Platform Internals
Detailed documentation of how Hoziron's internal subsystems work — agent execution, workflow orchestration, the mediated ingress/egress seams, tool dispatch, PII policy, and provider routing. Grounded in crates/platform/hoziron-core/src/, its two sibling extraction crates crates/platform/hoziron-core-package/ (package build/init/inspect/lint/transform/validation/verify logic) and crates/platform/hoziron-core-audit/ (audit config/emitter/export/hook-handler/unified-chain logic), the pure-data-type crate crates/platform/hoziron-core-types/, and the ADRs each document links to. hoziron-core re-exports everything from the two logic-extraction crates at its usual crate::package::.../crate::audit::... paths, so existing internal references still resolve even where the implementation now lives in a sibling crate.
These documents are for engineers and technical evaluators who need to understand runtime behavior for debugging, capacity planning, or advanced configuration. They describe the actual implementation, not just the API surface, and are more technical/conceptual than the task-oriented docs/guides/.
Documents
| Document | Covers |
|---|---|
| agent-execution.md | Message flow, the kernel dispatch seam, LLM loop, tool execution + outbound mediation, memory recall, session management |
| workflow-engine.md | Multi-step orchestration, step modes, mediated egress + billable value events, tracing spans, run lifecycle |
| competency-system.md | Multi-competency composition, contract-based integration resolution, tool visibility, pre-invocation reshaping |
| invocation-model.md | Unified invocation layer, source types, triggers, mediated ingress authority, rate limiting, idempotency |
| provider-routing.md | Closed provider/model inventory, driver-declared sovereignty, preference resolution, circuit breakers |
| permission-model.md | The agent-level Permission enum, enforcement points, child agent inheritance — distinct from RBAC |
| security-auth.md | Auth modes, API keys, RBAC (6 roles), OIDC/JWT, brute force protection, TLS, credential handling |
| pii-data-protection.md | PII pipeline, the unified carrier PII policy, the two mediated egress/ingress seams, memory isolation, unified audit chain |
| catalog-packages.md | Package format, the lifecycle trait contract, contract-based integrations, dependency resolution, multi-registry, signing, taxonomy |
Architecture Overview
There is no Trigger Registry — agents do not declare any
invocation-gating config at all. Channel invocations are gated by
AgentBinding/router config plus per-channel RBAC upstream; workflow
invocations are pinned by the workflow step's own agent reference; Cron
and API invocations are platform-managed. See
invocation-model.md for the current ingress model.