forgo.cloud
Sign in
Repo workspace

forkjoin-ai/gnosis

distributed-inference-host tests

distributed-inference-host/__tests__/README.md
forkjoin-ai/gnosis
Repo overview Open in workbench

distributed-inference-host tests

Parent: distributed-inference-host

Focused Vitest coverage for the host shim, station transports, pipeline orchestration, carrier medium/image adapters, formal helper surfaces, and the agentic superserver MCP chat wrapper.

Getting Started

  • What: test fixtures and validation notes for the tests area.
  • Why: it shows which behavior is locked by tests instead of only described in prose.
  • How: read the local test map, then run the narrow package or crate test command named by the parent README.
  • Next: use these files when changing behavior in the sibling source directory.

tts-audio.test.ts covers the local OpenAI-compatible speech endpoint helpers: audio mode resolution, deterministic WAV encoding, and invalid request handling without touching a real host audio device.

openai-agentic-mode.test.ts covers the bare-by-default Moonshine chat routing contract, including X-Zedge-Agentic, body fallback fields, env defaults, and optional MCP URL overrides.

thoth-npc-halogram.test.ts covers the shared-ui Halogram/metacognition adapter used to convert NPC, halogram, and hologram voice/body summaries into bounded Thoth prompt metrics.

style-transfer-any-image.test.ts, style-transfer-adain-adapter.test.ts, and forkRaceFoldStyleTransfer.test.ts cover the host-side style-transfer surface: classical transfer output constraints, AdaIN adapter contract behavior, and fork/race/fold candidate ranking and ceiling rejection.

style-transfer-golden-png.test.ts compares deterministic 16×16 classical outputs to committed PNG bytes under golden/style-transfer/ (regenerate with UPDATE_STYLE_TRANSFER_GOLDEN_SNAPSHOTS=1).

style-transfer-frf-golden-png.test.ts guards the weighted FRF style fold on the same fixtures (frf-mesh-fold-16.png).

frf-race-speed.bench.test.ts benchmarks the race-speed payoff of the ternary fork verdict (Accept/Abstain/Decline) over the legacy sawRejection boolean, both built on the real forkRaceFold. The bool forced a complete-result consumer to re-run the whole fork batch on any rejection; the trit's Abstain (partial survivors) ships the fold and backfills only the failed slots. Seeded N×failure sweep, real concurrent wall-clock. Finding: the binary whole-batch retry blows up geometrically (~`1/(1-p)^N`) — at N=16, p=0.20 it costs 22x the fork-work and 15x the wall-clock (27 rounds vs ~2); the policies tie only when nothing rejects. Asserts ternary work/rounds never exceed binary and ≥1.5x at the stressed corner.