All runbooks
planuapiredacted view

UAPI Infrastructure Redesign

Reviewed UAPI infrastructure redesign plan submitted through the AI feedback lane.

Updated 2026-06-22·Freshness: Reference·Sensitive tokens or credentials were hidden before display.

UAPI Infrastructure Redesign

Reviewed UAPI infrastructure redesign plan submitted through the AI feedback lane.

Promotion Metadata

  • Source feedback id: 66
  • Source feedback ids: 66
  • Promotion request id: 1
  • Feedback category: beneficial
  • Feedback source: grok-explorer-session-2026-06-22
  • Feedback created at: 2026-06-22T14:39:39.304602+00:00
  • Target slug: [REDACTED:high-entropy]
  • Review state: review
  • Reviewer note: AI lane auto-promoted redacted feedback into internal catalog context.

Source Feedback Body

Feedback 66

  • Product: uapi
  • Category: beneficial
  • Related slug: uai-doc-docs-plans-explorer-2.0-ux-blueprint
  • Source: grok-explorer-session-2026-06-22
  • Created at: 2026-06-22T14:39:39.304602+00:00

3801

created_at: 2026-06-22T15:10:00Z source_of_truth_status: draft namespace: uapi kind: plan tags: [plan, uapi, infrastructure, cadence, conductor, redesign]

UAPI Infrastructure Redesign (Full Plan + Tree from Session)

Agent: Grok (full session incl. GPU Cloud Architect + Codex + user clarifs) Product: uapi | Authority: plan only (AI lane) **Explorer v0.2.3 public_safe=true server-redaction | Grounded in primary contexts below

Fetched Primary Contexts (audits): - feedback-convention (515) - documentation-source-of-truth (518) - ai-memory-knowledge-beacon-mesh (517) - content-promotion-workflow (520) - explorer-2.0-redaction-embedding-lanes (521) - explorer-2.0-ux-blueprint - agents, rules, sandbox, public-safe, dev-standards, qa-gates

Strict Rules: AI lane only. No local copies. Intake via this feedback for uapi. Promotion request via uai-explorer feedback (Explorer catalog workflow only). Do not claim live until /api/ai/context/slug succeeds. Cortex=cataloged memory only (no raw).

Tree/TOC (1-24 from session agent replies)

  1. Executive: UAPI = unified central control (auth/routing/compute/embeddings)
  2. Current problems: Cortex overload, separate logins, no dedicated Front Desk, conflated Cadence role
  3. Principles: Explorer SoT, public-safe redaction, catalog memory, unified control, AI-visible only
  4. Layers: L0=UX(Cadence+Explorer+Cortex), L1=UAPI central, L2=Schedulers, L3=Serving, L4=Fleet
  5. UAPI scope: login/auth, OPA policy, auto-route (cloud/local GPU/NUC), compute, embeddings jobs
  6. Cadence (user clarified): separate operator PA/Front Desk/voice controller/entity mgr. Modular over APIs/MCPs. NOT scheduler/ infra owner
  7. Conductor: factory/builder (creates agents/tasks from plans)
  8. Unified login: single auth across Explorer/Cortex/UAPI/Cadence. Zero separate logins
  9. Routing: UAPI owns intelligent placement across heterogeneous resources + priority lanes
  10. Embeddings: move from Cortex -> UAPI (per ai-memory + redaction-embedding-lanes rules: sanitized tiered AI-visible packets only)
  11. Priorities: critical-gang-training, prod-shared-inference, interactive, dev/best-effort, burst (Kueue CQ + flavors)
  12. Schemas central in UAPI: Device/WR/Policy/RunRecord + immutable audit
  13. Migration: facade for dual legacy-Cadence-router + new UAPI; parallel; deprecate later
  14. Critiques fixed: prior UAI2 over-scope; Cortex doing non-memory work; missing PA separation
  15. Risks: storage pressure, ephemeral Colab, dual-path, Intel Arc maturity, operator gates
  16. Governance: Explorer catalog SoT; promotion via workflow; operator packets; lineage
  17. Verification path: ready+bootstrap+catalog+context; sandbox first
  18. For other AIs: 1. ready 2. bootstrap 3. catalog 4. context/{this-slug} 5. read related plans+memory+convention
  19. Open: full device registry, Cadence exact contracts, Intel flags
  20. Impl start: UAPI facade + auth + routing core; promote this plan
  21. Safety: server redaction only; public-safe; no secrets ever in catalog
  22. Evidence: audits listed above + ready catalog_count=67
  23. Done when: slug in catalog + fetchable context + feedback recorded for both products
  24. Next: uai-explorer promotion request; operator review to catalog kind=plan slug= [REDACTED:high-entropy]

Key Details

UAPI unifies what was fragmented. Cadence is the PA layer (Front Desk), Conductor the creator, UAPI the single API surface for everything infra (incl embeddings relocated from Cortex memory service). Priorities + unified auth ensure fair, simple access. All grounded in fetched Explorer contexts. This is the complete session redesign tree+content for intake.

Submit via feedback only. Promote through uai-explorer content/catalog to exact slug for AI visibility (Codex etc).