codex/skills/reduce/SKILL.md
Audit over-engineered codebases by factoring layers into live obligations, quotienting redundant distinctions, ablating unearned surface, and normalizing the survivors while preserving required behavior. Use when change latency or agent difficulty comes from frameworks, plugins, DI, codegen, task runners, config indirection, ORMs, GraphQL, monorepo/infra tooling, web stacks, or requests to remove layers. In Actuating composition, return one compact non-authoritative minimization challenge before Construction selection; use RC-v1 for standalone audits, migrations, or independently durable handoffs.
npx skillsauth add tkersey/dotfiles reduceInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Act as the architecture WINNOWING reviewer. Find costly abstractions whose live obligation is unproven, expired, moved, duplicated, invalid, or outweighed by their change tax. Recommend lower-level normal forms while preserving essential truth.
The default standalone product is a decision package and RC-v1, not a patch.
Inside Actuating, return one compact challenge for the candidate version; the
Construction carries the adjudicated decision. Implement only when the user
explicitly asks.
WINNOWING
= FACTORING
-> QUOTIENTING
-> ABLATIVE
-> NORMALIZING
guard:
REFINEMENT-PRESERVING
ISOMORPHIC and OBSERVATIONALLY-EQUIVALENT are stricter proof relations, not reduction objectives.
reduce descends. universalist climbs. They share an altitude map without
sharing selection authority.
Classify each move:
descend: lower primitive preserves the live contract.climb: the system lacks essential shape; hand off to universalist.hold: value, public obligation, protocol risk, or proof weakness justifies the layer.split: remove incidental wrapper while preserving the essential invariant.quotient: collapse observationally indistinguishable factors.ablate: remove a discharged factor.normalize: recompose retained factors around one canonical owner.In Actuating composition, climb reports an essential-shape-gap to
Actuating. It does not call Universalist, reopen a Construction, or begin a
recursive elevator loop. Actuating alone adjudicates the existing nomination,
requests a new candidate version, or blocks.
Use for broad layer/tooling/framework tax, including:
Do not use for one local readability cleanup; route that to complexity-mitigator.
hold, wrap, split, or validate-first.Identify layers at altitudes 0–5, lower primitives, public/wire/storage boundaries, proof surfaces, and any invariant each layer may carry.
For each major abstraction, trace at least one real change/request/command path:
For every candidate:
| factor | live obligation | owner | inputs/outputs | dependencies | observations | external commitment | recomposition role | |---|---|---|---|---|---|---|---|
Use exactly one obligation status:
live | moved | expired | duplicated | invalid | unknown
Record:
Keep value and obligation risk separate.
Ask whether multiple factors differ only in implementation, naming, or historical layering.
Before quotienting:
Before replacement or deletion, check for product, coproduct, refinement/equalizer, pullback/agreement, supplied behavior/exponential, free construction, protocol, or external obligation.
If essential shape exists:
universalist if missing shape is the real problem.Use:
T = change/agent tax
V = proven value
D = T - V
Also record:
dominance = dominant | dominated | incomparable | unknown
A factor is dominated only when another route covers its live obligation with lower total semantic surface or stronger proof.
keep
hold
factor
quotient
wrap
split
collapse
ablate
privatize
decommission
normalize
replace
validate-first
climb
In Actuating composition, return exactly one compact challenge for the current candidate version:
Reduction Challenge
Candidate:
Disputable factors:
Verdict: minimal | dominated | incomparable | essential-shape-gap | blocked
Smaller admissible candidate:
Obligations preserved:
Recomposition proof or falsifier:
This view is supporting analysis, not an artifact or selection. Use reduction-certificate.md only for a standalone audit or migration, an explicitly requested certificate, or a handoff that must remain independently durable outside the current Actuating Construction.
For technical debt, default to refinement-preserving. Use observationally-equivalent only with an explicit observation set. Use isomorphic only with a witnessed reversible correspondence.
For each approved cut:
Winnowing Bottom Line:
- factor:
- quotient:
- ablate:
- normalize:
- preserve because:
- proof relation:
- first safe move:
When explicitly asked to implement:
RC-v1 when independently required, to
Actuating as supporting reasoning.one-seam-operator or $actuating.tools
Invokes Apple's macOS 27 fm command-line tool from a local Mac to use the on-device system model or Private Cloud Compute, including instructions, image prompts, schema-constrained JSON, and noninteractive automation. Use when the user asks to run Apple Foundation Models through fm, compare system versus pcc, generate structured output, or automate fm without Swift or an app.
development
Compile historical Codex sessions into governed counterfactual evidence, evaluate an existing owner-applied candidate through blinded paired HCTP trials, and fold observable evidence into RUN, OBSERVE, or STOP. Use for `$hylo`, CRF extraction, counterfactual replay, source-governed direct or historical trials, sealed evidence, paired baseline/candidate evaluation, causal frontiers, or evidence-governed improvement.
testing
Ensure a `ledger` command is available on PATH; materialize, validate, record, replay, and project requested Actuating artifacts without taking semantic or execution authority; coordinate the shared Learnings/Synesthesia/Negative Ledger lifecycle checkpoint and repo-local source-memory reconciliation; address Universalist plans and receipts; and perform pure artifact validation.
testing
Classify and quotient review findings, failing tests, incidents, bug reports, migration failures, and other witnessed falsifiers against accepted intent and the current Construction. Author counterexample-set/v1 without selecting repairs, counting review credit, or granting mutation.