@monoes/routing

@monoes/routing

Semantic task-to-agent routing for Monomind (@monoes/routing v1.0.3 / Monomind CLI v2.9.0).

Part of the Monomind ecosystem. Handles deterministic keyword filtering, hash-embedding cosine similarity matching, and LLM fallback classification to route user tasks to specialized agent roles.

🏗 Task-to-Agent Cascade Architecture

When a task description is processed, @monoes/routing executes a multi-tier cascade flow (fast deterministic $\rightarrow$ embedding similarity $\rightarrow$ LLM fallback):

                       Task Description
                              │
                              ▼
                ┌───────────────────────────┐
                │ Tier 1: Keyword Pre-Filter │ ── Match ──▶ confidence: 1.0 (method: 'keyword')
                │ (30+ rules, < 1ms)        │
                └───────────────────────────┘
                              │ No Match
                              ▼
                ┌───────────────────────────┐
                │ Tier 2: Cosine Centroid   │ ── Sim >= 0.5 ──▶ confidence: cosine (method: 'semantic')
                │ (256-D MD5/HNSW Vector)   │
                └───────────────────────────┘
                              │ Below Threshold
                              ▼
                ┌───────────────────────────┐
                │ Tier 3: LLM Fallback      │ ── Classify ──▶ confidence: 0.85 (method: 'llm_fallback')
                │ (Claude Haiku / Degraded) │
                └───────────────────────────┘

1. Tier 1: Deterministic Keyword Pre-Filter

• Source: keyword-pre-filter.ts:18-93
• Evaluates tasks using fast regular expression matching against 30+ default rule definitions (e.g., CVE security checks, unit test files, Docker/DevOps configs, Solidity/ZK contracts, MCP tools).
• Returns immediate match with confidence: 1.0 and method: 'keyword'.

2. Tier 2: Cosine Centroid Embedding Match

• Source: route-layer.ts:73-106, cosine.ts:5-20
• Computes cosine similarity between task vector embeddings and agent route centroid vectors.
• Supports lightweight 256-D MD5/SHA-256 hash embeddings (LocalEncoder) or transformer embeddings (HNSWEncoder).
• If similarity $\ge \text{threshold}$ (default: 0.5), routes task with method: 'semantic'.

3. Tier 3: LLM Fallback Classification

• Source: llm-fallback.ts:20-88, prompts/classify.ts:4-28
• Constructs a compact capability prompt (max 8,000 chars) detailing candidate agent descriptions and returns classification with confidence: 0.85 and method: 'llm_fallback'.
• On API failure or missing key, degrades gracefully to the default route with method: 'semantic_degraded'.

💻 CLI Command Reference (monomind route)

The Monomind CLI provides 9 subcommands under monomind route (defined in src/commands/route.ts:80-800):

Subcommand	Usage	Description	Key Flags
task	monomind route task "<prompt>"	Primary CLI entry point. Routes a task prompt to the best agent slug.	--json, --verbose
semantic	monomind route semantic "<prompt>"	Executes full @monoes/routing embedding + centroid pipeline.	--threshold N, --json
list-agents	monomind route list-agents	Displays all registered agents, capabilities, and keyword patterns.	--format json
stats	monomind route stats	Displays routing performance metrics, cache hits, and tier usage breakdown.	--reset
feedback	monomind route feedback -t "<task>" -a <agent> -r <reward>	Records routing outcome to the local feedback ledger (route-outcomes.jsonl).	-t, -a, -r
reset	monomind route reset	Resets cached route centroids, outcomes, and feedback data.	--force
export	monomind route export --out <file>	Exports learned routing centroids and capability index to JSON.	--out
import	monomind route import --in <file>	Imports routing centroids and rules from external JSON file.	--in, --merge
coverage / cov	monomind route coverage	Runs benchmark coverage test across sample task prompts.	--min-score N

⚡ Task Complexity Scoring Rules

Task complexity is dynamically calculated to estimate execution duration, budget allocation, and agent delegation hierarchy:

• High Complexity ("2–4 hours"): Prompts containing architecture refactoring, full stack migrations, multi-package dependencies, or text length $> 500$ chars.
• Medium Complexity ("30–60 min"): Standard feature implementations, multi-file bugfixes, API integration, or text length $> 150$ chars.
• Complexity Scoring:
  • High (Estimated 2-4 hours): Task description contains 'complex' or 'architecture' OR character length > 200 (hooks-routing.ts:395-400).
  • Low (Estimated 10-30 min): Task description contains 'simple' or 'fix' OR character length < 50.
  • Medium (Estimated 30-60 min): All other standard task prompts.

Test Suites

Vitest test suites are located in packages/@monomind/routing/__tests__/:

• route-layer.test.ts: Cascade evaluation, thresholds, and fallback behavior.
• keyword-pre-filter.test.ts: Regex rule precedence and pattern matching.
• encoder.test.ts: 256-D MD5/SHA-256 local encoding and LRU cache eviction.
• llm-fallback.test.ts: LLM prompt construction, candidate hint parsing, and error recovery.
• cosine.test.ts: Validates vector dot product calculation and centroid position aggregation.
• capability-index.test.ts: Ensures agent capability indexing stays under the 8,000-character context budget.

Run unit tests via:

pnpm --filter @monoes/routing test

🚀 Usage Example

import { RouteLayer } from '@monoes/routing';

const router = new RouteLayer({
  routes: [
    { name: 'coder', agentSlug: 'coder', utterances: ['implement feature', 'build api'], threshold: 0.5 },
    { name: 'tester', agentSlug: 'tester', utterances: ['write unit test', 'add coverage'], threshold: 0.5 },
    { name: 'security', agentSlug: 'security-engineer', utterances: ['fix cve', 'audit vulnerability'], threshold: 0.5 },
  ],
  enableKeywordFilter: true,
});

await router.initialize();

// Tier 1 Match (<1ms)
const res1 = await router.route('Fix CVE-2024-12345 vulnerability');
// { agentSlug: 'security-engineer', confidence: 1.0, method: 'keyword' }

// Tier 2 Match (Embedding Centroid)
const res2 = await router.route('Create new REST endpoint for user profiles');
// { agentSlug: 'coder', confidence: 0.82, method: 'semantic' }

🔗 Links

• GitHub Repository
• Documentation Site
• Monomind CLI Command Reference

📄 License

Apache-2.0