0.5.1: TUI cleanup + accurate skill cost

CHANGELOG.md

@@ -7,6 +7,58 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.5.1] — 2026-05-22
+
+### Fixed (TUI honesty + correctness)
+
+- **Focus on tab-switch.** Each screen now calls `.focus()` on its
+  primary `DataTable` via `on_show`, so screen-level BINDINGS (Sessions
+  `d`, Perms `p`/`D`, Schedule `a`/`r`/`n`/`p`, Hooks `t`/`l`/`v`,
+  Memory `r`/`a`/`f`, Audit `s`) actually fire after a tab switch — they
+  previously silently no-op'd because focus stayed on `TabbedContent`'s
+  tab strip.
+- **Schedule key naming aligned.** Header text, F1 help and the empty-
+  state CTA now agree with the actual BINDINGS: `a` add, `r` remove, `n`
+  run-now, `p` promote. The header previously said `n` new job / Del
+  remove / `r` run-now — all three of which were wrong.
+
+### Changed (TUI now honest about its scope)
+
+- **Sessions `d` is real.** Highlight a row, press `d` → `ConfirmModal`
+  with session id / size / message count → on Yes, soft-deletes via
+  `core.sessions.delete_session` inside `tui_confirmed()` + `audit_action
+  (mode="tui")`. Restoration remains a CLI path
+  (`cc-janitor trash restore`).
+- **Perms `p` and `D` are real.** New `PrunePreviewModal` and
+  `DedupPreviewModal` list affected rules + source paths before applying;
+  on confirm, `core.permissions.remove_rule` is called per rule (each
+  call backs up the settings.json to `~/.cc-janitor/backups/` and writes
+  an audit entry).
+- **Sham bindings dropped.** Sessions `/` search and Audit `u` undo are
+  removed from BINDINGS, headers, and F1 help. Use `cc-janitor session
+  find` and `cc-janitor undo` from the CLI.
+
+### Fixed (cost calculation)
+
+- **Skill cost reflects per-request reality.** `core.context.enabled_skills`
+  used to count the full `SKILL.md` size in tokens, but Claude Code
+  truncates each skill body at `skillListingMaxDescChars` (default 1536).
+  cc-janitor now reads that setting from `~/.claude/settings.json` and
+  reports the truncated-slice cost. Also honours `skillOverrides`
+  (`"off"` excludes; `"name-only"` keeps only YAML frontmatter). Heavy
+  skills like graphify previously over-reported ~16x.
+
+### Tests
+
+- `tests/tui/test_focus_on_mount.py` — focus moves to DataTable on tab-
+  activate for every screen.
+- `tests/tui/test_sessions_delete.py` — `d` opens ConfirmModal; on Y the
+  row is removed.
+- `tests/tui/test_perms_actions.py` — `p` and `D` open the new preview
+  modals.
+- `tests/unit/test_context_skill_cost.py` — cap respected, overrides
+  applied, default value used when settings absent.
+
 ## [0.5.0] — 2026-05-21
 
 ### Fixed (Critical — TUI rendering)

README.md

@@ -169,21 +169,24 @@ cc-janitor stats sleep-hygiene      # memory hygiene metrics
 [F1] anywhere opens a tab-specific help modal):
 
 - **Sessions** — every Claude Code conversation in `~/.claude/projects/`,
-  sorted by last activity. `d` soft-deletes to trash. `/` filters.
+  sorted by last activity. `d` soft-deletes the highlighted session
+  (preview + confirm modal; restorable via `cc-janitor trash restore`).
 - **Permissions** — Bash/Edit/etc. allow rules merged from all settings
   layers, annotated with `Used90d` match-counts and STALE/DUP flags.
+  `p` prune-stale and `D` dedup both open preview modals before mutating.
 - **Context** — what gets injected into every request (CLAUDE.md, memory,
-  skills) with token totals and per-request cost.
+  skills) with token totals and per-request cost. Skill tokens reflect
+  the `skillListingMaxDescChars` cap (default 1536), not full file size.
 - **Memory** — your CLAUDE.md / MEMORY.md / feedback / project-state /
   reference files. `r` reinjects, `a` archives, `f` finds duplicates.
 - **Hooks** — PreToolUse / PostToolUse / Stop hooks merged across layers.
-  `l` toggles logging, `t` simulates the command.
-- **Schedule** — maintenance jobs via cron / schtasks. Empty? The right
-  pane lists the five built-in templates so you know what to add.
-- **Audit** — every mutation cc-janitor made, plus daily-stats sparklines
-  (toggle with `s`).
+  `l` toggles logging, `t` simulates the command, `v` opens source in $EDITOR.
+- **Schedule** — maintenance jobs via cron / schtasks. `a` add (from
+  template), `r` remove, `n` run-now, `p` promote-from-dry-run.
+- **Audit** — daily-stats sparklines (toggle with `s`); undo via
+  `cc-janitor undo` CLI.
 - **Dream** — Auto Dream before/after snapshot pairs with file-level
-  diffs and one-key rollback.
+  diffs (rollback via `cc-janitor dream rollback` CLI).
 
 On first launch a Welcome modal lists all eight tabs and what each shows;
 dismissing it touches `~/.cc-janitor/state/seen-welcome` so it never
@@ -406,9 +409,9 @@ cc-janitor completions install [bash|zsh|fish|powershell]
 
 -----
 
-## Status — `v0.4.0` Dream safety net
+## Status — `v0.5.1` honest TUI
 
-Shipped 2026-05-11. 202 passing tests. CI matrix Python 3.11 × 3.12 × Ubuntu × Windows.
+Shipped 2026-05-22. CI matrix Python 3.11 × 3.12 × Ubuntu × Windows.
 
 |Phase        |Released                 |Highlights                                                                                            |
 |-------------|-------------------------|------------------------------------------------------------------------------------------------------|
@@ -417,6 +420,7 @@ Shipped 2026-05-11. 202 passing tests. CI matrix Python 3.11 × 3.12 × Ubuntu 
 |**Phase 3**  |v0.3.0 (2026-05-11)      |Monorepo discovery / watcher daemon / stats dashboard / bundle export-import / shell completions      |
 |**Phase 3.x**|v0.3.1–0.3.3 (2026-05-11)|TUI safety-gate fix, undo, schedule audit, memory delete, backups list/prune                          |
 |**Phase 4**  |v0.4.0 (2026-05-11)      |Dream snapshot harness / dream doctor / sleep-hygiene metrics / settings audit hook                   |
+|**0.5.x**    |v0.5.0–0.5.1 (2026-05-22)|TUI rendering fix (1fr CSS), F1 help, focus on tab-switch, real Sessions delete / Perms prune+dedup, accurate skill cost |
 |**Phase 5**  |planned                  |Cross-platform hook fixers, `dream fix-stale-lock`, mutating Dream TUI actions, fuller I10/I11 closure|
 
 **Looking for early users on Windows.** CI covers the matrix but real-world Windows hook compatibility (PowerShell vs Git Bash vs WSL) has reported edge cases. The dedicated [issue template](.github/ISSUE_TEMPLATE/os-compatibility-report.md) takes ~5 min.

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "cc-janitor"
-version = "0.5.0"
+version = "0.5.1"
 description = "Tidy up your Claude Code environment — sessions, permissions, context, hooks, schedule."
 readme = "README.md"
 requires-python = ">=3.11"

src/cc_janitor/cli/__init__.py

@@ -21,7 +21,7 @@
 from .commands.undo import undo as _undo
 from .commands.watch import watch_app
 
-__VERSION__ = "0.5.0"
+__VERSION__ = "0.5.1"
 
 app = typer.Typer(no_args_is_help=False, help="cc-janitor — Tidy Claude Code")
 

src/cc_janitor/core/context.py

@@ -1,9 +1,13 @@
 from __future__ import annotations
 
+import json
 from dataclasses import dataclass
 from pathlib import Path
 
-from .tokens import count_file_tokens
+from .tokens import count_file_tokens, count_tokens
+
+# Claude Code default for `skillListingMaxDescChars` in user settings.
+DEFAULT_SKILL_LISTING_CAP = 1536
 
 
 @dataclass
@@ -68,18 +72,101 @@ def memory_files(*, claude_project_dir: str) -> list[ContextFile]:
     return out
 
 
+def _read_user_settings() -> dict:
+    """Read ~/.claude/settings.json, returning {} on missing/malformed."""
+    p = _user_home() / ".claude" / "settings.json"
+    if not p.exists():
+        return {}
+    try:
+        return json.loads(p.read_text(encoding="utf-8"))
+    except (json.JSONDecodeError, OSError):
+        return {}
+
+
+def _skill_listing_cap() -> int:
+    """Claude Code truncates each skill's body at this many chars per request.
+
+    Default is :data:`DEFAULT_SKILL_LISTING_CAP` (1536) — Claude Code reads
+    this from ``skillListingMaxDescChars`` in ~/.claude/settings.json.
+    """
+    raw = _read_user_settings().get("skillListingMaxDescChars")
+    try:
+        v = int(raw)
+        if v > 0:
+            return v
+    except (TypeError, ValueError):
+        pass
+    return DEFAULT_SKILL_LISTING_CAP
+
+
+def _skill_overrides() -> dict[str, str]:
+    """Claude Code's per-skill override map.
+
+    Values: ``"off"`` excludes the skill entirely; ``"name-only"`` hides the
+    description body so only the YAML frontmatter contributes tokens.
+    """
+    raw = _read_user_settings().get("skillOverrides") or {}
+    if isinstance(raw, dict):
+        return {str(k): str(v) for k, v in raw.items()}
+    return {}
+
+
+def _skill_name_from_path(skill_md: Path) -> str:
+    """Skill name = parent dir of SKILL.md (e.g. .claude/skills/graphify/ → 'graphify')."""
+    return skill_md.parent.name
+
+
+def _skill_text_for_request(body: str, cap: int, override: str | None) -> str:
+    """Return the slice of SKILL.md that actually loads into a request.
+
+    ``override="off"`` returns "" (skill excluded). ``override="name-only"``
+    returns only the YAML frontmatter (description hidden). Otherwise the
+    body is truncated to ``cap`` characters.
+    """
+    if override == "off":
+        return ""
+    if override == "name-only":
+        # Keep only the leading YAML frontmatter (between '---' delimiters).
+        if body.startswith("---"):
+            end = body.find("\n---", 3)
+            if end != -1:
+                return body[: end + 4]
+        # No frontmatter → fall back to first line (the name reference).
+        return body.split("\n", 1)[0]
+    return body[:cap]
+
+
 def enabled_skills() -> list[ContextFile]:
+    """Skill files actually injected per request.
+
+    Each SKILL.md is truncated at ``skillListingMaxDescChars`` (default
+    1536 — Claude Code's hard cap). Skills marked ``"off"`` in
+    ``skillOverrides`` are excluded; ``"name-only"`` keeps only the YAML
+    frontmatter.
+    """
     home = _user_home()
     out: list[ContextFile] = []
     skills_root = home / ".claude" / "skills"
-    if skills_root.exists():
-        for skill_md in skills_root.rglob("SKILL.md"):
-            out.append(ContextFile(
-                path=skill_md,
-                size_bytes=skill_md.stat().st_size,
-                tokens=count_file_tokens(skill_md),
-                kind="skill",
-            ))
+    if not skills_root.exists():
+        return out
+    cap = _skill_listing_cap()
+    overrides = _skill_overrides()
+    for skill_md in skills_root.rglob("SKILL.md"):
+        name = _skill_name_from_path(skill_md)
+        override = overrides.get(name)
+        if override == "off":
+            continue
+        try:
+            body = skill_md.read_text(encoding="utf-8", errors="replace")
+        except OSError:
+            continue
+        effective = _skill_text_for_request(body, cap, override)
+        out.append(ContextFile(
+            path=skill_md,
+            size_bytes=skill_md.stat().st_size,
+            tokens=count_tokens(effective),
+            kind="skill",
+        ))
     return out
 
 

src/cc_janitor/tui/_help.py

@@ -17,11 +17,9 @@
         "~/.claude/projects/<project>/<uuid>.jsonl. This tab lists all sessions,\n"
         "sorted by last activity.\n\n"
         "[bold]Keys[/bold]\n"
-        "  ↑↓        navigate rows\n"
-        "  Enter     show full preview in the bottom pane\n"
-        "  d         soft-delete the highlighted session (goes to trash,\n"
+        "  ↑↓        navigate rows (preview updates on highlight)\n"
+        "  d         soft-delete the highlighted session (with confirm;\n"
         "            recoverable via `cc-janitor trash restore <id>`)\n"
-        "  /         filter by message content\n"
         "  q         quit cc-janitor\n\n"
         "[bold]Tip[/bold]\n"
         "Heavy sessions (> 5 MB) inflate Claude Code's index. Prune old\n"
@@ -33,9 +31,8 @@
         "layers (user, project, local) with usage stats.\n\n"
         "[bold]Keys[/bold]\n"
         "  ↑↓        navigate rows\n"
-        "  p         prune stale rules (no matches in 90d)\n"
-        "  D         dedup overlapping rules\n"
-        "  /         filter by tool or pattern\n\n"
+        "  p         prune stale rules (no matches in 90d) — preview + confirm\n"
+        "  D         dedup overlapping rules — preview + confirm\n\n"
         "[bold]Columns[/bold]\n"
         "  Used90d   number of tool_use events matching this rule\n"
         "  Flags     STALE = no recent matches, DUP = overlaps another rule"
@@ -72,19 +69,19 @@
         "Maintenance jobs running via cron (Unix) or schtasks (Windows).\n"
         "All new jobs start in dry-run mode — verify, then promote.\n\n"
         "[bold]Keys[/bold]\n"
-        "  n         add a new job from a template\n"
-        "  Del       remove the highlighted job\n"
-        "  r         run the highlighted job now\n"
+        "  a         add a new job from a template\n"
+        "  r         remove the highlighted job\n"
+        "  n         run the highlighted job now\n"
         "  p         promote a dry-run job to live mode"
     ),
     "audit": (
         "[bold]Audit[/bold]\n\n"
         "Every mutation cc-janitor made — when, why, and how to revert.\n"
         "Daily snapshots feed the sparklines panel.\n\n"
         "[bold]Keys[/bold]\n"
-        "  s         toggle the sparklines panel\n"
-        "  u         show undo instructions for the most recent action\n\n"
-        "Take a snapshot with `cc-janitor stats snapshot`."
+        "  s         toggle the sparklines panel\n\n"
+        "Take a snapshot with `cc-janitor stats snapshot`.\n"
+        "Use `cc-janitor undo` from the CLI to revert the most recent mutation."
     ),
     "dream": (
         "[bold]Dream[/bold]\n\n"

src/cc_janitor/tui/screens/audit_screen.py

@@ -26,13 +26,12 @@ class AuditScreen(Widget):
 
     BINDINGS = [
         ("s", "toggle_sparklines", "Toggle stats"),
-        ("u", "undo", "Undo recent"),
     ]
 
     def compose(self) -> ComposeResult:
         yield Static(
             "[bold]What cc-janitor has done on your behalf[/bold] [dim]"
-            "· s toggle sparklines · u undo recent[/dim]",
+            "· s toggle sparklines · undo via `cc-janitor undo` CLI[/dim]",
             id="audit-header",
         )
         yield Static("Audit log viewer", id="audit-body")
@@ -69,8 +68,3 @@ def _refresh_sparklines(self) -> None:
     def action_toggle_sparklines(self) -> None:
         panel = self.query_one("#audit-sparklines", Static)
         panel.display = not panel.display
-
-    def action_undo(self) -> None:
-        self.notify(
-            "Run `cc-janitor undo` from the CLI to revert the most recent mutation.",
-        )

src/cc_janitor/tui/screens/context_screen.py

@@ -51,3 +51,6 @@ def on_mount(self) -> None:
             f"[b]≈ ${dollars:.4f}[/] per request at Opus input rate"
         )
         self.query_one("#context-totals", Static).update(text)
+
+    def on_show(self) -> None:
+        self.query_one("#context-table", DataTable).focus()

src/cc_janitor/tui/screens/dream_screen.py

@@ -51,6 +51,9 @@ def on_mount(self) -> None:
             )
         self._show_diff_for(None)
 
+    def on_show(self) -> None:
+        self.query_one("#dream-list", DataTable).focus()
+
     def on_data_table_row_highlighted(
         self, event: DataTable.RowHighlighted
     ) -> None:

src/cc_janitor/tui/screens/hooks_screen.py

@@ -51,6 +51,9 @@ def on_mount(self) -> None:
         self._source_filter = "real"
         self._reload()
 
+    def on_show(self) -> None:
+        self.query_one("#hooks-table", DataTable).focus()
+
     def on_select_changed(self, event: Select.Changed) -> None:
         if event.select.id != "hooks-source-filter":
             return

src/cc_janitor/tui/screens/memory_screen.py

@@ -58,6 +58,9 @@ def on_mount(self) -> None:
         self._source_filter = "real"
         self._reload()
 
+    def on_show(self) -> None:
+        self.query_one("#memory-table", DataTable).focus()
+
     def on_select_changed(self, event: Select.Changed) -> None:
         if event.select.id != "memory-source-filter":
             return