Execution plan for superseding human-era source control
A staged plan that moves from a production-ready JJ remote toward one continuous, governed rebuild loop for humans and agents — tracked against real implementation and verification evidence.
Synced from the internal delivery index — last updated 2026-07-10
Recently shipped
Two-way native GitHub materialization sync
2026-07-07
Bidirectional native-to-GitHub sync is complete and live: scheduled auto-sync has run 16+ consecutive clean sweeps in production, including a self-healed transient failure.
Fast initial clone materialization
2026-07-02
Deployed to production and validated live on the trunk repository: cold clone time cut 2.9x (82.8s to ~28s), with byte-identical checkouts and near-zero per-file blob fetches.
Vex CDN storage spaces
2026-06-14
Org-scoped CDN spaces backed by Bunny shipped end to end; production smoke uploaded and fetched a real object through the live CDN URL.
Managed infra add-ons, Modal CI, PlanetScale Postgres
Jun 2026
The first three infrastructure roadmap items shipped and were verified against live production resources.
Delivery phases
Phase 0: Live graph foundation
Complete
100%
Updated Foundational
The monorepo layout, JJ-native control-plane primitives, backend CAS behavior, and object-store contracts are live and everything else in this roadmap builds on them.
jj-backend gRPC service, transactional SQL control plane, and content-addressed object store are in production use.
Public vision and roadmap documentation surfaces are live.
Phase 1: Production-ready remote graph
Verifying
85%
Updated 2026-07-02
Most of this phase is implemented and passing local verification. What remains is external, live production gates rather than repo-owned engineering.
CI event runner, WireGuard VPC data plane, and secure node dispatch all reached implementation-complete "Verifying" status, pending live E2E gates.
GitHub Actions-compatible remote CI with a real object-store-backed cache is verifying.
Fast clone materialization deployed to production 2026-07-02: 2.9x faster cold clones, live-validated.
Object-store garbage collection (reachability sweep, op-log horizon, pack compaction) is live in default shadow mode.
Phase 2: High-throughput agent workspaces
In progress
55%
Updated 2026-06-26
Repo-owned engineering for the virtual working copy is complete. The remaining work is external product decisions plus continued VexFS throughput tuning.
The Vexd virtual working copy is implementation-complete and MacFUSE-verified on Darwin; it is blocked only on production object-store, KMS, and cross-machine authorization decisions.
VexFS performance work has shipped a local metadata index, FUSE fast path, and range-read chunk caching; hot-path hydration and generation-aware invalidation are still in progress.
JJ-alignment for the VFS (lazy rebase, snapshot, ephemeral workspaces) is drafted and grounded against upstream jj design docs.
Phase 3: Continuous rebuild loop
In progress
65%
Updated 2026-07-10
Sync, CI status, deploy, and rollback loops are live in production. Stack-first submit ergonomics and GitHub merge-queue coexistence are still in design.
Two-way native GitHub sync is complete and live; native stacked diffs with GitHub PR projection are verifying.
The live deploy TUI and native deployment rollback are verifying, with one production rollback drill remaining.
Open-source folder publishing and guided GitHub onboarding import are verifying.
Stack-first submit ergonomics and GitHub PR/merge-queue coexistence remain in draft design.
Phase 4: Scale and supervised autonomy
Early
20%
Updated 2026-06-26
Guardrails for true multi-tenant scale are mostly still in the design stage, with a few foundational pieces already live.
PlanetScale Vitess managed MySQL is deployed to production and smoke-tested, blocked only on a live backup/restore drill.
User-uploaded images and organization logos are verifying in production.
Bookmark protections, per-tenant history backups, and edge DDoS mitigation are grounded PRDs awaiting implementation.
CI automation identity (vex-action) and managed Docker remote builders are drafted.
Delivery principles
Keep each stage independently shippable and useful to humans before raising autonomy.
Prefer explicit consistency guarantees over opaque heuristics or prompt-only policy.
Move from source-of-truth docs to source-of-truth config, manifests, leases, and graph events.
Measure every optimization: cache hit rate, operation latency, remote API tail, and loop completion time.
Every automated action should be explainable from actor identity, policy, graph state, and runtime signal.
How to read "Verifying"
A phase or plan marked Verifying means the implementation and local test suites are complete; what remains is a live, external gate — a production deploy, a real third-party credential, or a signed-in UI smoke test — before it is marked Complete.