0.6.x stabilization
Proof Bundle v1 becomes tamper-evident, false completion gets an adversarial release corpus, and real-model comparisons establish a measurable baseline.
The current 0.6.x line hardens proof integrity and establishes a real evaluation baseline. Portable CI contracts and safe workspaces come before any larger delegation loop.
Proof Bundle v1 becomes tamper-evident, false completion gets an adversarial release corpus, and real-model comparisons establish a measurable baseline.
The path to 1.0
Proof Bundle v1 becomes tamper-evident, false completion gets an adversarial release corpus, and real-model comparisons establish a measurable baseline.
The same versioned proof contract becomes usable from a public API, external run manifests, and generic CI without model credentials.
Isolated worktrees, fail-closed headless policy, budgets, and durable run states make unattended execution bounded and recoverable.
Fresh-context iterations repeat one bounded task until controller-owned deterministic gates accept it, with resume and idempotency.
Adoption hardening, compatibility, selected integrations, and reproducible release evals turn the trust promise into a stable 1.0 contract.
What stays constant
A green answer should be backed by visible work, commands, evidence, or a clear unverified status.
MCP and skills expand the toolset while the runtime keeps permissions and results understandable.
Sessions, capsules, checkpoints, and memory should preserve the task state across long work.