Transformer Pipelines

hivemind-core can run OVOS transformer plugins on the text/bus path, so plain-text clients (CLI, webchat, bridges — anything without the audio binary protocol) get the same rewriting machinery voice clients get.

• Utterance transformers rewrite recognizer_loop:utterance messages after policy admission and before they are injected into the agent bus.
• Metadata transformers enrich the message context at the same point.
• Dialog transformers rewrite QUERY/CASCADE answer chunks (speak messages) before they are streamed back to clients.

Cancellation follows OVOS-TRANSFORM §8.1/§8.2: a transformer returning a valid canceled/cancel_reason pair terminates the lifecycle — the utterance never reaches the agent, and the client receives ovos.utterance.cancelled followed by ovos.utterance.handled.

Configuration

New blocks in ~/.config/hivemind-core/server.json. Loading is opt-in: a plugin only runs when named in its section; empty sections (the default) are no-ops.

{
  "utterance_transformers": {
    "ovos-utterance-corrections-plugin": {},
    "ovos-utterance-cancel-plugin": {}
  },
  "metadata_transformers": {},
  "dialog_transformers": {}
}

Chains run in ascending priority order (OVOS-TRANSFORM §4); an explicit "order" list wins over priorities. See the ovos-plugin-manager transformer docs for the full contract.

When to use — and the surprise factor

Transformers on the hivemind server apply to every client in the mesh — that is the point, and also the surprise: a client that sends "turn on teh lights" and watches the bus will see its own utterance arrive corrected, and a QUERY client may receive answer text the agent never literally said (because a dialog transformer rewrote the chunk).

Deliberate uses:

• Mesh-wide policy stop-words: ovos-utterance-cancel-plugin here cancels utterances for every client, uniformly.
• Shared corrections/normalization: fix domain vocabulary once for all text clients.
• Uniform answer tone: a dialog transformer here rewrites QUERY/CASCADE answers for the whole mesh.

Watch for double-processing in split deployments: the OVOS agent behind this server (ovos-core) also runs utterance/metadata transformers, and hivemind-audio-binary-protocol runs its own chains for voice clients. Enable each plugin in exactly one place. As a rule: text-client-facing rewrites belong here; skill-facing rewrites belong in ovos-core; voice-specific chains belong in the audio binary protocol or on the satellite.