[rNoz]

Summary

Two defects made rtk init -g --agent cursor fail / pollute the home dir:

1. It fell through to the shared init path with install_claude derived only from !opencode, so it created RTK.md, CLAUDE.md, and patched settings.json under ~/.claude in addition to patching .cursor/hooks.json. On machines without ~/.claude it failed outright: Failed to create temp file in ~/.claude: No such file or directory.
2. Even after removing the .claude writes, the Cursor path still failed on a fresh machine because install_cursor_hooks never created ~/.cursor before atomic_write placed its temp file there.

Fixes:

• Gate install_claude on the absence of a sibling agent target (Cursor, Windsurf, Cline), so only the default or an explicit Claude target installs Claude Code files.
• Create ~/.cursor (non-dry-run) before patching hooks.json.
• Should complete/close rtk init -g --agent cursor fails #2097 (same root cause as the previously-closed bug: "rtk init -g --agent cursor" errors due to expecting $HOME/.claude #1465)

Root cause

• src/main.rs, Commands::Init handler: install_claude = !opencode. Cursor is the only sibling agent that reaches the shared init path (Windsurf/Cline short-circuit earlier in hooks::init::run), so it incorrectly triggered the Claude install branch.
• src/hooks/init.rs, install_cursor_hooks: wrote hooks.json via atomic_write without ensuring the parent directory exists.

Fix

// src/main.rs
let install_cursor = agent == Some(AgentTarget::Cursor);
let install_windsurf = agent == Some(AgentTarget::Windsurf);
let install_cline = agent == Some(AgentTarget::Cline);
let install_claude =
    !opencode && !install_cursor && !install_windsurf && !install_cline;

// src/hooks/init.rs (install_cursor_hooks)
if !dry_run {
    fs::create_dir_all(&cursor_dir)?;
}

Test plan

• cargo fmt --all && cargo clippy --all-targets && cargo test --all: pass.
• New tests/init_agent_cursor_test.rs:
  • agent_cursor_dry_run_never_touches_claude: asserts dry-run output contains no /.claude/ path and does patch /.cursor/.
  • agent_cursor_real_write_creates_no_claude_dir: non-dry-run install in a temp HOME: exits 0, never creates ~/.claude, and writes .cursor/hooks.json.
  • default_dry_run_still_installs_claude: regression: default init still installs Claude Code files.
• cargo test --bin rtk hooks::init: 169 existing init tests pass (no regression).

E2E Validation

Tested locally and providing proofs (Screenshot). All working as expected:

[rNoz]

Tested E2E locally with my local rnoz/integration branch. All working as expected. See the capture. Ready for review/merge.

[rNoz]

@pszymkowiak please, let me know if I have to do anything else here (or it's just waiting), since another PR was closed in favor of my current one. I have also another one reviewed by some independent folk. Have a good one!

[theabhinavdas]

@rNoz Curious if this fixes rtk-ai usage on the Cursor IDE vs. the Cursor CLI tool? Or is rtk-ai meant only for terminal-based coding agents like Claude Code and Cursor CLI?

[rNoz]

@rNoz Curious if this fixes rtk-ai usage on the Cursor IDE vs. the Cursor CLI tool? Or is rtk-ai meant only for terminal-based coding agents like Claude Code and Cursor CLI?

I focused on the cli for testing, but if I have to guess, it will work for both.

[theabhinavdas]

LGTM 👍

I've reviewed this PR and tested on my machine:

Macbook Pro M1 Max - Mac OS 15.5 (24F74) - Sequoia

Cursor Desktop (non CLI):
Version: 3.13.10
VS Code Extension API: 1.128.0
Commit: 4f02290ccd9304f0e6bf8ee85f6e9106f02ac1f0
Date: 2026-07-23T21:41:07.333Z
Layout: IDE
Build Type: Stable
Release Track: Default
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.256
OS: Darwin arm64 24.5.0

These changes expertly fix the original issue(s) described.

However, these changes alone do not solve the wider rtk usage on Cursor IDE (non CLI).

It seems that the rewrites made by rtk use "permission":"ask" in update_input which doesn't actually show a permission interrupt for users to decide. But it looks like #1173 solves for this issue (untested).

[rNoz]

Friendly ping @aeppling. This has been open since July 14 with a community approval (@theabhinavdas tested it on Cursor IDE, MacOS Sequoia) but no maintainer review yet. It closes #2097: rtk init -g --agent cursor was writing RTK.md/CLAUDE.md/settings.json into ~/.claude and failing on fresh machines where ~/.cursor did not exist. Rebased on current develop; fmt/clippy/tests green. Happy to adjust anything if you see something.

[PedroS11]

Is there any ETA on this MR? Cursor users would really appreciate it 😞

[rNoz]

Yeah this is depressing (not only in this repo but others AI-powered as well), ETA hopefully by 2030.

[KuSh]

Round 1 of 3 — APPROVED, follow-ups filed

Claim: rtk init -g --agent cursor installs the Cursor hook only (no RTK.md, CLAUDE.md or settings.json under ~/.claude) and creates ~/.cursor before writing hooks.json, so it works on a machine that has neither directory (#2097).
Scope: accept, as a Cursor-only fix. Making every installer create its own parent directory is a separate follow-up and is not asked of this PR. (round 1, frozen)
Ran: all 3064 rtk init forms × 6 home seeds (empty, user content, env dirs, empty env, invalid JSON, existing backups) × install / re-run / uninstall, develop 91872ab vs this PR merged onto it. Every one of the 215 differences is an --agent cursor row, 20 of them go from exit 1 to exit 0, there are no crashes, and no other agent changes. On the seeds with user content, develop rewrites the user's ~/.claude/CLAUDE.md and this PR leaves it byte-identical. Also by hand: cursor with --claude-md, --hook-only, --opencode, --no-patch; ~/.cursor as a symlink, a dangling symlink and a regular file; project-scoped refusal; dry-run creates nothing. Five mutations (gate reverted, create_dir_all dropped, dir created in dry-run, gate narrowed to agent.is_none(), Cursor skipped with --opencode) are each caught by a test. Invariants: non-default targets, every decision, user content. No filter involved, so no savings figure.
Matrix: executed this round
CI @f29882be: all green (11/11, incl. macOS, Ubuntu, Windows)

Blocks merge (0, frozen at round 1)

None.

Optional — will not hold merge

• src/main.rs: install_claude lists what to exclude (!install_cursor && !install_windsurf && !install_cline). The Windsurf and Cline terms never take effect, because hooks::init::run sends both to their own modes before it reads install_claude. An allow-list, !opencode && matches!(agent, None | Some(AgentTarget::Claude)), has the same behaviour today and would stop a future agent that falls through to this branch from picking up Claude files.
• Users who ran --agent cursor before this fix still have RTK.md, the @RTK.md line and the settings.json hook under ~/.claude. rtk init -g --uninstall removes them. That is worth a line in the release notes, not code here.

Filed as follow-ups

• #913 (false "No hook installed" warning): the check only looks for the Claude hook. After this PR, a Cursor-only install on a machine that has a ~/.claude directory hits it: rtk ls prints the warning and suggests rtk init -g. Gemini-only installs already get it on develop. Evidence added there.
• #4248: after this PR, --agent cursor silently ignores --claude-md and --hook-only, as --agent windsurf and --agent cline already do on develop. --no-patch still patches Cursor's hooks.json, also as on develop. These are symptoms of the way rtk init chooses its target, so #4248 proposes one --agent <name> form with a per-agent list of supported options, rather than one more guard per combination.

Checked and correct — no need to re-verify

• The install_claude = false, install_cursor = true case lands in the (false, false, _, _) arm of run, which develop already keeps from bailing when Cursor is requested. --opencode installs the plugin and then the Cursor hook.
• create_dir_all runs only outside dry-run and before the legacy-script migration. A symlinked ~/.cursor is followed exactly as develop's atomic_write follows it. A dangling symlink or a regular file at ~/.cursor fails with Failed to create …/.cursor: File exists; develop failed on those too.
• --agent cursor --uninstall was already Cursor-only on develop, so install and uninstall now match.
• --agent claude and a plain -g install exactly the same Claude files as before.
• --agent cursor no longer writes ~/.config/rtk/filters.toml. Only the Claude path generates that template, and --opencode on its own already skips it on develop, so Cursor now behaves like the other non-Claude installs.

Hypotheses built and dropped

• Reusing #2521's ensure_parent_dir here instead of create_dir_all: this belongs to the follow-up that makes all the installers create their directories the same way, so it is not asked of this PR.

Your questions

• "please, let me know if I have to do anything else here" → nothing. I pushed the fixup below, which carries the rebase.

Next

Fixup pushed, then approving:

• a2ab981f merges develop and moves the ~/.cursor hunk into src/hooks/init/cursor.rs after the init.rs split (#3878).
• f29882be pins the new tests' environment inside the temp home. With CLAUDE_CONFIG_DIR exported, default_dry_run_still_installs_claude failed, and the real install wrote history.db into the caller's XDG_DATA_HOME. The commit also checks the dry runs' exit status, adds tests for --agent claude and --agent cursor --opencode, and corrects the "Cursor hooks (additive, installed alongside Claude Code)" comment in init::run.

Rounds: 1/3. Threads: 0 open, 0 resolved this round.

[rNoz]

Not bad: 2 MONTHS AND 11 DAYS to merge it. Good job, what a repo.