claude-flow

What's new in 3.15.0

This is a MINOR release adding agenticow@~0.2.3 — Copy-On-Write memory branching — as 4 new MCP tools. No breaking changes; the new dep is in optionalDependencies with graceful-degraded fallbacks.

New MCP tools (4)

| Tool | Purpose |
|------|---------|
| agenticow_branch | COW-fork a base .rvf memory at ~162 bytes regardless of base size |
| agenticow_checkpoint | Freeze a labelled restore point that survives close+reopen |
| agenticow_rollback | Discard edits since the most recent checkpoint, O(edits) not O(N) |
| agenticow_promote | Atomically merge a branch's edits (with tombstones) into a base or target |

Read-through semantics on every branch: parent ∪ edits, child wins. Persisted via <file>.agenticow.json lineage manifest.

Motivation

v3.14.4 uncovered a 3.3 GB tarball-bloat regression where Darwin loops' git-worktree-per-agent pattern accumulated full-copy snapshots. agenticow is the structural fix — 162-byte branches instead of full copies.

Measured perf vs. published claims

Full bench harness: scripts/bench-agenticow.mjs
Findings doc: docs/agenticow/findings.md (gist mirror)

| Claim | Result | Detail |
|-------|--------|--------|
| 162-byte branches | ✅ Confirmed exact | Constant at N=1k, 10k, 50k |
| ≥3,000× smaller than full-copy | ✅ Confirmed and exceeded | 3,214× at N=1k → 181,930× at N=50k |
| "0.5ms branch" | ❌ Not reproduced | Measured ~10 ms (fixed cost, not size-proportional) |
| "83× faster" | ❌ Crossover N≈30k | Below crossover, full-copy wins |

We measured, not parroted — per CLAUDE.md's source-of-truth rule.

Architectural constraint (zero hard runtime dep)

agenticow is in optionalDependencies. When missing, every tool returns {success: true, degraded: true, reason: 'agenticow-not-found'} and exits 0. Same posture as the metaharness integration (ADR-150). Verified by the no-metaharness-smoke CI gate on every PR.

What this enables in ruflo

• Darwin loops without worktree bloat — branch-per-iteration at 162 B instead of full git-worktree clones
• Speculative agent edits with O(1) rollback — checkpoint → try → rollback if the JUDGE step rejects
• Per-user / per-session memory personalization — federation primitive for fork-and-merge
• A/B routing experiments on real memory — branch candidate routers, bench each, promote the winner

Tests

• 7/7 vitest unit tests pass (v3/@claude-flow/cli/__tests__/agenticow-tools.test.ts)
• 8/8 bash smoke contract pass (scripts/smoke-agenticow.sh)
• 108/108 CI checks green on PR #2500 before merge

Install / upgrade

npx ruflo@latest             # 3.15.0
npx @claude-flow/cli@latest  # 3.15.0
npx claude-flow@latest       # 3.15.0

Or via package.json:

{
  "dependencies": {
    "@claude-flow/cli": "^3.15.0"
  }
}

Verified npm dist-tags

@claude-flow/cli   latest=3.15.0  alpha=3.15.0  v3alpha=3.15.0
claude-flow        latest=3.15.0  alpha=3.15.0  v3alpha=3.15.0
ruflo              latest=3.15.0  alpha=3.15.0  v3alpha=3.15.0

Related

• PR #2500 — feat(mcp): integrate agenticow@~0.2.3 — COW memory branching (4 MCP tools)
• PR #2501 — chore(release): 3.14.4 → 3.15.0