← Catalog·Tooling·2026

triage

triage is a Python service that sits between an agent (or a CLI, or a web UI) and `fusion-web`. The job: given a prompt + context + constraints, return a `RoutingDecision` that says *which* model or *which* panel of models should answer it, and at what cost and latency. The agent doesn't have to guess whether a query is worth a 3-model MoA call. A fast-path heuristics layer (token count, code-block presence, language, stakes markers, perspective markers, tool-call hints) handles ~80% of prompts in under 5ms with no network call. The remaining ~20% go to a single DeepSeek classifier call that returns a structured JSON decision. If the classifier fails twice, triage falls back to a conservative rule-based default — the classifier is never the failure point. Every decision is logged to `~/.triage/decisions.jsonl` (prompt preview, features, decision, classifier path, latency, cost) for offline analysis. A 35-case golden regression set ships with the repo and runs in CI. Triage is built around the same philosophy as `fusion-web`: stdlib-only Python, no pip install, single-file HTTP server with SSE streaming, model registry in a single Python dict. It direct-imports `fusion-web`'s `lib.fusion_core` for execution and never touches OpenRouter except when the routing decision itself requires it.

live
  • ai
  • routing
  • moa
  • deliberaton
  • classifier
  • deepseek
  • fusion-web
  • tooling
Open in new tab
Language
Python 3 stdlib only — no pip dependencies
HTTP layer
stdlib `http.server` + Server-Sent Events, single-file `server.py` on port 8760
Classifier
DeepSeek V4 Pro via OpenAI-format endpoint, temperature=0, structured JSON output
Execution backend
Direct import of `fusion-web/lib/fusion_core.py` for single calls and MoA deliberation
Decision model
Hybrid: fast-path heuristics (<5ms) + slow-path LLM classifier + rule-based fallback
Logging
Append-only JSONL at `~/.triage/decisions.jsonl`, 100MB rotation, 10 files retained
Testing
35-case golden regression set (`examples.jsonl`), unit + e2e tests, live smoke tests gated on cost cap
Models in registry
12 models across 5 vendors (DeepSeek, Moonshot, Z.AI, Qwen, MiniMax, Anthropic) in 3 roles (panel / synth / single)
CLI
`triage decide` for ad-hoc routing, `triage replay` for diff-testing against logged decisions
Bridgeworks integration
Catalog entry on bridgeworks.dev, sibling project to fusion-web on the LAN service

Live demohttp://192.168.1.184:8760/
Loading triage — live demo