DJ — On-demand journal-track composer

A nested swiss-knife reference for composing one music track per project journal entry. Use this when the user asks for music tied to what they've been working on — "make a track for today's journal", "give me a bossa nova for last week", "the journal mentioned X — try a piece in style Y".

This skill assumes you already have file/bash/skills capabilities. It does not install or configure any media provider — it discovers what is available in your skills catalog and composes via whichever skill matches the user's saved presets.

When to load this skill

• User asks for music tied to a specific project, journal date, or session.
• User asks what genre would fit the project's mood.
• User asks why a previous track sounded the way it did, or wants another take on the same journal entry.

If the request is ambiguous (which journal? which genre? which project?), ask first. Don't guess and burn a generation call.

Preflight — find a usable provider

Before composing, you need to know what providers you can actually reach. Run this the first time the user asks for music in a session, and re-run when the request would use a media path you haven't tried this session.

Step A — enumerate available media-creation skills. Look in your skills catalog for skills tagged media-creation. If your catalog only shows name + description, the description still mentions "media-creation" for tagged skills — grep that.

If you want the canonical list, walk the configured skills paths shown by the skills catalog:

bash
for path in <skills paths from skills(action="info")>; do
  for skill in "$path"/*/SKILL.md; do
    [ -f "$skill" ] || continue
    grep -l -E '(^|, )media-creation(,|]| )' "$skill" 2>/dev/null
  done
done

For each one, the skill itself is the source of truth on what providers it talks to and what env-var key it expects. For MiniMax specifically, load the sibling Swiss Knife minimax-cli reference (../minimax-cli/SKILL.md); it is the canonical provider reference and explains how to scan TUI presets recursively, pick the declared MiniMax slot, export it without printing the key, and match the region.

Step B — cross-check against the user's saved presets. Each media-creation skill expects an API key. Scan ~/.lingtai-tui/presets/**/*.json recursively for each preset's manifest.llm provider + api_key_env (slot name) — the sibling ../minimax-cli/SKILL.md §3 carries the canonical JSONC-safe scanner; reuse it rather than re-implementing the comment-stripper here. The saved keys themselves live in ~/.lingtai-tui/.env; list key names only, never values:

bash
grep -E '^[A-Z0-9_]+_API_KEY=' ~/.lingtai-tui/.env | cut -d= -f1

Step C — intersect. A media-creation skill is usable only if (a) its provider matches a saved preset's provider AND (b) its expected env-var key is present in .env. Build the list of usable providers.

Step D — decide.

• Any usable provider exists → pick one (prefer the user's stated provider; otherwise pick whichever matches a current preset they're using; otherwise pick the first). Load its skill, follow its instructions, compose.
• No usable provider → reply plainly. Tell them what skills you found, which providers they imply, and which presets they'd need to add for those skills to work. For MiniMax, suggest concretely: "save a MiniMax preset via the TUI preset library and paste your sk-cp-… key — this will populate that preset's slot in ~/.lingtai-tui/.env and unlock the minimax-cli skill." Do not produce a fake track. Do not pretend.

Genre palette

Start from this palette unless the user specifies otherwise:

• 符合周礼 / 雅乐 — court-ritual music in the Zhou-li tradition; ceremonial, restrained, modal, suitable for milestones and decisions of weight.
• Bossa nova — gentle, syncopated, warm; for sessions that flowed easily.
• Jazz — small-combo or trio; for sessions full of improvisation, exploration, course-correction.
• Lo-fi hip-hop — relaxed, instrumental, low-stakes maintenance work and refactors.
• Ambient / drone — long-form thinking, deep architecture work, contemplative sessions.
• Classical chamber — careful, structured engineering; quartet textures.

The user may request anything outside this palette — Ravel, Coltrane, City Pop, 戏曲, gamelan, anything. Honor specific requests. The palette is a starting point, not a fence.

Composition working order

1. Parse the request. Which project? (Default: the current project — its hash is the one matching the user's working directory in the registry.) Which journal date or hour? (Default: most recent.) Which genre? (If unspecified, propose 2–3 from the palette that fit the journal's mood and ask, OR pick one and explain your choice in the reply.)

1. Read the journal. Project journals live at ~/.lingtai-tui/brief/projects/<hash>/journal.md. The brief.md / profile.md files in the same tree give you context on the user. If the user points at a specific date or hour, also consult the matching history/<YYYY-MM-DD-HH>.md. Distill: what did the user do? What was the emotional arc? What instrumentation, tempo, key, mood would honor this session?

1. Load the chosen media-creation skill by reading its SKILL.md from the skills catalog. For MiniMax, this is the sibling Swiss Knife minimax-cli reference (../minimax-cli/SKILL.md); use its preflight instead of inventing credential or region rules here. Follow the provider skill's live-doc / --help guidance so you have the current schema, model list, response shape, and expected wait time.

1. Compose the prompt. Translate the journal's mood into a music-generation prompt: genre, instruments, tempo, key, mood adjectives, optional structure (intro / verse / breakdown / outro), reference artists if useful. Keep it under whatever the API limit is per the live docs.

1. Call the provider. Use the provider skill's current mechanism (for MiniMax, normally mmx music ... via minimax-cli; for other providers, whatever their skill documents). If the response is a URL, curl -o it down; if a base64 blob, decode to file; if the CLI writes an output directory, move/copy the final audio into place. Save to ~/.lingtai-tui/brief/projects/<hash>/music/<YYYY-MM-DD>-<genre-slug>-<short-title-slug>.<ext>. Create the music/ folder if it doesn't exist. Do not overwrite an existing track for the same journal date — append a counter (-2, -3) if the user asks for another take.

1. Append the index entry. ~/.lingtai-tui/brief/projects/<hash>/music/index.md:

markdown

• 2026-04-29 — bossa nova — "Refactor in B♭" · journal 2026-04-29 · gentle syncopation for a clean refactor day

Create the index file if it doesn't exist.

1. Reply to the user. Tell them what you composed, where the file is, and one or two sentences on why this genre fit. Do not include the API request/response payload — keep the reply concise.

What this skill does NOT do

• No journal mutation. You read journals; you do not edit them.
• No retries on long-running media calls. Provider music endpoints can take 1–10 minutes — wait it out, do not retry.
• No fake tracks. If no usable provider is configured, say so plainly. Silence beats deception.

> Found a bug or issue? If you encounter any problems with this skill, load the lingtai-issue-report skill and follow its instructions to report it.