fix(init): fail closed on missing .claude dir and hoist OpenCode install (fixes #4046)

[TechWizard9999]

Summary

Fixes #4046 regressions where --opencode was riding the Claude install path, so every Claude-path failure or early-exit took the OpenCode plugin down with it.

Changes

1. src/hooks/init/claude.rs:
   • Add fs::create_dir_all(&claude_dir) in patch_settings_json_command and run_default_mode to create the .claude directory if missing (fixes bug(init): global Claude init fails when ~/.claude is missing #2519, rtk init --opencode appears to skip Claude hook installation #4046 Windows regression)
   • Hoist OpenCode plugin installation block before Unchanged early return in run_claude_md_mode_with so OpenCode installs even when CLAUDE.md block is already up to date

Fixes Three Regressions

Command	Before (regression)	After (fixed)
init -g --opencode --auto-patch	exit 1, nothing installed	exit 0, both installed
init -g --opencode --hook-only --auto-patch	exit 1	exit 0, both installed
init -g --opencode --claude-md, re-run	exit 0, "already up to date", plugin missing	exit 0, both installed

Test Coverage

Existing test test_global_opencode_installs_when_claude_dir_missing (already #[cfg(unix)]) verifies the missing-dir fix. The hoisted OpenCode install ensures plugin installs even on re-run with unchanged CLAUDE.md block.

[rtk-wshm-sync-bot]

wshm · Automated triage by AI

📊 Automated PR Analysis

🐛 Type	bug-fix
🟡 Risk	medium

Summary

Fixes three regressions around global init --opencode: creates the .claude config directory if missing before writing settings/CLAUDE.md, and hoists the OpenCode plugin installation before the early return that fires when the CLAUDE.md block is already up to date. This ensures OpenCode installs correctly regardless of directory existence or whether the rtk block changed.

Review Checklist

• Tests present
• Breaking change
• Docs updated

Linked issues: #4046, #2519

---

Analyzed automatically by wshm · This is an automated analysis, not a human review.

[TechWizard9999]

@KuSh review pls , This fixes #4046 regressions 1) fs::create_dir_all only in non-dry-run mode (fixes #2519, #4046 Windows regression), 2) Hoist OpenCode install before Unchanged early return. All three regressions fixed.

[KuSh]

Round 2 of 3 — APPROVED (round 1 was posted on #4052, which this PR carries forward)

Claim: rtk init -g --opencode installs the Claude hook and the OpenCode plugin together (Fixes #4046).
Scope: accept (frozen at #4052 round 1). #4052 is closed in favour of this PR; no other open PR claims #4046.
Ran: 3 binaries (merge-base 7efbaef46, head 523b86e0, head merged onto current develop 083a9c5d), one CARGO_TARGET_DIR each, md5s distinct; 16-row rtk init matrix per binary under an isolated HOME, LC_ALL=C, stdin /dev/null, plus 8 probes (malformed CLAUDE.md block, read-only ~/.claude, --hook-only --no-patch, re-run output of every non-opencode mode, sibling agents, declined prompt); mutations: routing revert, hoist revert, redundant-line removal. Invariants: every decision branch of the new routing (default, --hook-only, --claude-md, --dry-run, local rejection, --agent cursor), idempotence, user settings.json/CLAUDE.md content preserved. Savings floor N/A (init change, no filtered output).
Matrix: executed this round on this head (never run on #4323 before; #4052's matrix ran on a different head).
CI @56a2c33da: 11/11 pass. Local gate on the fixup merged onto current develop: fmt, clippy, cargo test --all (4097) green.

Blocks merge (0)

The two #4052 mechanisms are fixed on this head: a fresh HOME with --opencode (default and --hook-only) exits 0 with hook + RTK.md + plugin, and re-running --claude-md --opencode restores a deleted plugin.

Fixed by my fixup (b95e5b755, 56a2c33da; low-hanging; threads closed by it)

• The two create_dir_all(&claude_dir)? lines are redundant. write_if_changed and the settings patch already create ~/.claude (#2521, merged 2026-09-26, after #4052's base — my "verified fix" part (a) in that review is superseded by it). With both lines removed, 14 of 16 matrix rows are byte-identical; the two that differ are ~/.claude as a file or dangling symlink, where the error now reads Failed to create directory: <path>: File exists instead of a bare File exists (os error 17). They also left an empty ~/.claude behind on --hook-only --no-patch (develop leaves nothing).
• Every no-op rtk init re-run printed an extra Claude Code will now use rtk… line (plain local rtk init, init -g --claude-md, and the --opencode variants); develop is silent. The Unchanged return keeps its silence now; the OpenCode install stays hoisted above it.
• No test pinned the change. Reverting the main.rs routing or the hoist left 310 init unit tests and 5 init integration suites green. tests/init_opencode_test.rs (spawning rtk through common::rtk_command(), as core::test_isolation requires) adds 5 tests (fresh home, --hook-only, unchanged-block re-run restores the plugin and stays silent, --hook-only --no-patch creates no ~/.claude, --dry-run writes nothing); the routing revert fails 5/5, the hoist revert fails the re-run test, head as authored fails exactly the two tests that pin its defects.
• Comment rewrite: no issue numbers or history, says why the install sits above the early return.

Filed as follow-ups

• NEW: a broken Claude side now aborts --opencode before the plugin is installed. --claude-md --opencode with a malformed <!-- rtk-instructions block, or --opencode with an unwritable ~/.claude, exits 1 with a clear message and no plugin (develop exits 0 and installs it). Loud rather than silent, and the class is the one named in #4052 round 1; filed as #4413, not blocking. Pre-existing: the same input already fails plain rtk init -g that way.

Checked and correct — no need to re-verify

Local --opencode rejected (exit 1); --agent cursor --opencode installs OpenCode + Cursor and never creates ~/.claude; dry-run output lists all six actions and writes nothing; existing settings.json keys and CLAUDE.md content survive byte for byte; second --opencode --auto-patch leaves settings.json identical; CLAUDE_CONFIG_DIR pointing at a missing nested dir works; --uninstall --opencode after an install removes both; docs wording (README, supported-agents, troubleshooting, --show text) matches behaviour; clean merge with #4343 and #4269.

Your questions

None open. Note for the description: the "before" column in the PR body describes #4052's head, not develop (develop installs only the plugin), and test_global_opencode_installs_when_claude_dir_missing is #2521's test; it passes without this PR's create_dir_all lines.

Next

Nothing from you: the fixup (tests, two removed lines, restored re-run silence) is pushed on top of your head, code-reviewed twice, CI 11/11 on 56a2c33da. Approving.
Rounds: 2/3. Threads: 0 open, 0 resolved this round.