Cursor plugin¶

What ships in the plugin¶

Skills (directory vs chat command)¶

Chat commands use the name: field in each SKILL.md (not always the folder name on disk):

Codex and Claude Code plugins ship six shared skills (includes platform-observability) but not standalone production-triage or blast-radius skills.

Install¶

Install from the Cursor marketplace¶

The public storefront is orenlab/codeclone-cursor.

If CodeClone is already listed in your marketplace panel, select CodeClone, choose user or project scope, and install it.

To expose the repository as a team marketplace:

1. Open Cursor Dashboard → Settings → Plugins.
2. Under Team Marketplaces, select Add Marketplace.
3. Select Import from Repo and enter https://github.com/orenlab/codeclone-cursor.
4. Add CodeClone, configure team access, and save.
5. Install CodeClone from Cursor's marketplace panel.

Install codeclone[mcp] separately so the bundled launcher can resolve codeclone-mcp:

uv tool install "codeclone[mcp]"
codeclone-mcp --help

Local development only¶

Use a local symlink only while developing the plugin:

ln -sfn /path/to/codeclone/plugins/cursor-codeclone ~/.cursor/plugins/local/codeclone

Reload Cursor after changing the local source. Do not present this path to normal users as the installation flow.

Project hooks (Hooks UI)¶

uv run python plugins/cursor-codeclone/scripts/install-project-hooks.py
# full-repo gate:
uv run python plugins/cursor-codeclone/scripts/install-project-hooks.py --enforce-scope repo

Writes:

• .cursor/hooks.json — shown in Settings → Hooks
• .cursor/codeclone-hooks.json — enforce_scope (python default, or repo)

Do not commit generated files (machine-local Python paths). This monorepo ignores /.cursor/ in .gitignore.

Skills¶

codeclone-production-triage¶

Two MCP calls: analyze_repository then get_production_triage. Baseline-relative triage — not patch-local verify. Suggests codeclone-review for a deeper session.

codeclone-hotspots¶

Cheapest ad-hoc snapshot after analyze_repository; prefer list_hotspots / check_* before broad list_findings. Optional help(topic="coverage") when Coverage Join semantics matter.

codeclone-blast-radius¶

Read-only: get_blast_radius after analysis. Does not call start_controlled_change. Use codeclone-change-control for edits.

codeclone-review¶

Conservative-first full review; optional deeper pass with explicit user request. Does not declare intent by itself.

codeclone-change-control¶

Normal edit cycle uses workflow tools (not legacy-only atomic path):

analyze_repository → start_controlled_change → get_relevant_memory → edit in scope → analyze_repository (when after-run required) → optional record_candidate → finish_controlled_change.

Queue/recovery: manage_change_intent (promote, recover, …). Atomic check_patch_contract / create_review_receipt are advanced/debug only when workflow tools are unavailable.

codeclone-implementation-context¶

Bounded structural, call-graph, contract, and change-control evidence from one stored MCP run. Call after start_controlled_change with intent_id before editing scoped Python work. Read-only — does not declare intent.

codeclone-engineering-memory¶

Scope memory before edits; optional semantic=true on mode=search when [tool.codeclone.memory.semantic] is enabled, the semantic sidecar is installed, and semantic index rebuild succeeded (manage_engineering_memory action=rebuild_semantic_index or CLI memory semantic rebuild). Use codeclone[semantic-local] plus embedding_provider = "fastembed" for local semantic-quality recall; codeclone[semantic-lancedb] alone supports only the deterministic diagnostic provider. Human approve/reject: VS Code Memory view (preferred) or CLI codeclone memory approve|reject|archive --i-know-what-im-doing (MCP agents cannot approve).

Full contract: Engineering Memory.

codeclone-platform-observability¶

Maintainer-only — not for users reviewing their Python repository.

Diagnose CodeClone's own runtime (MCP latency, DB cost, memory pipeline) via query_platform_observability after explicit observer setup:

export CODECLONE_OBSERVABILITY_ENABLED=1
# restart codeclone-mcp / CLI with this env, reproduce, then query sections

Without enablement the tool returns status=disabled or no_store. Never treat observer metrics as repository quality or edit authorization.

Playbook: Maintainer workflow.

Agent¶

codeclone-structural-reviewer¶

Defined in agents/structural-reviewer.md with frontmatter name: codeclone-structural-reviewer. Read-only review protocol; does not declare intent or modify files. The structural reviewer agent uses CodeClone MCP tools exclusively for evidence, does not modify files or declare change intent, and does not treat report-only signals as CI failures or vulnerability claims.

Distribution¶

• Monorepo source: plugins/cursor-codeclone/
• Marketplace source: https://github.com/orenlab/codeclone-cursor
• Install: Cursor marketplace panel; local symlink only for development
• Standalone releases: ship full plugins/codeclone/scripts/launch_mcp.py body

Runtime model¶

Additive: local MCP via launch_mcp.py, eight skills, three rules (two alwaysApply + one Python glob), optional hooks. The full default agent MCP surface is passed through — the launcher does not pass --ide-governance-channel (VS Code adds +2 IDE-only tools and Memory governance). New server tools from upgraded codeclone-mcp pass through unfiltered.

Monorepo: plugins/cursor-codeclone/scripts/launch_mcp.py delegates to plugins/codeclone/scripts/launch_mcp.py. Standalone releases must embed the full launcher body.