0.2.0
Universal LayerStack
Portable root identity and native materialization
Back to roadmap overview
04 · Delivery plan
Evidence-gated migration, not a flag-day rewrite.
0.2.x is divided into dependency-ordered release slices. Every slice leaves the current runtime usable, adds one durable contract, and must pass correctness, recovery, portability, and performance gates before the next layer relies on it.
- Cadence
- Gate-driven
- Migration
- Incremental
- Compatibility
- Existing sessions preserved
- Release rule
- Evidence before promotion
Delivery strategy
Build the durable spine before the orchestration surface.
The critical sequence starts with runtime-neutral roots and atomic checkpoint publication. Graph APIs, rollout scheduling, and heterogeneous executors layer on only after the state invariants are proven.
0.2.1
SandboxGraph Runtime
Durable nodes, attempts, OCC, leases, recovery
0.2.2
Rollout Fabric
Bounded scheduling, evaluators, trajectories, idempotency
0.2.3+
Sandbox Fabric
OCI/Linux + WASI adapters and locality-aware workers
Patch labels communicate ordering, not a promise that every capability lands in exactly one patch. Exit evidence decides the boundary.
Release slices
Each slice has deliverables, migration work, and a hard exit gate.
0.2.0
Universal LayerStack
Deliverables
- Runtime-neutral RootId and logical tree manifests
- CAS object integrity and provenance schema
- Worker-local native root materialization and verified cache
- OCI command path with no per-read reconstruction
Migration and compatibility
- Read existing LayerStack roots through a compatibility adapter
- Dual-write new portable metadata while current session APIs remain authoritative
- Backfill provenance lazily when an existing root is selected
Exit gate
The same root verifies identically through the portable manifest walker and native OCI projection; a cold root hydrates once and hot reuse adds no reconstructed read path.
0.2.1
SandboxGraph Runtime
Deliverables
- SandboxGroup, SandboxNode, ExecutionAttempt, and Evaluation records
- Fork, activate, suspend, retry, inspect, promote, merge, and prune
- Lease/journal recovery and bounded graph-aware GC
- Private writable delta per active attempt
Migration and compatibility
- Wrap current workspace sessions as execution attempts
- Mirror successful publication into sealed nodes
- Keep old session identifiers as aliases during the compatibility window
Exit gate
Crash and race tests prove atomic sealing, sibling isolation, zero lower payload duplication for hot forks, stale-promotion rejection, and lease-safe GC.
0.2.2
Rollout Fabric
Deliverables
- Quota- and locality-aware placement leases
- Versioned evaluator, policy, seed, trajectory, and reward records
- Idempotent retry and exactly-once accepted backpropagation
- Frontier retention and subtree pruning policies
Migration and compatibility
- Adopt graph operations first in internal rollout coordinators
- Store search statistics beside—not inside—immutable checkpoint manifests
- Allow existing direct session orchestration as a fallback client
Exit gate
A large graph runs through a bounded worker pool; forced retries do not double-count outcomes; inactive nodes consume no process or mount residency; recovery preserves frontier leases.
0.2.3+
Ephemeral Sandbox Fabric
Deliverables
- OCI/Linux and WASI workspace adapters
- Capability-aware worker registration and placement
- Pinned environment semantics and executor attestation
- Cross-backend observability and result sealing
Migration and compatibility
- Keep OCI/Linux as the native compatibility path
- Introduce WASI only for workloads whose syscall contract fits
- Schedule by declared requirements rather than guessing backend equivalence
Exit gate
A graph selects supported backends reproducibly from declared capabilities; cross-backend logical roots verify equivalently; no target image supplies LayerStack utilities or durable checkpoint truth.
Validation program
One evidence matrix follows every release candidate.
The suite separates algorithm viability from environment setup. Every failure is classified as semantic, recovery, compatibility, capability, performance, or harness/setup before a storage approach is rejected.
| Lane | What it proves | Representative evidence | Blocks release |
|---|---|---|---|
| Correctness | Isolation, immutable roots, OCC, merge and promotion semantics | Concurrent sibling writes, conflicting publishes, manifest equivalence | Yes |
| Recovery | Journal, lease, seal, retry, evaluation and GC idempotency | Crash injection at every durable transition | Yes |
| Portability | Supported Docker hosts/images and WASI logical equivalence | Ubuntu Linux/arm64 and amd64 containers; minimal target images | Yes |
| Capability | Explicit behavior with and without required executor privilege | Mount probe, namespace probe, fail-closed diagnostics | Yes |
| Performance | Fork, activation, command I/O, seal, hydration and GC costs | Median, tail, bytes allocated, page cache, CPU and wall time | Regression budget |
| Scale | Bounded residency under large durable graphs | Frontier width, history depth, retry storms, worker loss | Yes |
Benchmark contract
Measure the real resource curve, not only wall-clock latency.
Fork
Inputs
workspace size · history depth · fan-out
Measures
metadata bytes · payload bytes · latency
Activate
Inputs
hot/cold root · image size · backend
Measures
hydration bytes · mount/start latency · CPU
Execute
Inputs
read/write mix · file count · working set
Measures
throughput · p50/p95/p99 · page-cache impact
Seal
Inputs
changed bytes · path count · metadata churn
Measures
capture time · CAS writes · manifest size
Recover
Inputs
crash point · graph size · lease count
Measures
recovery time · orphan count · integrity violations
Collect
Inputs
reachable set · pruned subtrees · cache pressure
Measures
reclaimed bytes · pause time · false reclamation
Required environment matrix
| Docker Desktop host | macOS arm64 | Linux VM / arm64 | OCI fast path |
|---|---|---|---|
| Native Linux host | Ubuntu arm64 | Linux / arm64 | OCI fast path |
| Native Linux host | Ubuntu amd64 | Linux / amd64 | OCI fast path |
| Minimal target image | Any supported Docker host | Linux / matching arch | No helper utilities |
| WASI runtime | Host-agnostic runtime | wasm32-wasi | Capability-scoped filesystem |
Risk register
Risks are retired with evidence, not hidden by orchestration.
| Risk | Signal | Mitigation | Gate owner |
|---|---|---|---|
| Mount capability varies by runtime | Overlay probe fails or produces misleading EPERM/EINVAL. | Preflight capabilities, fail closed, keep privilege inside executor boundary, document supported runtime contract. | OCI executor |
| Portable manifests leak local filesystem identity | Cross-worker validation differs on inode/path/whiteout semantics. | Canonical schema, backend conformance corpus, no local paths or inode numbers as identity. | State plane |
| O(1) fork shifts cost into activation | Cold or deep roots hydrate repeatedly and dominate rollout time. | Verified worker cache, locality-aware placement, measure hot and cold curves separately. | Scheduler |
| GC races with selection | A frontier node loses an ancestor or object during activation. | Lease ancestor closure, epoch/reachability barriers, adversarial concurrent GC tests. | Storage |
| Retries corrupt search statistics | Rewards or visits increase more than accepted rollout count. | Idempotency keys, append-only acceptance records, transactional backpropagation. | Rollout service |
| WASI semantics are over-promised | Workload depends on syscalls, links, modes, or process behavior outside contract. | Capability declaration, compatibility tests, explicit OCI-only workload classification. | Executor API |
| Compatibility layer becomes permanent | Old session-only paths gain new features or bypass graph provenance. | Deprecation telemetry, one-way feature policy, removal gate after client migration. | Product API |
| Benchmarks hide setup defects | Approaches fail before exercising their algorithm. | Classify setup vs semantic failures, publish probes and artifacts, require self-review before rejection. | Release engineering |
Definition of done
0.2.x is complete only when the system—not the diagram—has these properties.
- One durable SandboxGroup graph serves multiagent development, RL rollout, and MCTS orchestration.
- Hot local forks add O(1) metadata and zero duplicated lower-directory payload bytes.
- The live worker pool remains bounded while durable history grows independently.
- OCI/Linux execution uses native filesystem I/O over a LayerStack-backed immutable lower plus a private writable delta.
- WASI reuses RootId, manifest, CAS, node, evaluation, lease, and provenance contracts without Linux mount assumptions.
- Crash recovery, retry, promotion, backpropagation, and GC are idempotent under fault injection.
- The compatibility path is documented, instrumented, and removable rather than becoming a second architecture.