v0.2.0 — Temporal Branching

AOSP v0.2.0 — AI-Operable Software Protocol

What's New

Temporal Branching — fork execution into multiple parallel timelines, score all outcomes, commit only the winner.

• Fork one intent into multiple competing strategies
• Shadow-execute all timelines in parallel
• 6-dimension weighted scoring (confidence, safety, minimality, completeness, speed, reversibility) with auto-normalizing weights
• Risk-aware ranking with 3-level tiebreaker (score → risk → reversibility)
• Reality Merge Requests with quorum-based approval and auto-expiry
• Pairwise timeline diffs with per-dimension breakdown
• Counterfactual analysis ("what if we'd picked the other timeline?")
• Prevented futures reporting — tracks blocked dangerous timelines
• Non-selection proofs — SHA-256 hashed evidence chain of why timelines lost
• Branch budgeting (timeline count, fork depth, total intents, wall-clock time)
• 4 pruning strategies (score_threshold, confidence_decay, risk_ceiling, diminishing_returns)
• Commit-time revalidation (re-check reality drift before merging)
• Sandbox mode (explore futures without mutations)
• Future search engine (query/filter by score, risk, confidence, phase)
• Custom safety gate checks (register dynamic validation functions)
• Action diffs with content preview and size estimation
• Auto-strategy inference from intent analysis
• Exploration statistics (avg score, std dev, risk distribution, budget utilization)

Core Platform (from v0.1)

• Shadow-first execution — preview every action before committing
• 10-step kernel lifecycle — identity → state → risk → policy → blast radius → shadow → approval → commit → state update → trace
• Policy DSL — declarative rules (DENY WHEN action MATCHES "delete_*")
• Blast radius analysis — graph-based impact analysis with transitive dependencies
• Approval gateway — human-in-the-loop for high-risk actions
• Merkle trace chain — tamper-evident audit trail with per-entry proofs
• Drift detection — detect unauthorized out-of-band changes
• Transaction coordinator — atomic multi-step execution with ordered rollback
• Pipeline engine — multi-stage pipelines with conditional branching
• Filesystem adapter — reference implementation with full shadow support

Test Suite

77 tests across 11 test files covering all modules: identity, policy, DSL, approval, blast radius, Merkle chain, state, drift, filesystem adapter, temporal branching, and end-to-end integration.

Stats

• ~8,400 lines of TypeScript across 31 source files
• 3 runtime dependencies (chalk, commander, uuid)
• MIT licensed