claude-flow

Turns agent execution into a governed learning loop with provenance on every decision.

New

• agenticow COW memory substrate (ADR-170) — 9 MCP verbs (branch/checkpoint/rollback/promote + ingest/query/diff/lineage/status), 162-byte branches, nativeAnn Rust fast-path, one shared loader
• Per-agent COW branches (opt-in), speculative branch-and-promote (agenticow_speculate), checkpoint/rollback gate for autopilot loops
• Provenance-tiered evaluation oracle (ADR-171) — execution-oracle → Fable judge → structural proxy, each label tagged resolvedBy; promote-gate (proxy can NEVER clear); causal failure receipts
• Fable advisor harness (ADR-172) — cost-disciplined claude -p (clean-cwd + batching = $1.56→~$0.02/item); judge + GEPA reflector
• weight-eft $0 distill slice + remote-GPU train (ADR-173) — export/plan/eval + neural distill train --remote (OFFLINE dry-run default, --execute --yes gated, --preflight opt-in)

Safety (adversarial RC pass — all fail-closed)

Promote unresolved branch → ineligible · proxy-as-gold → can't clear · SSH without --execute --yes → offline/refused · Fable without opt-in → no call · rollback after partial write → restores.

Bundles: agenticow 0.2.4, @ruvector/ruvllm 2.6.0. ADR-150 optional-deps + graceful degradation throughout. Startup 0.12s. New CI lockfile-drift guard (#2563).

PRs: #2562 · ADRs 170-173