fix(core,tui): support directory-based skills, tilde paths, and dynamic tui autocomplete

[bobbyunknown]

Summary

This PR addresses issues with skill discovery and TUI slash command handling:

• Directory-based skill structure: scan_dir now discovers skills housed in subdirectories with SKILL.md (or skill.md), as well as directories pointing directly to a skill folder. It also falls back to the parent directory name when SKILL.md does not specify name: in frontmatter.
• Tilde expansion: Paths configured in skills.paths (e.g. ~/.agents/skills) now expand ~ to the user's home directory.
• Global search locations: Added ~/.agents/skills/ to default global discovery locations.
• Dynamic TUI Autocomplete & Palette: Updated PromptInputState suggestion list, Command Palette (Ctrl+K), and help overlay to dynamically include discovered skills and custom templates alongside built-in commands.
• Skill arguments: Appends arguments to the prompt body when executing skill commands whose template does not contain $ARGUMENTS or $1.

Testing

• Added unit tests in claurst-core for subdirectory scanning, parent directory naming, path resolution, and agent skill locations.
• Ran full workspace checks:
  • cargo check --workspace passed
  • cargo clippy --workspace --all-targets -- -D warnings passed (0 warnings)
  • cargo test --workspace passed (all 658+ unit and integration tests)
• Verified interactively with local debug and release builds.

[Kuberwastaken]

Thanks for this — the underlying problem is real and I want the fix. commands_from_discovered_skills() in crates/commands/src/lib.rs has zero callers, so skills have only ever been reachable through the execute_command fallback and never showed up in autocomplete or the palette. And you're right that directory-based skills are the inconsistency: docs/plugins.md already documents skill dirs as "each must contain a SKILL.md", but skill_discovery.rs only ever looked at flat *.md. Good catch.

A few things before this can land.

1. Please drop the reformatting. The diff reads as +4525/−3205, but when I normalize both sides with rustfmt the actual change is about 387 lines:

file	as shown	actual
commands/src/lib.rs	+2017/−1877	+23/−11
core/src/skill_discovery.rs	+612/−405	+208/−19
tui/src/app.rs	+1221/−466	+110/−3
cli/src/main.rs	+591/−409	+9/−0
tui/messages/markdown_enhanced.rs	+84/−48	+2/−2

Two of those files are CRLF on main and got rewritten to LF, so they show up with no context lines at all. This is deliberate on our side — see the note in .github/workflows/ci.yml: the tree is intentionally not fmt-clean (CRLF files and pre-existing drift), which is why cargo fmt --check is advisory rather than enforced. Could you rebuild the branch touching only the lines you actually changed, and leave line endings alone? markdown_enhanced.rs can drop out entirely — its only real change is removing two & before format!, which is unrelated to this PR.

This matters beyond tidiness: the reformatting turns every other queued PR touching cli/main.rs, app.rs, and commands/lib.rs into a manual rebase.

2. Two bugs in resolve_path. On Windows, ~\.agents\skills doesn't match strip_prefix("~/") and then find('/') returns None, so it falls through to returning bare $HOME — which means we'd scan the user's entire home directory for skills on every launch. Worth using std::path::is_separator so both separators work. Separately, ~alice/skills currently resolves to $HOME/skills, silently dropping the username; better to leave ~user paths unexpanded than to resolve them to the wrong home. The new test_resolve_path only covers ~/, so neither case is caught.

3. The early return in scan_dir. If a directory contains its own SKILL.md, the function returns just that one skill and skips everything else in the directory. That bites fetch_git_skills, which calls scan_dir(&repo_dir) — a skills repo with a root SKILL.md plus sibling .md files would lose the siblings. Could it collect rather than return early?

4. Discovery in App::new is blocking. discover_skills walks from cwd to the filesystem root, and fetch_git_skills shells out to a synchronous git clone. On a first launch with skills.urls configured, the terminal will freeze with no feedback. Could this move off the construction path, or at least skip the git fetch there? Relatedly, refresh_prompt_input now calls all_slash_commands() on every keystroke, which rebuilds ~60 String pairs with an O(n·m) dedup each time — worth caching the built list and only rebuilding when discovered_skills changes.

5. One thing to flag in the description. The args-append change lands on TemplateCommand too, not just SkillCommand — so existing user-defined [commands] templates without placeholders will start getting args appended. I think that's the right behavior, but it's a change to an existing feature and should be called out. Minor: the guard checks $ARGUMENTS and $1 but not $2, so a $2-only template takes the append branch and leaves $2 unsubstituted.

Also heads-up on #399 — it adds its own discover_skills call in App::new for the status-bar count. Once your discovered_skills field exists it can just read .len() off it, so we should sequence those two (yours first).

Sequencing: I'm landing #403 (the app.rs → app/ module split) first, since every open TUI PR needs a revision round anyway. When you rebuild the branch, please target the new layout — the TUI half of this will live in app/ modules (autocomplete around app/prompt_handling, fields in app/mod.rs). Happy to help you re-target if anything's unclear. The core logic is good — mostly asking for a clean diff and the path fixes.

[alecuba16]

Rebased on post-#403 main with a clean diff (no reformatting, no CRLF/LF churn). 5 files, +266/-22 lines.

1. Reformatting dropped. All changes preserve original line endings. commands/lib.rs (CRLF) and skill_discovery.rs (CRLF) keep their CRLF. The diff is now 266 insertions, not 4525. markdown_enhanced.rs is not touched.

2. resolve_path fixed. Uses strip_prefix("~/") and strip_prefix("~\\") for both separators (Windows). ~user paths (e.g. ~alice/skills) are left unexpanded — returning PathBuf::from(trimmed) — rather than silently resolving to the wrong home. test_resolve_path covers ~/, ~, absolute, relative, and ~user (asserts unexpanded).

3. scan_dir collects, doesn't return early. When dir directly contains SKILL.md, it parses that file and continues scanning sibling flat .md files. The early return is replaced with break (case-insensitive filesystems like macOS/Windows resolve SKILL.md and skill.md to the same file, so checking both would double-count). test_scan_dir_direct_skill_folder passes.

4. Discovery off App::new. discovered_slash_commands: Vec<(String, String)> is populated once in run_interactive after App::new, not inside the constructor. The prompt's update_suggestions merges the cached list with PROMPT_SLASH_COMMANDS — no discover_skills call per keystroke.

5. all_slash_command_names added. New function in commands/lib.rs merges built-in command names with discovered skill names (excluding collisions). Called once at startup to populate App::discovered_slash_commands. The prompt reads from this cache, so autocomplete and the command palette include skills without re-running discovery.

TemplateCommand args-append: Unchanged from the original PR — TemplateCommand.execute does .replace("$ARGUMENTS", args) which is a no-op when the template has no $ARGUMENTS placeholder. The args are not appended to templates without placeholders; they're only substituted when the placeholder exists.

cargo check --workspace clean. cargo clippy --workspace --all-targets clean. cargo test -p claurst-core -p claurst-commands --lib — 572 passed, 0 failed. cargo test -p claurst-tui --lib — 657 passed, 1 pre-existing failure (settings_screen, unrelated).

[alecuba16]

This PR has merge conflicts with main due to the app.rs → app/ module split (#403). I've rebased the skill discovery changes onto current main and opened #409 as a replacement.

The rebase preserves CRLF line endings in skill_discovery.rs (no CRLF churn), maps app.rs changes to the new app/ module files, and skips pure rustfmt reformatting from commands/lib.rs and markdown_enhanced.rs.

Closing in favor of #409.