Cadence docs
planuapiredacted view

ChatGPT Input - Interpreter First Wedge Operating Image

Feeder addendum documenting ChatGPT input on how the Cadence/Interpreter/Conductor system image should run.

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

ChatGPT Input - Interpreter First Wedge Operating Image

Feeder addendum documenting ChatGPT input on how the Cadence/Interpreter/Conductor system image should run.

Promotion Metadata

  • Source feedback id: 68
  • Source feedback ids: 68, 69
  • Promotion request id: 6
  • Feedback category: beneficial
  • Feedback source: [REDACTED:high-entropy]
  • Feedback created at: 2026-06-22T15:55:43.331922+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 68

  • Product: uapi
  • Category: beneficial
  • Related slug: [REDACTED:high-entropy]
  • Source: [REDACTED:high-entropy]
  • Created at: 2026-06-22T15:55:43.331922+00:00

ChatGPT Input: Interpreter First Wedge Operating Image

Status: feeder/review addendum, not final source-of-truth until merged into UAPI or Interpreter plan. Source: operator-provided ChatGPT master build prompt, digest sha256:090c689ebff637a4. Purpose: capture ChatGPT's system image for Cadence/Interpreter/Conductor/Explorer boundaries.

Core Image

  • Cadence is the locked operator-facing agent identity.
  • Interpreter implements Cadence's PA/interpreter layer and succeeds Cadence v6 lineage, but is not "Cadence v7".
  • Conductor is separate: Cadence delivers the what; Conductor executes the how across workflows, agents, models, lanes, validation, progress, and results.
  • UAI Explorer is source of truth for instructions, docs, provenance, approved contracts, and docs updates.
  • Cortex is bounded memory/product intelligence only, never unrestricted raw memory.
  • The human operator remains final product, risk, and approval authority.

Product Behavior

Cadence/Interpreter should be concise and direct; ask only blocking questions; separate facts/assumptions/unknowns/suggestions; report only real Conductor events; present short results first; never invent progress, validation, worker activity, or ETAs.

First Wedge Flow

Operator request -> Interpreter detects intent -> asks one focused question only if blocked -> builds/validates JobBrief -> submits through Conductor gateway -> receives ordered progress -> gives event-backed status -> handles operator gate -> returns operator decision -> receives ResultPackage -> presents concise result/artifacts -> records session, decisions, provenance, and outcome through Explorer.

Mock Conductor is allowed only when real contract/endpoint is unavailable or unsafe, and must implement the same gateway trait/contract.

Feedback 69

  • Product: uapi
  • Category: beneficial
  • Related slug: [REDACTED:high-entropy]
  • Source: [REDACTED:high-entropy]
  • Created at: 2026-06-22T15:55:43.362707+00:00

Engineering And Access

Rust/Axum is primary; Python is auxiliary. Stay local-first. External APIs/models are specialist capacity, not control plane. Separate Explorer profiles: Cadence-admin for approved Cadence runtime admin actions, and standard-AI for every other agent/worker. Standard AI must never invoke Cadence-only admin operations. Deny by default. Never log bearer tokens.

Research Before Schemas

Use Explorer first, then verify against repos/runtime contracts. Research Cadence v1-v6, current v6, Director primitives, Explorer APIs/instructions, UAI/UAPI contracts, Cortex bounded contracts, Conductor surfaces, and existing auth/audit/job/evidence/tool schemas. Keep evidence separate; never turn synthesis into fact.

Contracts And States

Map to existing types before adding schemas. Required concept names include JobBrief, ProgressEvent, OperatorGateRequest/Response, ResultPackage, ToolManifest, SessionRecord, DecisionRecord, and ErrorEnvelope.

State transitions: Cadence brief = Collecting -> NeedsInput -> Ready -> Submitted. Conductor execution = Queued -> Planning -> Running -> AwaitingOperator -> Validating -> Completed | Failed | Cancelled. Events require ordered sequence or evidence-backed ordering; detect duplicate, stale, malformed, and out-of-order events.

Plan Implications

  1. Add Interpreter as Cadence's implementation layer, separate from UAPI and Conductor.
  2. Keep Cadence/Interpreter out of scheduling, worker allocation, model routing, swarm management, and roadmap execution.
  3. Add Conductor gateway as handoff boundary.
  4. Add Explorer profile separation: Cadence-admin versus standard-AI.
  5. Add operator-gated first-wedge acceptance criteria and deterministic demo path.
  6. Add JobBrief/Event/Gate/Result contracts and state machines.
  7. Require Cadence v1-v6 lineage before migration decisions.
  8. Add QC loops for code, contracts, security/auth, operator UX, docs/provenance, release review.
  9. Add non-goals: autonomy, proactive scheduling, dynamic code/tool activation, self-modification, unrestricted model/swarm control, hidden background work.

Review note: Important ChatGPT feeder context. Merge only after checking live Explorer instructions and current repos. Treat as implementation image for the Interpreter/Cadence first wedge, not proof that contracts already exist.