Plugin matrix

Hashes the normalized query against the local result cache. A hit returns the stored answer and skips every downstream plugin — the cheapest possible hop.

Strips emails, keys, SSNs and phone numbers from text that might leave the device. In air-gapped mode it blocks every network plugin outright.

Replaces raw conversation dumps with compact int8 embeddings plus short extractive summaries. Typical prompt cut: 70–90%.

Cosine search over the on-device 96-dim quantized store. Only the most relevant snippets are eligible for the prompt.

Matches the query against persisted skill vectors so the pipeline can bias retrieval and routing toward the right capability.

Clips retrieved snippets to a configured token budget so a noisy store cannot inflate the prompt.

Scores complexity, privacy, cache state and estimated cost. Simple factual lookups stay local; hard reasoning may take a hybrid hop.

Builds an answer from the highest-scoring local snippets. Used whenever the router stays on-device or air-gap is armed.

Sends only retrieved, redacted snippets to a cloud reasoner. Blocked in air-gapped mode. Never receives raw history.

Writes the answer and routing trace back to IndexedDB so identical queries cost nothing next time.