Skip to content

README.md

Latest commit

 

History

History
104 lines (80 loc) · 5.25 KB

README.md

File metadata and controls

104 lines (80 loc) · 5.25 KB
Raw

NiakVIO Provider Engine V2 — ARCHI2

engine_v2/ is the canonical production control plane for NiakVIO. Since the ARCHI2 cutover, it is no longer an isolated experiment: publication, provider catalog preservation, runtime repair planning, evidence and device-contract validation are coordinated around this architecture.

Production pipeline

3 upstreams + published LKG
        |
        v
provider discovery / sibling selection
        |
hubs / DNS / public address discovery / LKG
        |
        v
canonical ProviderSpec + staged provider bundles
        |
        v
NiakVIO Brain
classify cause -> choose bounded skill -> probe -> validate -> learn/defer
        |
        v
strict identity + media validation
        |
        v
Mobile / Desktop / TV evidence
        |
        v
canonical provider catalog -> general + VF manifests
        |
        v
atomic versions / integrity / LKG publication

no_streams, blocked, unavailable, provider_unreachable and runtime errors are observations, not terminal repair decisions. Quarantine remains reserved for positive safety/identity evidence.

NiakVIO Brain

The Brain is the repair orchestrator. It is intentionally separated from the validators: it may propose/create/adapt repair skills, but it cannot relax the rules that decide whether a result is safe and playable.

Production world

  • classify the first causal gap rather than patching from a global status;
  • prefer a healthy sibling when it already proves the canonical provider;
  • reuse a learned skill when compatible;
  • otherwise select a targeted built-in skill;
  • allow validated production discoveries to become learned skills immediately;
  • promote a skill to cross-provider auto-application only after repeated success and enough provider diversity;
  • keep Quick bounded and give unresolved providers another chance on the next cron;
  • LKG is a safe fallback after the repair budget, never a fake diagnosis or a blind upstream/default-provider shortcut.

Back doors / stop conditions

The Brain must stop and emit deferred_retry when it detects a loop, exceeds its mutation/time/generated-code budget, or would need to rewrite its own core during production. A later cron can retry with new evidence.

The production Brain can evolve skills and provider-facing primitives. Core invariants (classification contract, validation boundary, budgets, fail-closed publication) are proposal-only from production so a difficult provider cannot teach the Brain to weaken its own tests.

Learning Lab world

A daily bounded Learning Lab runs separately from publication. It can explore failed patterns, compose candidate skills, propose primitive improvements and suggest core changes. Its outputs are artifacts and, when useful, a reviewable PR. It cannot publish manifests/providers or mutate production state.

The Learning Lab intentionally persists only generalized technical knowledge. Raw URLs/tokens/header values, private notes and source spreadsheet/prompt text are excluded from proposals.

Failure classes

The Brain distinguishes at least:

  • not_invoked
  • dns_unreachable
  • transport_blocked
  • search_gap
  • identity_mismatch
  • detail_gap
  • episode_gap
  • player_gap
  • media_extraction_gap
  • playback_context_gap
  • media_validation_gap
  • runtime_contract_drift
  • unknown_failure
  • healthy

A provider that has no usable API can still be repaired through the HTML/player path: catalogue/search -> detail -> embed/player -> scripts/XHR/JSON -> final HLS/DASH/direct media, while preserving Referer/Origin/cookies when required.

Domain / hub recovery

Known hubs, redirects, Telegram public pages, direct candidates and historical LKG routes are tried first. If those routes fail, the bounded search fallback can discover a provider address source through a public search engine; a Telegram result is opened, the latest relevant message id is preferred, and the announced terminal is validated before it can update routing state.

Device Lab visual sub-lab

The Native corpus Device Lab remains the real Mobile/Desktop/TV proof. A visual sub-lab renders its sanitized logs into JSON/HTML timelines so invocation -> provider result -> transport can be inspected quickly. Internal catalogue/detail/player stages are shown only when actually observed; the visualizer never invents missing hops or persists raw stream URLs/tokens/header values.

Non-negotiable rules

  1. Published LKG remains authoritative when a candidate is not demonstrably better.
  2. Upstreams are knowledge sources, not automatically trusted final bundles.
  3. Domain discovery precedes structural provider mutation.
  4. Broken providers are repair inputs; safety quarantine requires positive evidence.
  5. A healthy sibling is a useful canonical source, not an excuse to stack endless patches.
  6. Every accepted repair requires real playable evidence and no positive identity/duration contradiction.
  7. File extension alone is never final-media proof.
  8. Movie, TV and anime are first-class regression dimensions; Breaking Bad remains a mandatory TV fixture.
  9. Mobile/Desktop/TV contracts are separate evidence dimensions around one canonical request/result model.
  10. Versions/cache identifiers are finalized atomically with publication so changed provider bytes cannot keep stale Nuvio identifiers.