complexity-triage

name	complexity-triage
description	T-shirt sizing for tasks before dispatch. Runs heuristics and optionally LLM classification to determine task size (S/M/L/XL), recommend local vs cloud routing, and set tool budgets.
version	1.0.0
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["triage","complexity","routing","dispatch","planning","t-shirt-sizing"],"related_skills":["local-model-router","writing-plans","architect","skill-discovery"]}}

Complexity Triage

Run before planning to classify task complexity and set dispatch parameters.

When to use

Before every coding engine dispatch. The triage result determines:

• Whether to plan (M/L/XL) or dispatch directly (S)
• Which model to use (local gemma4 for S, cloud gemini-3.5-flash for M+)
• Tool budget (max skills, max MCP tools, max turns)
• Which skills to inject via --append-system-prompt

Dispatch

terminal(
    command="python3 ~/.hermes-coder/scripts/dynamic_curator.py --task '<task summary>' --repo '<project-dir>'",
    workdir="~/.hermes-coder",
    timeout=30
)

Reading the output

The script outputs JSON:

• size: S, M, L, or XL
• confidence: high, medium, or low
• method: heuristic, llm, or fallback
• routing.engine: "local" or "cloud"
• routing.model: model identifier
• tool_budget: max_skills, max_mcp_tools, max_turns
• skill_matches: top matched skills to inject

Size definitions

Size	Characteristics	Action
S	Single file, trivial verb, CSS-only, typo	Skip planning, raw dispatch, local model OK
M	Multi-file logic, new component, helper utility	Plan, selective skill injection, cloud model
L	New feature, major refactor, multi-file integration	Full plan, full skill injection, cloud model
XL	Architecture change, system redesign	Full plan, architect review, cloud model

Every size dispatches to the coding engine. Size never licenses the coordinator to edit repo files itself — an S task is a small dispatch, not an inline edit. (2026-07-10: inline implementation by the coordinator burned two full turn budgets and shipped below the quality bar.)

Fallback

The triage LLM pass runs through the active coding harness (resolved from --engine/coding.default_engine). If the harness is unavailable, heuristic-only triage still works (exit code 3). If heuristics are inconclusive and no LLM is available, the script defaults to M sizing with cloud routing.