🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
Managed dynamic workflows validation
This records the initial adapter stage. See the console and verified learning validation for the latest expanded integration results.
Date: 2026-10-24. Baseline: 973d0ab497c9138d7bee1dc3c7a4d3985e37bf9e.
Scope: existing Managed Agents REST adapter, incremental event reducer,
plugin guidance, and isolated CI contract. No dependency or lockfile changes.
Results
| Check | Result |
|---|---|
| Provider fixtures and existing no-key tool contract | 87 tests passed |
| Adjacent persisted workflow validation/resume/ownership | 21 tests passed |
| Strict TypeScript check of adapter and imported service | Passed |
| ruflo-agent plugin smoke | 16 passed |
| ruflo-workflows plugin smoke | 18 passed |
| CLI/MCP registration audit | 413 registered, zero dangling references |
| Tool discoverability audit | 411 scanned, zero missing guidance/short/duplicate descriptions |
| Ruflo 3.25.6 deep code security scan, adapter and seven hook files | Zero scanner findings |
| Independent source review | Six confirmed defects corrected and covered by fixtures |
| Whitespace/diff integrity | Passed |
Scanner coverage is heuristic and scoped. It is not a dependency advisory audit or proof of security. No dependency changes were introduced. Existing whole repository dependency exposure is outside this patch's assessment. Advisory feeds were not queried. The threat model and specific attack regressions are in ADR 0003.
The independent reviewer identified stale final-idle completion, interruption completion, valid confirmation result unions, historical-run contamination, missing run IDs and nested usage accounting. All six have regression fixtures.
Reproduction
With repository development dependencies installed, from v3/@claude-flow/cli:
npx vitest run --config vitest.config.ts \
__tests__/managed-agent-tools.test.ts \
__tests__/managed-workflows.test.ts \
__tests__/managed-mod-integration.test.ts \
__tests__/workflow-validation-runtime.test.ts \
__tests__/workflow-resume-dispatch-r50.test.ts \
__tests__/workflow-progress-ownership-r50.test.ts
From the repository root:
npx tsc --target es2022 --module nodenext --moduleResolution nodenext \
--strict --skipLibCheck --types node --noEmit \
v3/@claude-flow/cli/src/mcp-tools/managed-agent-tools.ts
bash plugins/ruflo-agent/scripts/smoke.sh
bash plugins/ruflo-workflows/scripts/smoke.sh
node scripts/audit-cli-mcp-tools.mjs
node scripts/audit-tool-descriptions.mjs
git diff --check
Actual local tooling: Node 24.22.1, Vitest 3.2.6, TypeScript 5.9.3, and the Ruflo security scanner pinned to 3.25.6. The GitHub workflow installs isolated Vitest 3.2.6, TypeScript 5.9.3, Node 22 types and cli-core 3.7.5 with lifecycle scripts disabled. No provider key is needed by CI.
The scanner was run read-only against copies of
managed-agent-tools.ts, managed-workflows.ts, and the seven changed hook
files in the agent and workflow mods, using:
npx -y @claude-flow/[email protected] security scan \
--target /tmp/ruflo-managed-audit --depth deep --type code
Measured optimization
Run on Node >=22.18, with the baseline commit available locally:
node scripts/bench-managed-workflows.mjs
The benchmark executes the exact original summarizer extracted from the frozen baseline source, compares incremental reduction with repeated full-history summarization, verifies equal assistant text and alternates timing order after warmup. Both timings include output assembly. Nine measured repetitions, 5,001 events, 51 observations. The source hashes are in the JSON receipt.
| Metric | Baseline | Candidate |
|---|---|---|
| Median local processing | 10.287 ms | 6.097 ms |
| Events visited | 132,501 | 5,001 |
This sample is 1.69 times faster with 96.2% fewer event visits. It does not measure provider latency, tokens, billing, model quality or bug-finding recall. Inclusive watermarks can replay multiple events sharing a timestamp; those events are deduplicated rather than counted twice.
Limits and next adoption gate
The live Anthropic API was not invoked. All provider tests use deterministic HTTP fixtures. The full monorepo suite was not run; the focused boundary and adjacent workflows were tested. Hosted GitHub CI is a separate observation.
Before broad adoption, the maintainer should validate one live session with a small cents budget, minimal network access and a non-sensitive fixture, inspect every worker transcript, independently reproduce results, and explicitly delete the session and owned resources. Then compare single agent, current Ruflo swarm and managed workflows on the same frozen corpus at equal dollar budgets over three runs. Suggested adoption gate: at least 10% more reproducible findings per dollar, no higher false-positive rate, and verified interruption/deletion.
No claim of improved model quality or production readiness follows from the synthetic optimization. Keep the PR in draft until live validation is recorded.
Mod and plugin refinement
The follow-up adds 27 tests exercising real registered hook functions, policy settings, local capability discovery, no-network runtime introspection, workflow mod routing and compatibility with the existing console parser. Combined focused and adjacent suite: 108 passing tests. The original hosted Managed dynamic workflows contract passed before this refinement; new hosted checks must run on the updated commit. Plugin versions are raised to agent 0.4.0 and workflows 0.7.0 so installed copies receive the shipped behavior.
The native Claude plugin test runner was not invoked. Hook execution is covered by a deterministic test harness using the existing hook call pattern; the real host remains a separate compatibility check. Console parser integration is tested without changing the console plugin. The mod is a local admission constraint, not an organization-wide spend controller.