system-prompts-and-models-of-ai-tools/docs/governance/planes-and-runtime.md

Planes and runtime guidance

Canonical: MASTER_OPERATING_PROMPT.md (same rules; this file is navigable detail).

Three governing layers (stack posture)

1. Exploration intelligence — discovery, analysis, triage, scenarios, recommendations, decision memos, risk synthesis, forecasting. Outputs are structured (schema-first) with provenance and freshness.
2. Committed execution — the only layer that may create durable commitments, long-lived workflows, and external business actions (signatures, rollouts, partner flows, release promotion, PMI steps, etc.).
3. Trust fabric — policy, approvals, authorization, audit, security gates, tool verification, evidence packs, model/provider governance, traceability, evaluation.

Primary rule: AI may recommend; systems commit; humans approve critical decisions.

Five planes (reference architecture)

Plane	Owns	Must not
Decision	Cognition, memos, structured outputs, scenarios, review interrupts	Durable external side effects as “narration only”
Execution	Deterministic workflows, retries, compensation, idempotency, versioning, external commitments	Unstructured trust-me execution
Control	Policy, approvals, RBAC/ReBAC, secrets, environment promotion, audit, release gates	Ad-hoc rules living only in prompts
Data	Operational truth, semantic metrics, embeddings, contracts, lineage, quality, ingestion	Conflicting definitions of the same business metric
Trust	Verification, security evidence, model evaluation, provenance, freshness, reversibility	Claims without proof

Rule: cognition loops belong in the decision layer; commitments belong in the execution layer.

Technology mapping (when to use what)

• Repo-native coding agents — repository inspection, edits, refactors, tests, release prep, architecture discovery.
• Agent runtimes — analysis loops, memos, recommendations, structured outputs, tool-based exploration, review interrupts.
• Stateful graph runtimes — persisted cognition, checkpoints, interrupts, resumable reasoning, human-in-the-loop pauses.
• Durable workflow runtimes — flows lasting hours/days/weeks; retries, rollback, compensation; cross-system commitments; must survive crashes, restarts, and deployments.

Long-running business workflows must not live only inside ephemeral agent graphs without a durable execution owner.

Dealix code pointers (illustrative)

• Agent routing / execution: salesflow-saas/backend/app/services/agents/ (router.py, executor.py).
• Graph-style flows: salesflow-saas/backend/app/agents/, salesflow-saas/backend/app/flows/.
• Policy direction in code: salesflow-saas/backend/app/services/dealix_os/ (e.g. policy engine) — keep policy out of markdown prompts.

See also: approval-policy.md, trust-fabric.md, discovery-and-output-checklist.md.