Provides pre_tool_call hooks for schema validation, loop detection, circuit breaker, context budget, and routing awareness in small-model agentic workflows via Hermes plugin.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Defensive harness for running 1B–12B parameter models in production agentic workflows.
Five-layer architecture: routing → validation → constraint → circuit break → context management. Delivered as a Hermes plugin (deterministic enforcement) + MCP server (analysis tools).
Small models (1B–12B) have dramatically closed the quality gap with frontier models, but the reliability gap persists:
Failure Mode | Impact | Source |
Doom Loops — repetitive death spirals under greedy sampling | Qwen3.5-4B: 22.9% loop rate | Antidoom (FTPO) |
Context Rot — degradation with input length, not position | Effective window = ~1/3 of stated window | Chroma/NVIDIA |
Format Drift — tool call fragility on long chains | 87% accuracy vs GPT-4o's 92% | Qwen3-32B eval |
Related MCP server: RelayPlane
┌─────────────────────────────────────────┐
│ ENTRY POINT / TASK DISPATCH │
└────────────────┬────────────────────────┘
│
┌────────────────▼────────────────────────┐
LAYER 1 │ TASK CLASSIFIER │
ROUTING │ Classify complexity, route to model │
│ T1 (<4B) → T2 (4-8B) → T3 (9-12B) → │
│ T4 (cloud) cascade with confidence │
└────────────────┬────────────────────────┘
│
┌────────────────▼────────────────────────┐
LAYER 2 │ PRE-CALL VALIDATION │
GUARDRAIL │ Schema check, loop detection, budget │
│ Blocks bad calls BEFORE execution │
└────────────────┬────────────────────────┘
│
┌────────────────▼────────────────────────┐
LAYER 3 │ CONSTRAINED DECODING ENGINE │
OUTPUT │ XGrammar token masking (40μs) │
ENFORCEMENT │ Guarantee valid JSON/tool call output │
│ (NOT YET IMPLEMENTED) │
└────────────────┬────────────────────────┘
│
┌────────────────▼────────────────────────┐
LAYER 4 │ CIRCUIT BREAKER │
LOOP │ 3-state: closed → open → half-open │
DETECTION │ Detect loops, break circuits, escalate │
└────────────────┬────────────────────────┘
│
┌────────────────▼────────────────────────┐
LAYER 5 │ CONTEXT BUDGET │
CONTEXT │ Sliding window compaction │
│ 1/3 effective window rule │
└─────────────────────────────────────────┘Component | What It Does | Status |
Hermes Plugin |
| ✅ Phase 1 |
MCP Server | 5 tools: context status, compaction, task classification, routing, reset | ✅ Phase 2+3 |
Task Classifier | Rule-based complexity scoring and model tier assignment | ✅ Phase 3 |
Context Router | Tier cascade (T1→T2→T3→T4) with failure-based escalation | ✅ Phase 3 |
Output Enforcement | Constrained decoding (XGrammar/Outlines) for guaranteed valid output | ⏳ Planned |
Output Verifier | Post-generation validation of tool call correctness | 📋 Planned |
# Copy plugin to Hermes plugins directory
cp -r hermes-plugin ~/.hermes/profiles/dev/plugins/small-model-harness
# Enable it
hermes plugins enable small-model-harnessAdd to ~/.hermes/config.yaml or ~/.hermes/profiles/dev/config.yaml:
mcp_servers:
small-model-harness:
command: python3
args: ["/path/to/mcp-server/server.py"]
enabled: trueTool | Description |
| Query context budget utilization for a session |
| Compact session context (sliding window summarization) |
| Classify task complexity and suggest model tier |
| Route task to model tier with cascade logic |
| Reset all harness state for a session |
Tier | Model Size | Example Models | Use Case |
T1 | <4B | SmolLM3, phi-4-mini | Simple extraction, classification |
T2 | 4-8B | Qwen3-8B, Llama 3.2-8B | Single tool calls, basic routing |
T3 | 9-12B | Ornith-1.0-9B, Qwen3-30B-A3B | Multi-step, reasoning, planning |
T4 | Cloud | DeepSeek V4, GPT-4o | Complex chains, security-critical |
# Setup
uv sync --extra dev
# Run tests
uv run pytest tests/ -v
# Test MCP server
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2026-07-28","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}' | python3 mcp-server/server.pyWorkflow | Status |
CodeQL |
|
MIT
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
AI routing, memory, guardrails, and governance. Routes across Claude, GPT, Gemini.
Intelligent context infrastructure for AI teams: knowledge graph, sessions, tasks, documents.
Shared task layer for AI agents: tasks, handoffs and decisions that persist across tools.
Context and control for any AI. Agent identity, scoped access, and proof across tools.
