dsh-midia

Configuration

Precedence is profile config → environment variable → default. ~ is expanded to your home directory. Defaults follow the reference layout under your home directory; set yours to match where your media and engine actually live.

Key	Environment variable	Default	Meaning
musicas	DSH_MIDIA_MUSICAS	~/workspaces/midia/musicas	music library root (guard and returned paths)
videos	DSH_MIDIA_VIDEOS	~/workspaces/midia/videos	video library root
gemma	DSH_MIDIA_GEMMA	~/workspaces/tools/embeddinggemma/bin/embeddinggemma	EmbeddingGemma binary
estado	DSH_MIDIA_ESTADO	~/workspaces/harness/global/scripts/embeddings/.estado	optional index state sidecars (granularity note)
memsearch	DSH_MIDIA_MEMSEARCH	~/.local/bin/memsearch	memsearch CLI, for -bgem3-v1
colecao	—	midia-gemma-v1	collection searched by default
n	—	10	items returned when the model does not pass n
guarda	—	true	library guard on/off

Override by id in your profile patch (~/.dsh/profiles/web/cordis.patch.yml). A patch replaces the whole config of the line, so restate every key you want to keep:

- id: midia
  config:
    colecao: mymusic-gemma-v1
    n: 10
    guarda: true
    musicas: ~/Media/music
    videos: ~/Media/video
    gemma: ~/tools/embeddinggemma/bin/embeddinggemma

Environment variables are handy when the engine lives outside the default place:

export DSH_MIDIA_GEMMA="$HOME/tools/embeddinggemma/bin/embeddinggemma"
export DSH_MIDIA_MUSICAS="$HOME/Media/music"

Usage

In the chat:

/midia slow jazz for studying -n 5
/midia --mais calm --menos drums
/midia-parecidas "/home/you/Media/music/kind-of-blue.flac"
/midia-duplicatas --camada todas --propor

Or let the agent call midia_buscar with a description. Each returned item has a name, an absolute path when it can be resolved, a score, a type (faixa, trecho, quadro) and, for slices, the timecode (HH:MM:SS plus the start in seconds).

The direction arithmetic (--mais / --menos) is not normalized by the engine: only the order is meaningful there, not the absolute scores.

/midia-duplicatas runs three layers: identical bytes (md5), same recording in another format (SimHash over strings), and near-duplicate vectors (cosine). It writes a proposal under ${XDG_STATE_HOME:-~/.local/state}/dsh-midia/-duplicatas/ and never touches the media. The optional model verdict (--llm) is not implemented; the three mechanical layers are.

The guard

While guarda is true, the plugin listens on tools/pre-execute and denies a bash command whose verb is destructive (rm, mv, rename, unlink, shred, truncate, dd) when it targets a media file under musicas or videos. Read-only commands (ffprobe, ls, find, indexers) pass. The guard covers the bash tool only — other tools are not inspected — and paths outside the configured roots are not its business.