feat: add Cursor Agent support via --agent flag

[moisei]

Summary

Adds Cursor Agent support to RTK via a new --agent <name> flag on rtk init.

• rtk init -g --agent cursor installs a preToolUse hook in ~/.cursor/hooks.json that rewrites shell commands through rtk rewrite, using Cursor's JSON format ({permission, updated_input})
• Works with both Cursor editor and cursor-cli — they share ~/.cursor/hooks.json, so a single installation covers both
• rtk init --show now reports Cursor hook and hooks.json status alongside existing Claude Code / OpenCode status
• rtk init -g --uninstall also removes Cursor artifacts (hook script + hooks.json entry)
• rtk discover adds a note when Cursor hooks are detected, directing users to rtk gain for Cursor session analytics (Cursor transcripts use a text-only format without structured tool_use/tool_result blocks, so discover cannot extract commands from them)

Design decisions

• --agent is an extensible enum (claude, cursor), not a boolean flag — ready for future agents (Windsurf, Zed, etc.)
• Additive behavior: --agent cursor installs Cursor hooks ON TOP of the default Claude Code installation. Without --agent, behavior is 100% unchanged.
• --opencode stays separate for backward compatibility
• Cursor hook uses preToolUse with matcher: "Shell" and returns updated_input to rewrite commands — functionally identical to Claude Code's PreToolUse but with Cursor's JSON schema

cursor-cli note

The preToolUse hook mechanism is shared between Cursor's editor agent and its CLI agent (cursor-cli). Both read hooks from ~/.cursor/hooks.json, so rtk init -g --agent cursor covers both without any additional setup.

Files changed

File	Change
hooks/cursor-rtk-rewrite.sh	New: Cursor-format hook script (thin delegator to rtk rewrite)
src/main.rs	Add --agent <name> flag with AgentTarget enum
src/init.rs	Cursor install/uninstall/show functions, hooks.json patching, unit tests
src/discover/provider.rs	Doc comment update noting Cursor limitation
src/discover/report.rs	Cursor note in discover output when hook detected

Test plan

• cargo build — 0 errors
• cargo test — 899 passed, 0 failed
• rtk init --help shows --agent with claude and cursor values
• rtk init -g --agent cursor --auto-patch installs hook + patches hooks.json
• rtk init --show reports Cursor hook status
• rtk init -g --uninstall removes Cursor artifacts
• Cursor agent session with hook active rewrites git status → rtk git status
• rtk gain --history shows tracked commands from Cursor session

[pszymkowiak]

Please retarget this PR to develop instead of master. We use develop as the integration branch — master is for releases only. You can change the base branch in the PR settings. Thanks!

[pszymkowiak]

@moisei — Rebased your PR on latest develop (which now includes Gemini #573, Codex #377, Copilot #605, and OpenClaw #358).

Resolved merge conflicts in init.rs and main.rs:

• Added cursor: bool to uninstall() signature (alongside gemini and codex)
• Updated run() to accept install_cursor parameter
• --agent cursor works alongside existing --gemini, --codex, --opencode flags
• Fixed 2 unit tests to match new function signatures

Tested:

• rtk init -g --agent cursor installs hook + hooks.json correctly
• 970 tests passing (including Cursor tests)
• Compiles cleanly

Ready to merge. Thanks for the clean implementation — the --agent enum approach is exactly the right pattern for extensibility!

[moisei]

@pszymkowiak this is great.
what's the release process - how do I know which version contains this change?

[pszymkowiak]

we add feature to develop then there are tests on develop, then release

[fingon]

Does @RTK.md really work for you guys? when I run codex --ask-for-approval never "Summarize the current instructions." I do not get RTK instructions there, unless I inline the RTK.md contents to AGENTS.md (at least for global file). (sometimes it says 'it should include RTK.md' but never content of it so it does not really include it. Funnily enough sometimes later it looks for RTK.md in current project directory, which does not work.)

There is open issue about it here: openai/codex#6038

(I tested this commit with develop branch and wondered why rtk gain showed no use)