex Loop, cross-model review design, runner and packaging: Chase AI (chaseai-yt/claudex-loop), MIT. See ACKNOWLEDGMENTS.md for the upstream community contributions.
• Legacy interview skills in legacy/: © Matt Pocock, MIT.
• Cursor adapter: okkindel.

MIT license