feat: visible_only mode to limit diffs to open buffers

[kam-hak]

Problem

If you're focused on one file but Claude is making small edits elsewhere (wiring, imports, config), nvim mirrors every edit with a diff tab and neo-tree reveal. There's no way to get full-context review for the file you care about while letting the rest flow through the CLI.

Design

visible_only gives finer control: files open in nvim get the full diff preview, everything else falls through to the Claude CLI's normal review flow.

Changes

• Add diff.visible_only config (default false, no behavior change)
• When enabled, only files open in a visible nvim window get diff previews and neo-tree reveals
• Add :ClaudePreviewToggleVisibleOnly for runtime toggling
• Add hook_context() to batch config + file visibility into one RPC call

Note: permissionDecision removal (separable)

This PR also removes the permissionDecision output from the PreToolUse hook. This is separable from the visible_only feature if you'd prefer to handle it differently.

Currently the hook unconditionally returns "ask", which overrides Claude Code's own permission system — including bypass mode. A user who has explicitly configured bypass still gets prompted on every edit. A preview hook should be a UI layer that shows diffs, not an approval gate. Happy to split this out if you'd rather discuss it separately.

Test plan

• Default config (visible_only = false) — all edits show diffs in nvim
• Enable visible_only = true — only open files show diffs, others go through CLI
• Toggle with :ClaudePreviewToggleVisibleOnly — takes effect on next edit
• Claude's own permission settings (bypass/ask) work independently of this hook

[kam-hak]

Self-review caught two issues. Fixed and pushed:

1. Same walk-up regression as feat: configurable neo-tree reveal with cross-CWD support #21 — dropped the nearest-existing-parent logic for created files, broke new-file reveal. Restored.
2. The is_mac + :lower() branch I had was wrong — fs_realpath already returns the filesystem's canonical form, so case-insensitive volumes normalize without per-OS code. Worse, lowercasing would produce false positives on case-sensitive filesystems (Linux, case-sensitive APFS). Dropped it.

[kam-hak]

Rebased onto main. Two commits:

1. feat: visible_only mode
2. refactor: drop permissionDecision output. My perspective is that an nvim plugin should not override CC's permissions. Currently, this triggers even if nvim is not open if there is a stale socket (which is common at least for me). If intentional, perhaps we add a respect CC's permission flag, and happy to split this as a separate issue and design discussion.

[kam-hak]

looking into these failing tests

[kam-hak]

Also opened #26 / #27 for a separate test-infrastructure bug.

The e2e stale-socket tests weren't actually exercising stale-socket discovery (they bypassed the scan logic via the NVIM_LISTEN_ADDRESS fast path, and had dead wrapper code). Independent of this PR, but relevant for test coverage around these hooks. Happy to rebase this PR on top of #27 if you merge that first.

[kam-hak]

Note on interaction with #27: if #27 merges first, this PR has a one-line merge conflict in test_stale_socket.sh (both PRs touch the same assertion in the test_no_nvim_graceful test). Resolution should be simple. Drop this PR's change to that file specifically, since #27 restructured the test entirely and its new version doesn't assert on the printf output. test_edit.sh changes here are independent and apply cleanly either way. Happy to do this if 27 is merged!

[Cannon07]

Great work on the visible_only feature — tested it and it works well. The runtime toggle is a nice touch too.

Regarding the permissionDecision removal — I'd prefer to keep it as the default since it's a core part of the review workflow for most users. However, the use case you described (bypass mode users getting prompted on every edit) is valid.

Would you be open to making it configurable instead of removing it? Something like:

diff = {
override_permission = true, -- default, preserves current behavior
}

When true (default), the hook returns "ask" as it does in main. When false, the hook produces no permissionDecision output and defers to Claude Code's own permission settings.

This way existing users get no behavior change, and power users who want bypass mode can opt in.
Let me know your thoughts.

[kam-hak]

Regarding the permissionDecision removal — I'd prefer to keep it as the default since it's a core part of the review workflow for most users. However, the use case you described (bypass mode users getting prompted on every edit) is valid.

Would you be open to making it configurable instead of removing it? Something like:

diff = { override_permission = true, -- default, preserves current behavior }

When true (default), the hook returns "ask" as it does in main. When false, the hook produces no permissionDecision output and defers to Claude Code's own permission settings.

This way existing users get no behavior change, and power users who want bypass mode can opt in. Let me know your thoughts.

Sure! Can put it behind a flag. Just so the intended behavior is clear, the flag might be "defer_claude_permissions = true" if turned on.

This would never override but defer to claude, avoiding having two different permission toggles which interact: one in this nvim addon, which injects 'ask' directly to claude (overriding CC's permission setting), and a second toggle in claude's session between vanilla/accept edits/plan/bypass window, which at least for me, gets overridden by this nvim.

If that sounds good can make the update.

[Cannon07]

That works! defer_claude_permissions with a default of false is a clean name — makes it clear what you're opting into.

[kam-hak]

Rebased onto main (now includes #27). Added defer_claude_permissions flag in a third commit:

1. feat: visible_only mode (unchanged)
2. refactor: drop permissionDecision output (unchanged)
3. feat: add defer_claude_permissions config flag (new)

defer_claude_permissions defaults to false, preserving the current "ask" behavior. When true, the hook produces no permissionDecision output and defers to Claude Code's own settings. Updated test assertions to match.

[Cannon07]

Great addition! Both features are well implemented.

LGTM. Merging!