Documentation

The canonical documents live in the Bivdi repository. The pages on this site are a readable presentation of the same material; where they disagree, the repository is authoritative.

Start here#

• README

  The authoritative synthesis: goals, the model, decisions and open questions. Read this first.

• ARCHITECTURE.md

  The technical elaboration — layers, capability model, object store, IPC, drivers, boot chain, performance targets.

• Specification v0.1draft

  The model any implementation must satisfy, independent of the substrate. Normative for the decided items.

• Decision log

  The single source of truth for what is and is not decided.

• Glossary

  Ambient authority, attenuation, confused deputy, crypto-shredding, petname, powerbox, workload, and the rest.

Security#

• Threat model

  Assets, the trusted computing base, adversaries and mitigations, and what is explicitly out of scope for v1.

• Capabilities

  The heart of the security model — properties, the two tiers, the powerbox, leases.

• Durable authority tokensproposed

  Authority at rest and in transit. Carries no normative weight until an RFC is accepted.

• Provenance and audit

  The append-only log of authority — never content. The hash chain and signed checkpoints it specifies are not yet implemented.

• Boot and attestation

  Measured boot, sealing storage to the boot state, and telling an update apart from a tamper.

• Identity

  Identity kinds, petnames, selective disclosure.

The model#

• Object store

  Blob, cell and catalog; content addressing, immutability, transactions, snapshots, crypto-shredding.

• State engine

  Desired state, reconciliation, immutable generations, transactional activation.

• Events

  Typed, first-class events and the shared correlation identity.

• Resources and scheduling

  Workloads, budgets, and time as a capability.

• Workload manifestdraft

  The security boundary in human-readable form. Schema and encoding are open.

• ABI and interfacesdraft

  What the ABI must do. Which IDL, and its encoding, is open.

The system#

• AI agents

  Agents as constrained citizens; the delegation model.

• Drivers and hardware

  Driver strategy, the VM-first target, hardware tiers, isolation.

• Networking

  Identity-based networking, flow capabilities, distribution.

• Compatibility

  The four-level model, and the cannibalisation response.

• Observability

  Built-in metrics, tracing and correlation identity.

• Performance

  First-class performance and the proposed reference targets.

• Recovery and migration

  Declarative recovery; no machine-state persistence.

Project#

• Roadmap

  Milestones, deliverables, exit gates, dependencies and risks.

• Licensing

  The open-core model in full.

• Contributing

  Issue-first workflow, branch naming, Conventional Commits, CI-gated merges, and the RFC process.

• Security policy

  How to report an issue against the design.

• RFCs

  Where every non-trivial design decision is recorded. Four so far: the target niche and the interface definition language (both accepted), the platform guarantees matrix and the kernel choice (both deferred).

Further reading behind the design#

Bivdi reuses more than it invents, and the sources are worth reading directly.

• Klein and colleagues, seL4: Formal Verification of an OS Kernel (SOSP 2009)
• LeVasseur and colleagues, Unmodified Device Driver Reuse via Virtual Machines (OSDI 2004)
• Waldo and colleagues, A Note on Distributed Computing (1994)
• Hardy, The Confused Deputy (1988)
• seL4 · WASI · Fuchsia · Genode · Qubes OS · NixOS · Firecracker