profile
Ships the discipline
crux-config-wizard bundles a code-minimalism profile: one command puts the ladder in every agent's CLAUDE.md, versioned, with a drift test pinning it to the runtime overlay.
The code-minimalism plane
Less code, measured.
The cheapest code to maintain is the code that was never written. The Crux daemon turns that from a slogan into substrate: a discipline agents actually follow, services that do the discipline's legwork, and receipts for what it saves.
The idea
A ladder climbed before writing
Does this need to exist at all? Does it already exist in this codebase? Does the standard library, the platform, or an installed dependency cover it? Can it be one line? Only then: the minimum code that works. Safety is carved out explicitly: trust-boundary validation, data-loss handling, security and accessibility are never traded for a shorter diff.
The pattern was popularised by the MIT-licensed Ponytail skill, credit where due. A prompt can say all of it. A memory daemon can do it. That difference is this page.
What the daemon does
Four services, one discipline
engram
Serves it pre-execution
The same ladder ships as a built-in engram: a ~200-token overlay resolved by intent before the agent acts, gated by model capability class. Operators override it per-workspace with a fact; override wins by name and version.
reuse_check
Answers the hardest rung
"Does this already exist?" A prompt can only tell the agent to grep. The daemon's reuse_check tool answers from the substrate: ranked candidates from the retrieval index and the Features lens, with an honest nothing-found when the index is cold.
debt ledger
Remembers the shortcuts
Deliberate simplifications carry a crux-min: marker naming their ceiling and upgrade trigger. The code-health harvest turns markers into receipted debt facts, idempotently; resolved markers retire on the next run.
crux-config-wizard add code-minimalism # ladder into CLAUDE.md
# engram + reuse lookup (flag-gated, off by default):
CORECRUXD_FEATURE_ENGRAM_MCP=1 CORECRUXD_FEATURE_REUSE_CHECK=1Measured, not asserted
What it saves
Our profile, exactly as the wizard injects it, against same-day paired baselines on twelve real-repo feature prompts. Every measured cell produced a non-empty diff, but the pinned harness did not execute the generated tasks or their tests. These figures measure code volume and resource use, not correctness.
| model | code written | tokens | modelled cost | wall-clock |
|---|---|---|---|---|
| Fable 5 | −61.0% | −24.1% | −29.0% | −32.4% |
| Opus 4.8 | −59.9% | −38.2% | −35.4% | −34.6% |
Stronger models need it more
Left unconstrained, stronger models over-build more (baseline code volume grew roughly 43% from Sonnet to Opus on identical tasks) while disciplined output converges. The ladder closes a bigger gap on better models.
A code-volume win, not a token win
The ladder spends read tokens to avoid written lines; on some tasks input tokens rise while code falls. If tokens are your only metric, look twice; the durable saving is the code that never existed.
Full methodology, the upstream Ponytail numbers next to ours (including where their plugin beats our profile), every limitation we know about, and reproduce-it-yourself commands: scorecrux.com/code-minimalism.
The deeper economy
Context is the scarce resource
Token prices fall; context windows don't scale the same way, and an agent's effective attention is scarcer still. Code volume is a context liability with compound interest.
Code is read far more often than it is written
Every line shipped today is context spent in every future session that greps it, loads it, and reasons over it. The benchmark produced ~60% smaller diffs; whether those patches are correct still needs an execution-based gate.
The cheapest context is the context you never load
reuse_check substitutes substrate memory for window: the daemon's index answers "does this exist?" in one call, instead of the agent filling its window with speculative file reads.
Discipline delivered small
The engram carries the whole ladder in ~200 tokens, resolved once, pre-execution, instead of the agent re-deriving house style from thousands of tokens of examples each session.
Debt without re-discovery
The ledger keeps shortcuts as queryable facts with receipts, so no future session burns context re-scanning the tree to find what was deferred.
Less code isn't just cheaper to run; it's a smaller mortgage on every context window that will ever visit the repository.
One profile away.
The plane ships with the daemon. Enable the profile, flip the flags when you want the runtime services, and the receipts follow.
An AI agent (Claude) authored this plane's code and these measurements under human review; commits, receipts and run manifests carry the attribution.