AgentShekel/agentic-workflow

Русский · English

agentic-workflow

Multi-agent framework for Claude Code: 58 agents, 46 methodology skills, 16 + 3 Python orchestration scripts, 2 Workflow orchestration engines + 2 LangGraph human-gate engines, tier-aware acceptance (S/M/L), filesystem-isolated adversary review, cross-family second opinion via Codex MCP, human as supreme judge at critical transitions.

v0.3 (2026-06-05): engagement orchestration unified under the engagement-workflow Workflow — the main loop conducts a single pre-gate cascade (plan → deliver in isolated git-worktree waves → validate → handoff → gate) and stops at the handoff seam; the LangGraph human-gate (consilium → directive → manager) remains the acceptance path after the seam. Domain leads are planning-only; specialist coordination is structural — waves in the lead's plan. See CHANGELOG.md.

v0.2.4 (2026-05-28): Windows compatibility — three latent issues surfaced on Max-subscription claude CLI: claude.CMD npm-wrapper truncates multiline argv at the first newline (CMD line-parsing), subprocess.run(text=True) decodes UTF-8 Russian as cp1251 on Russian-locale Windows, and consilium_synth_completed ledger emit was passing raw natural verdict to a schema expecting ACCEPT/REJECT/DIRECTED. All three fixed across 4 scripts (find_claude_cmd() resolves .CMD → claude.exe; 10 subprocess sites got encoding="utf-8", errors="replace"; inline VERDICT_MAP mirror in _make_finalize_node). All --invoker mock tests passed pre-fix; latent risk lived in real subscription mode untested on Windows until now.

v0.2.3 (2026-05-28): engagement_lg.py end-to-end across all 11 nodes in three execution modes. NEW --mock mode runs the real graph paths but with canned-artefact subprocess wrappers — full end-to-end smoke testing without claude CLI required. Send fan-out to specialists, validator_lg.py + `adversary