Add smart loop detection as a stop condition for agent execution. When
the same tool call signature (name + input + output) appears more than
5 times within a 10-step window, the agent stops instead of running
until context window exhaustion.

https://github.com/charmbracelet/fantasy/pull/138

fix(ui): chat: use follow indicator to determine auto-scrolling behavior

refs #2129

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

Replace github.com/sabhiram/go-gitignore (regex-based) with
github.com/go-git/go-git/v5/plumbing/format/gitignore (glob-based).

Key optimizations:
- Two-level caching: per-directory pattern cache + combined matcher cache
- O(1) fast-path for common directories (node_modules, .git, __pycache__, etc.)
- Pre-build combined matchers to avoid O(depth) pattern walking per file
- Proper isDir parameter for directory-specific patterns (e.g., "backup/")

Profiling showed 80% CPU in regexp.tryBacktrack from the old library when
walking large monorepos (771k files). After this change, gitignore matching
drops to ~2% of CPU time.

💘 Generated with Crush

Assisted-by: AWS Claude Opus 4.5 via Crush <crush@charm.land>

detection

`ctrl+space` is used by some terminal emulators and operating system,
so some users were enable to use it. `ctrl+t` should work everywhere.

Fixes #1618

* feat(lsp): show user-configured LSPs in the UI

This will show the user-configured LSPs as stopped in the UI.

Maybe we should have a different state for "waiting"?

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

* chore(lsp): mark "unstarted" LSPs as such, use named styles

* fix: add unstarted state

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

---------

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>
Co-authored-by: Christian Rocha <christian@rocha.is>

This panic happen once in a while on CI on Windows specifically.
I personally never saw it happening myself, but I think it's possible
to happen for the end user on Windows as well.

Looks like a potential bug on the interpreter, but in the meantime let's
at least recover from the panic and gracefully handle it.

    panic: ended up with a non-nil exitStatus.err but a zero exitStatus.code

    goroutine 61 [running]:
    mvdan.cc/sh/v3/interp.(*Runner).Run(0xc000220848, {0x1415220e0, 0xc00021a1e0}, {0x14151e088, 0xc00025a600})
    	C:/Users/runneradmin/go/pkg/mod/mvdan.cc/sh/v3@v3.12.1-0.20250902163504-3cf4fd5717a5/interp/api.go:929 +0x6b2
    github.com/charmbracelet/crush/internal/shell.(*Shell).execCommon(0xc000256360, {0x1415220e0, 0xc00021a1e0}, {0x14135a250, 0x9}, {0x14151baa0, 0xc00025a540}, {0x14151baa0, 0xc00025a580})
    	D:/a/crush/crush/internal/shell/shell.go:273 +0x285
    github.com/charmbracelet/crush/internal/shell.(*Shell).execStream(...)
    	D:/a/crush/crush/internal/shell/shell.go:288
    github.com/charmbracelet/crush/internal/shell.(*Shell).ExecStream(0xc000256360, {0x1415220e0, 0xc00021a1e0}, {0x14135a250, 0x9}, {0x14151baa0, 0xc00025a540}, {0x14151baa0, 0xc00025a580})
    	D:/a/crush/crush/internal/shell/shell.go:111 +0x139
    github.com/charmbracelet/crush/internal/shell.(*BackgroundShellManager).Start.func1()
    	D:/a/crush/crush/internal/shell/background.go:122 +0x15f
    created by github.com/charmbracelet/crush/internal/shell.(*BackgroundShellManager).Start in goroutine 28
    	D:/a/crush/crush/internal/shell/background.go:119 +0x72a

* fix(grep): do not go outside cwd, add timeout

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

* fix: timeout config

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

---------

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

fix(ui): adjust sessions dialog size

This change fixes a UI issue where the cursor is off by one line when
attachments are present in the editor.

* refactor: use csync.Map for regex caches

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

* fix: lint

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

* ci: golangci-lint 2.9

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

---------

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

Use a more reasonable size for the sessions dialog.

* fix: respect disable_default_providers

closes #1949

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

* fix: use x/slice

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

* fix: fail if no providers

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

* test: fix

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

---------

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

closes #1401

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>

Also refactor so we have the same pragmas on both drivers.

I couldn't reproduce OP's issue, but they're likely trying to use many
Crush instances at the same time, which may cause this.

The timeout was 5s only, so it was kind of easy to hit under load, I
presume. Upping it to 30s should improve that. AFAIK there's no much
else we can do.

See https://www.sqlite.org/rescode.html#busy

Closes #2129

Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com>