task-router-lite

name	task-router-lite
description	Thin Phase PLAN router for codex-max. Select at most one primary execution skill and one optional secondary skill from the installed local skill surface.
risk	safe
source	codex-max
date_added	2026-04-19

Task Router Lite

Overview

task-router-lite is the canonical Phase PLAN router for this workspace. Its job is to make the execution surface lighter, more predictable, and more faithful to the installed local skill set.

It does not fan out into many agents or many skills. It does not rely on missing orchestration scripts. It does not invent process overhead when a narrow local skill is enough.

Core contract

Before implementation:

1. Identify the active repo.
2. Identify related repos for read-only context.
3. State allowed write scope and git scope.
4. Classify the task as easy, medium, or hard.
5. Select at most:
   • 0-1 primary execution skill
   • 0-1 optional secondary skill
6. Prefer installed local skills first.
7. If an ideal skill is not installed yet, fall back to the closest installed native behavior.

Hard limits

• Do not activate many skills in parallel.
• Do not call missing orchestration scripts.
• Do not assume external skill libraries are installed just because they are planned.
• Do not invent time estimates unless the user explicitly asks for an estimate.
• Do not replace execution skills; only route into them.

Primary routing table

Situation	Primary skill	Optional secondary	Fallback if planned skill is not installed
Bug, failing test, unexpected behavior	systematic-debugging	verification-before-completion	none
Multi-step or high-context task	planning-with-files	verification-before-completion	none
Medium planning task without heavy execution	concise-planning	none	none
Code changes that must be validated	lint-and-validate	verification-before-completion	none
Completion / pass / fixed claim	verification-before-completion	none	none
Underspecified feature or vague request	brainstorming	planning-with-files	use concise-planning or planning-with-files
Library / framework / API lookup	documentation-lookup	none	use local docs, repo evidence, or Context7
Before coding, dependency choice, or utility choice	search-first	none	use targeted codebase search and repo evidence
Unfamiliar repo onboarding	codebase-onboarding	planning-with-files	use targeted repository discovery
Significant architecture decision or trade-off	architecture-decision-records	planning-with-files	use ADR-style structured notes
Git branching / commit / PR / merge / release discipline	git-workflow	verification-before-completion	use direct repo git conventions
Critical user journey or browser-flow verification	e2e-testing	verification-before-completion	use existing test tools directly
Auth / secrets / sensitive endpoints / risky data handling	security-review	verification-before-completion	use direct security checklist

Difficulty guidance

• easy: direct execution, usually no subagent
• medium: direct execution unless the task separates cleanly
• hard: prefer at least one subagent when the work is separable and the current tool/runtime constraints allow it

Output contract

Before implementation, produce the readiness table required by workspace policy:

Context Element	Details
Active repo	[path]
Related repos	[paths]
Context readiness	Ready / Incomplete
Execution intent	[brief description]
Difficulty level	easy, medium, or hard
Subagent decision	Spawning Subagent / Direct Execution
Phase BUILD Skills	[selected skills]

Execution handoff rules

• Choose the narrowest skill that matches the task.
• Keep the BUILD phase small and explicit.
• If no installed skill matches perfectly, say which fallback behavior is being used.
• If the user asks for pure analysis only, stop at routing and analysis.

Workspace fit

This router exists because the workspace is curated. The runtime surface must stay aligned with .agents/skills/, while the source library stays aligned with skills/README.md, skills/CATALOG.md, and skills/manifest.yaml.