/Simplify

name: Simplify
metadata:
  short-description: "Makes a file, skill, tool, subsystem, or codebase smaller and clearer by mapping every dependent and cutting only what parity checks prove safe."
version: 1.0.1
description: "Research-first simplification of any file, skill, tool, subsystem or codebase against its ideal state: reads its ISA or derives its purpose, maps every dependent (callers, imports, references, hooks, settings, docs, tests), classifies each element with evidence as load-bearing, duplicate, dead, cruft or over-abstracted, then proposes and applies cuts that keep every behaviour a dependent relies on, proven by before-and-after parity checks. USE WHEN simplify, simplify this, simplify this system, remove cruft, remove duplication, dedupe, dead code, clean this up, this has gotten bloated, too complicated, reduce complexity, consolidate, strip this down without breaking it, what can we delete. NOT FOR always-on context files (Trim), cleanup of just the current diff (/simplify), harmony across many systems (Unify), auditing prompt rules for over-prompting (BitterLessonEngineering), bug hunting (/code-review)."
evals:
  manual:
    reason: "the claim that matters is that no behaviour a dependent relies on was lost, which only the target's own tests and probes can show, and those differ for every target"
    falsifier: "after an Apply run, a baseline check from the survey passes fewer cases than before, or a removed name still has a caller that a fresh reference search finds"

Simplify

Simplify makes a target smaller and clearer without breaking anything that depends on it. A target can be a file, a skill, a tool, a subsystem or a whole repo. The work comes in three intents, each building on the record from the one before: Survey maps the target read-only, Propose ranks the cuts and holds them for a yes, and Apply lands the approved cuts with proven parity. The central rule: a cut never drops a behaviour that a dependent or ISA criterion relies on, and anything the evidence cannot explain stays. Records sit next to the target's ISA when there is one.

Ideal State

• The target ends smaller and clearer, and every behaviour a dependent or ISA criterion relies on still works, shown by the same baseline checks passing before and after.
• No element is cut on looks: each cut rests on a survey finding with its searches cited, and anything the evidence cannot explain stays.
• Nothing changes until the principal approves the specific rows, and a delete of irreplaceable content gets its own yes.
• Each stage leaves its record under `LIFEOS/MEMORY/WORK/{slug}/simplify/`, so the next stage builds on evidence rather than on memory.

Intent Routing

| Intent | Trigger | File | |--------|---------|------| | **Survey** | "simplify X", "what can we delete", "map this before we cut", "is this dead" | `Intents/Survey.md` | | **Propose** | "what would you cut", "propose the simplifications", follows Survey in the same run | `Intents/Propose.md` | | **Apply** | "apply them", "y" or approved row numbers after a proposal | `Intents/Apply.md` |

A bare "simplify X" runs Survey and then Propose, and stops at the proposal. A Survey-only request, which is how a read-only analysis calls this skill, stops at the survey.

Data & Config

None. This skill reads no config and needs no account or key. Each stage writes its record under `LIFEOS/MEMORY/WORK/{slug}/simplify/` next to the target's ISA, if it has one; nothing else is created.

Gotchas

• **Zero code-search hits does not mean dead.** LifeOS calls a great deal of code by string: hook registrations in `settings.json`, launchd plists, workflow YAML, skill routing tables, spawned commands and dynamic imports. `ReferenceCheck.ts --shipped-paths` exists because plain path scans missed spawn and dynamic-import specifiers.
• **The file can change mid-run.** Memory loops and peer sessions write to the same trees. A file edited after its survey is re-read before it is touched (Trim, where a concurrent deploy-command fix was nearly reverted).

Examples

**Example 1: Simplify a tool**

User: "simplify LIFEOS/TOOLS/ReferenceCheck.ts"
→ Survey intent, then Propose
→ Purpose from its header and DocCheck lineage, every caller, findings per function, test baseline
→ A ranked table of cuts, each with its parity check, held for y

**Example 2: Land the approved cuts**

User: "y, but skip 4"
→ Apply intent
→ Rows 1–3 and 5 land as separate commits, tests show the same pass count before and after, and the measured reduction sits beside the estimate

**Example 3: An always-on context file**

User: "simplify CLAUDE.md"
→ Routed to Trim, which owns always-on context files and their directive-coverage gate