Host repositories where humans and AI agents edit at the same time.

avcshub is the cloud home for AVCS repositories. Intent, operations, evidence, and decisions are first-class objects — your code is the materialized projection of the operation graph, not a pile of conflicting diffs.

Explore repositories →Browse the demo repo AVCS on GitHub

the operation graphstate = reduce(base, operationDAG, decisions, policy, materializer)

intentgoal · scope

sessionhuman / agent

operationsemantic Δ

evidencetests ✓

decisioned25519

reduce()

policy +
materializer

PROJECTION

Code tree

materialized state

// the model

Code is a projection, not the source of truth

Git records when the code changed. AVCS records who changed it, with what intent, on what evidence, and through which decisions — then computes the working tree by reducing the operation graph.

intent

Why

The goal, constraints, and allowed scope for a unit of work.

session

Who

A human or AI agent work episode opened against that intent.

operation

What

A semantic change with declared effects — not a raw text diff.

evidence

Proof

Machine-checkable test, typecheck, lint, and scan results.

decision

Authority

A signed resolution that accepts the operation into the tree.

// object model

Everything a repo needs is a first-class object

Governance, provenance, and verification live in the data model — not in bolted-on process. Familiar Git nouns become queries and projections over the graph.

intent

Why a change is being made — its goal, constraints, and allowed scope. Agents stay inside the lines you set.

session

A single work episode by one actor against an intent. Human or AI agent, both first-class.

operation

One semantic change with declared effects — this is the real history, not a text diff.

evidence

Machine-checkable proof attached to operations: test, typecheck, lint, and scan results.

decision

A signed resolution of a conflict or design choice. Forged or unsigned approvals are dropped.

viewwas the branch

A declarative query over the operation graph — long-lived lines without copying state.

checkpointwas the commit

A verified state vector: operations + policy + materializer, reproducible on demand.

releasewas the tag

A signed, provenance-bearing checkpoint, bundled with evidence and an SBOM.

policy

The deterministic merge rules the reducer is parameterized by — trust-weighted, auditable.

// conflict resolution

Conflicts escalate by consequence, not by line number

The reducer classifies every concurrent change. Most resolve themselves; only genuine consequence reaches a human — and when it does, the decision is signed and reusable.

L0 · L1

Auto-merge

Different entities, or disjoint line regions in the same file — merged deterministically, with no markers and no human in the loop.

Policy auto-decision

Concurrent edits overlap the same line region. The policy (human-preferred, trust-weighted) resolves it and records the rationale.

Blocked

A behavior-changing operation with no trusted evidence cannot be accepted — it stays out of the tree until proof is attached.

Human decision

A public-API break is routed to the scope's owners for a signed decision. The rationale becomes reusable history.

// using avcs

From the model to muscle memory

Humans drive AVCS from the CLI; agents drive it through the MCP server. Same graph, same guarantees — every command is just an operation on it.

acme/demo — avcs — zsh

$ avcs status

on view main · checkpoint ck-2f81 · 142 operations

2 conflicts awaiting a decision · 0 blocked

$ avcs conflicts

L2 src/checkout/cart.ts · policy: human-preferred · auto

L4 src/api/types.ts · public-API break · → owners

$ avcs decide L4 --accept --sign ed25519

✓ decision recorded · reduced → checkpoint ck-2f82

intent_create

Open an intent

Declare the goal, constraints, and scopes.

session_start

Start a session

A human or agent begins working.

operation_propose

Propose operations

Author semantic changes with honest effects.

validate_run

Attach evidence

Tests and lint gate behavior changes.

decision_record

Decide & release

Signed decisions become checkpoints.

For AI agentsMCP · 21 tools

→Drive the repo through the MCP server — 21 first-class tools, no shell scraping.
→Hold a soft WorkLease so two agents never duplicate in-flight work.
→Get a minimal RepairContext packet back when validation fails.
→Every operation carries actor identity and ed25519 provenance.

For teamsCLI · web

→avcs status · conflicts · log — read the operation graph from the CLI.
→Own L4 merges through signed, auditable decisions.
→Blame and bisect by intent: who changed what, and why.
→Branches become views, commits checkpoints, tags releases.

$ avcs open acme/demo

See the operation graph in action

Browse public repositories on avcshub, or open the demo to explore its code projection, proposal graph, and members — no account required.

Explore repositories →Open acme/demo Read the AVCS spec