Integrations and capabilities

Fifty-eight open-source libraries are now wired into DreamLayer as optional adapters — vector databases, local Whisper, spaCy, facial action-unit models, neural VAD, Skia rendering, FastAPI, offline translation, and more — without adding a single required dependency. The core runs identically with nothing installed; each adapter upgrades one seam when its library is present and falls back to the built-in behavior when it is not. docs/INTEGRATIONS.md is the full before/after table; this chapter explains the system and the switch-on story.

The pattern: add-alongside, try-import, fall back

There is deliberately no central registry to consult and no second gating mechanism. Each adapter is a sibling file that tries its import, exposes available, and degrades to the pre-existing built-in:

• memory/vector_store.py — with sqlite-vec, indexed on-disk vector recall; without it, the exact linear cosine scan the retriever always did.
• orchestrator/asr_faster_whisper.py — with faster-whisper, real local speech-to-text; without it, the ASR seam stays a seam.
• truth_lens/au_backends.py — with LibreFace or py-feat, real facial action-unit detection feeding the Truth Lens face channel; without it, the AU frame passes through untouched.
• ai_brain/server_fastapi.py — with FastAPI, an async ASGI mirror of the same handlers; without it, the stdlib server, unchanged.

The invariant the whole system rests on, enforced in CI: the suite stays green with zero optional dependencies installed.

Highlights of the catalog

By group (each entry: installed / absent): memory — sqlite-vec, chromadb, lancedb, usearch (ANN recall / linear scan), sentence-transformers (real local embeddings / mock), mem0 (dedup and decay — live today with a built-in fallback), networkx (graph algorithms / hand-rolled adjacency); voice — silero-vad (neural VAD / energy threshold), faster-whisper (local STT / none), whisperX (word timing / none); intelligence — spaCy (real NER for commitments / regex), speechbrain ECAPA (real speaker embeddings / hash), river (online per-user learning), dowhy (causal fusion / fixed weights), diart (live diarization / none), supervision (identity-stable tracking / nearest-centroid); vision — CLIP, ultralytics, moondream, coremltools (four real classifiers behind the object-recognizer seam); privacy — presidio (ML PII detection over an always-on regex redactor), pydantic (the Veil as a type invariant: a veiled memory cannot even be constructed — live today), cryptography (Ed25519 figment signatures / HMAC); infra — rich, watchdog, zeroconf (LAN auto-discovery — baked into the Mac app), datasette, rerun; platform — pluggy, pyee, argostranslate (offline translation / text unchanged), skia-python (GPU-crisp HUD rasterizing / PIL), fastapi, exo (cluster inference), MLX (overnight LoRA on Apple silicon), frame-sdk (a Brilliant Frame display adapter — the second device), and more.

A handful are live with no install at all: the PII regex redactor, the Veil type invariant, answer validation, the pairing rate-limiter, the presence ledger, and mem0-style dedup all run in their fallback forms today.

Knowing what is on: the capability report

capabilities.py is read-only introspection over the catalog — 42 named capabilities spanning the 58 wired libraries (a capability like local ASR bundles more than one library) — it probes what is importable without executing anything, honors DL_DISABLE_<KEY> kill-switch env vars, and reports each capability as active / off / missing / unsupported / external:

python -m dreamlayer.capabilities            # the per-machine report
python -m dreamlayer.capabilities --json     # for tooling
python -m dreamlayer.capabilities --probe    # also ping Ollama / exo

The same report is an endpoint (GET /dreamlayer/capabilities) and a whole view in the Mac app, where each row carries a live status dot, an impact rating, and either a one-click Turn on / Turn off (persisted as disabled_caps, the config twin of the env vars) or a copyable pip install "dreamlayer[extra]":

Deployment profiles

Four named installs compose the groups per machine (pip install -e ".[profile-mac]"):

docs/DEPLOYMENT.md is the operator guide. A test pins the profiles and extras in pyproject.toml to the capability catalog so they cannot drift.

Capability packs — the human-sized handle

Nobody installs 58 libraries one at a time, so the Mac app offers five curated packs, each with an honest download size and an impact rating, installed with one click on a source-run Brain (the sealed .dmg app cannot pip-install into itself, and says so):

A first-run nudge on the Home view points new Brains at Total Recall; pack installs run in the background and the view polls until the capabilities flip active.