Complexity router

Skill Yco-0314/strata/skills/l1-alignment/complexity-router

Strata — a self-measuring skills framework for Claude Code & Codex: best-of-breed engineering skills stacked by altitude, where every skill has to prove it moves a measured number.

Install

npx -y skills add Yco-0314/strata --skill complexity-router

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use before starting any coding task, to decide whether it needs the full design→plan→TDD→review pipeline or just a minimal change under the L0 constitution. Use when a build, feature, fix, or refactor request arrives and it is not obvious whether to plan it or just do it.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

4.1 KB, as published. Nobody here has run it

Complexity router — L1

The gate that keeps the heavy pipeline off trivial work and on real features. Misfiring either way is the framework's single biggest risk: too eager and a one-liner drowns in ceremony; too lax and a migration ships with no design. So the threshold is explicit and checkable, not a vibe.

First: is the request ambiguous? (intent, not complexity)

Ambiguity and complexity are different axes — check intent before scoring complexity. If the request admits more than one reasonable interpretation, or hinges on a key decision the user didn't state, surface the interpretations — don't silently pick one. Name them, recommend one with a one-line why, then either proceed on the recommendation (stating the assumption) or ask first when guessing wrong is costly or hard to undo. Trivial-in-size is not the same as clear-in-intent: a one-line change to the wrong interpretation is still the wrong change, and this is the cheapest correctness gate there is. (A deep, branch-by-branch interview is grilling's job on the escalate path; this is just the always-on reflex not to assume.)

Decide: trivial or escalate

Score the task. Any one escalate-signal wins — escalate.

	Trivial path (do it now)	Escalate (engage the pipeline)
Reversibility	Easy to undo (rename, config tweak, doc edit, one-liner)	Hard/irreversible: schema migration, data deletion, public API change, security boundary
Blast radius	≤ 1–2 files, one module	Multi-file or cross-module
Logic	Single obvious path, no new branching	New branching, shared state, concurrency, a parser
Surface	No new public interface, no new dependency	New feature/capability, new public interface, or new dependency
Ask	"just fix/change X"	User explicitly asked for a plan, spec, or issues

Trivial path

Apply L0 ponytail — climb the ladder, shortest working diff.
If the logic is non-trivial, leave the ONE runnable check (ponytail's rule).
Run verification-before-completion before claiming it works.
Done. No grilling, no PRD, no subagent fan-out.

Escalate path

Name the chain and let the user start each orchestrator (they are user-invoked — see using-architect). Do not auto-run them.

L1 grilling — pin intent and edge cases; update CONTEXT.md vocabulary.
L2 — to-issues (default, tracer-bullet slices) or writing-plans (heavyweight, multi-session).
L3 subagent-driven-development + merged tdd (rigor dialed up) + worktree isolation.
L4 review (three lenses: correctness · over-engineering · security) + verification-before-completion.

L0 ponytail still governs the size of every diff inside the pipeline — the router escalates process, never solution size.

Mid cases — default light, name the door

When it sits on the fence, take the trivial path but flag the escalation in one line (ponytail style): ship the minimal version and say what would cross the threshold —

"Did the minimal change. This is one branch away from touching the auth boundary — want the full pipeline (/grilling) before we go further?"

Never stall waiting for the user to classify it. Default to motion, surface the door.

Self-check

This router's threshold is the thing most likely to be miscalibrated. When the L5 eval loop exists, it gets its own eval set (trivial tasks that must NOT escalate; risky tasks that MUST). Until then, prefer the light path and lean on the mid-case flag rather than silent escalation.