- one portable Agent Skill packaged for Codex and Claude Code
- deterministic project inspection and generated Research Profile
- one plan-hash campaign approval
- isolated local CPU/light experiments
- authoritative metric parsing, compatibility checks, ledger, and recovery
- multiple local campaigns with immutable approved base revisions
- explicit v0 migration check and read-only legacy preservation
- scored hypothesis candidates and deterministic exploit/explore/confirm selection
- logical DAG provenance with scoped evidence and recombination metadata
- separate parent, champion, target, and identical-tree confirmation semantics
- approval-bound Strategy Contract with project-shape rationale
- diagnostic, balanced, and optimization Selectors behind a registry
- deterministic, pre-approved Selector transitions
- structured hypotheses and append-only evidence assessments
- unchanged schema v1 campaign execution and recovery
- validate evidence-guided DAG campaigns against a real optimization repository
- add richer candidate, strategy, hypothesis, and campaign visualization
- validate against a real retrieval/RAG repository without introducing a required retrieval module
- add richer evidence-based command discovery and profile confidence reporting
- support configurable seed scheduling and statistical confirmation
- integrate optional Bayesian Optimization and multi-fidelity Selectors
- add explicit user-approved GPU resource contracts
- add remote execution providers only behind the same Execution Contract boundary
- optional research knowledge packs as hypothesis-generation inputs
- user-driven promotion/merge workflows
- campaign visualization and multi-project summaries