Reflex v0.1.0

First public release.

Added

• @reflex-ai/core: the decision runtime. Zero runtime dependencies.
  • Decision ladder: rules → exact cache → pattern table → hashed head (+ difficulty head) → full-model backend → escalate.
  • Risk classes (safe, cheap, costly, stateful, destructive) with per-class thresholds and floors. destructive is never auto-executed.
  • Modes off / shadow / assist / auto; REFLEX_DISABLE kill switch and REFLEX_MODE override.
  • Budgets (steps, $, frontier calls, wall-clock, retries) and loop detection.
  • Online learning from labels, step outcomes and task outcomes; small-sample shrinkage; audit sampling.
  • Hedged escalation (onHedge), safe speculation, decideMany batching, hard timeoutMs, fail-open everywhere.
  • Traces (JSONL, secret redaction, retention), versioned policy registry, offline training with held-out calibration, per-risk thresholds and a promotion gate with a 95% Wilson lower bound.
• @reflex-ai/server: local HTTP sidecar with token auth and a live dashboard (/ui); dependency-free MCP server (stdio).
• @reflex-ai/cli: init, doctor --tune, trace, stats, train, eval, policies, promote, rollback, serve, mcp, laya, bench, version.
• @reflex-ai/bench: ReflexBench harness with simulated coding and support suites.
• Claude Code plugin: unchanged re-read and duplicate-fetch blocking, loop stopping, destructive-command gating for bash/PowerShell/cmd, opt-in auto-approval, /reflex-status.
• reflex-laya (Python): Laya full-model sidecar; batches decisions that share a state into one forward pass; --mock scorer.
• reflex-agent-client (Python): fail-open client for the sidecar.
• Codex integration via MCP, examples, and docs.