Skip to content

The other rooms

Beyond /tsukue, kanade has five back-of-house rooms. They’re not the work surface — they’re the chronicle of the work.

A live feed of every Claude Code hook event as it fires: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Notification, SessionEnd. Filterable by event type with a one-click chip row at the top (each chip shows the matching count on screen).

New in v0.2.12: click-through detail. Every row is a button now. Click any event to slide out a right-rail detail drawer with:

  • The prompt text verbatim (for UserPromptSubmit)
  • The full tool input (for PreToolUse) — formatted JSON
  • The tool output, including is_error, duration_ms, and any error payload (for PostToolUse)
  • A usage block with input/output/cache-read tokens and cost (for Stop)
  • The raw envelope at the bottom for spelunkers

Session ids in the drawer link straight to /sessions/<id>; skill names link to /skill/<name>.

The ledger of every past sitting, day-grouped. Each row carries the project, duration, event count, tool counts, and any skills that fired.

New in v0.2.12: goal-based titles. Each session card now derives its heading from the first prompt of the sitting rather than just the project folder name. So “kanade · 2026-05-19 · 32m” becomes “Build the inline skill editor and test save round-trip · kanade · 2026-05-19 · 32m”. The full first prompt is one click away on the detail page under “opening prompt”.

The search bar is smarter too — it tokenizes on whitespace, so “fix auth” matches any session whose title, opening prompt, project, or path contains both fix and auth.

Every skill in ~/.claude/skills/ enumerated as a card: fire count, last-used date, 30-day sparkline, source path. Filter by name or text.

New in v0.2.12: inline SKILL.md editor. Click any catalog card to open /skill/<name>. Hit edit in the SKILL.md section and the body becomes a plain textarea with the verbatim frontmatter + body. Save atomically writes to disk (temp file + os.replace) after validating the YAML frontmatter — bad frontmatter is rejected with a 400 so the on-disk file is never corrupted. The fire/last-fired chips are honest now: when Claude Code hooks haven’t surfaced a fire, the card explains that and falls back to the heuristic prompt-mention count.

The smith’s appraisal. Heuristic suggestions across your skill library:

  • Retire candidates — skills that haven’t fired in 21+ days.
  • Never lit — installed skills that have never fired at all.
  • Growing quiet — skills with declining fire counts.
  • Champions — skills firing significantly above the median.

Each suggestion comes with a one-sentence rationale and a longer line in Kanade’s voice. New in v0.2.12: cards link to the skill page. A click on any forge card jumps straight to /skill/<name> so you can edit, retire, or inspect the fire timeline immediately.

The nightly reading. If you’ve connected a local LLM, Yume narrates one paragraph across your aggregate activity each day — what you worked on, what patterns showed up, what’s been quiet. If you haven’t connected a local LLM, this page goes quiet. Every other surface still works.