xeno-harness 4.1.0

Multi-agent autonomy harness with multi-provider support, per-role permission tiers, and Agent Harness Engineering observability.

agent-harness

An autonomy-aware Claude harness with cross-session memory, skills, a streaming dashboard, and an MCP server. Built across a 7-lab series; see LABS/ and DEEP-DIVES/.

Layout

autonomy-lab/        # Canonical harness (sync CLI loop)
  autonomy_harness.py    # Budget, CacheMetrics, @tool registry, memory,
                         # compaction, sub-agents, skills, ~1600 lines
  memory_backend.py      # FilesystemBackend + SqliteBackend
  skill_loader.py        # Two-tier cascade + path-safe Read
  consolidation.py       # End-of-session Haiku consolidation pass
  mcp_server.py          # MCP stdio server exposing the registry
  skills/                # Skill cascade (3 starter skills)
  tests/                 # pytest suite (offline + API-gated)

ai-harness/          # Streaming dashboard (FastAPI + React)
  backend/agent.py       # 190-line streaming wrapper around autonomy_harness
  backend/main.py        # FastAPI app: auth, rate-limit, CORS, /api/*
  frontend/              # Vite + React + Tailwind v4 + shadcn

.github/workflows/   # CI: pytest + tsc + backend-import on every PR

Use it

# CLI single-shot (full machinery, no streaming)
cd autonomy-lab
uv run python autonomy_harness.py "Your goal here"

# Dashboard with streaming UI
cd ai-harness && ./dev.sh   # backend :8000, frontend :5173

# MCP server (Claude Code, Cursor, ...)
# In your client's MCP config:
#   "autonomy-harness": {
#     "command": "uv",
#     "args": ["run", "--directory",
#              "/abs/path/to/autonomy-lab",
#              "python", "mcp_server.py"]
#   }

Deploy the dashboard

cd deploy
cp .env.example .env   # set ANTHROPIC_API_KEY and DASHBOARD_TOKEN
docker compose up --build
# http://localhost:8080

See deploy/README.md for the three modes: local laptop (HTTP), Mac mini + Tailscale (HTTP over private mesh), or public domain (HTTPS via Let's Encrypt). Memory + replay state persist in deploy/data/.

Dev mode (no DASHBOARD_TOKEN) auto-disables auth.

Web tools

web_search and web_fetch are wired as Anthropic server-side tools (zero local impl, real-time index, billed per call). On by default; disable with WEB_TOOLS_ENABLED=0 before running the CLI or the dashboard backend.

Tests

cd autonomy-lab && uv run pytest -v

CI runs the offline subset (38 tests, ~2s) on every PR. API-using tests (test_skill_triggers, test_context_rot, test_claude_picks_right_tool) self-skip without ANTHROPIC_API_KEY.