v1.1.0

Benchmark-driven refinements (architect-reviewed): RunResult.toolCalls (backend-neutral per-run tool count), parallel gateway fan-out with partial-failure degradation, and opt-in limits.workspaceQuotaExcludes (default still counts the whole workspace so the disk quota stays honest).

v1.0.0 — first stable release

First stable release. Every roadmap feature shipped. New in 1.0: Prometheus GET /metrics export and a container image warm pool (ContainerSandboxProvider.warm() pre-pull). The claude/codex drivers are verified end-to-end against the real CLIs and the container sandbox + egress proxy against a real docker daemon. Full history in CHANGELOG.md.

v0.9.0

Domain-allowlist egress control for container runs (verified against real docker) and per-key tenant binding on the HTTP facade (a leaked key exposes one tenant, not the fleet).

v0.8.0

Multi-node scale-out: consistent-hash session→node affinity router and a gateway server (SSE proxying, fan-out lookups/cancel, aggregated stats). Copy-on-write workspace snapshots for fast session creation (build once with an optional prepare command, clone per session).

v0.7.0

HTTP facade auth (API keys), run history endpoints (GET /v1/runs, GET /v1/runs/{id}), resume-state persistence across restarts, session pre-warming, framework throughput benchmarks, and real-docker verification of the container sandbox.

v0.6.0

Core runtime (sessions, fair scheduling, sandboxing, three backend drivers), markdown harness authoring with hot reload, container isolation, run cancellation, operations layer (backpressure, retries, hooks, metrics, quotas, secret scoping), harness packs (git/npm/tarball/local) with CLI, artifact stores (local + dependency-free S3), and real-CLI verification of the claude/codex drivers including warm resume.