arcasilesgroup Simplify
Discoverable wrapper around the ai-simplify agent: dispatches it via the Agent tool, validates after every change, applies the self-check protocol, and reports the simplifications made. On-demand only — no scheduling, no PR, no auto-commit (the complement of /ai-simplify-sweep, which IS scheduled, repo-wide, and opens a draft PR).
Quick start
/ai-simplify # current diff, default aggressiveness
/ai-simplify src/auth/login.py # scoped to a single file
/ai-simplify src/auth/ --conservative # scoped to directory, conservative
/ai-simplify --diff --aggressive # current diff, aggressive
Workflow
Principles applied: §10.1 KISS (the simplified version must actually be simpler — not just different); §10.7 Clean Code (names tell the story, functions do one thing, cyclomatic complexity ≤ 8).
- Step 0 — load contexts — read
.ai-engineering/manifest.ymlproviders.stacks; apply.ai-engineering/overrides/<stack>/conventions.mdso the stack linter is wired before any edit. - Detect target —
$ARGUMENTSresolves to explicit paths,--diff(staged changes), or empty (current diff, the default). - Dependency preflight — verify
.claude/agents/ai-simplify.mdexists. STOP and report the exact missing path if absent. - Dispatch — invoke the
ai-simplifyagent via the Agent tool with{paths, aggressiveness}. It applies guard clauses, extracts methods, flattens nesting, removes dead code, simplifies conditionals — validating after EVERY change (stack linter + fast tests). - Self-check (per simplification; any No -> revert) — (a) is it actually simpler, or just different? (b) would a newcomer find it easier? (c) if it adds an abstraction, does that abstraction earn its existence? (d) is complexity reduced, or just moved?
- Render report — grouped by file:
File | Change | Complexity Before | After | Lines Saved. - No PR, no commit — the in-flight lane; the operator owns the next commit.
Distinction from /ai-simplify-sweep
| Aspect | /ai-simplify |
/ai-simplify-sweep |
|---|---|---|
| Invocation | On-demand by operator | On-demand (weekly cadence recommended) |
| Scope | Operator-chosen paths or current diff | Repo-wide sweep |
| PR | None (in-flight work) | Always opens a draft PR |
| Auto-commit | No (operator owns next commit) | Yes (/ai-commit before PR) |
| Aggressiveness | Operator-chosen | Conservative-only |
| Telemetry | kind=simplify_ondemand_run |
kind=simplify_sweep_* |
Both share the same ai-simplify agent engine — they differ in cadence, scope, and post-conditions, not in simplification logic.
Output Contract
## Simplification Report
| File | Change | Complexity Before | After | Lines Saved |
|------|--------|-------------------|-------|-------------|
### Summary
- Files simplified: N
- Total complexity reduction: N points
- Lines removed: N
- All tests passing: YES / NO
If All tests passing: NO: report the failing test, do NOT auto-revert (the operator decides), and emit framework_event kind=simplify_test_regression so the failure is auditable.
Boundaries
- Behaviour preserved — same inputs, same outputs.
- Tests pass after every change — run any fast (<30s) suite after each file.
- Never modifies test files — tests are the immutable spec; simplify production code only.
- Never churns compliant code — no value simplifying code already below thresholds.
- One file at a time, validate before the next — never batch across files without intermediate validation.
- External API signatures are immutable — public signatures, exported types, CLI contracts are off-limits; refactor internals only.
- No new abstractions — reduce complexity in existing structure; do not add protocols/base classes/extension points.
- No PR, no auto-commit — the operator owns the resulting diff.
Examples
Example — scoped simplification of a single file
User: "simplify src/auth/login.py — the nested if/else is hard to follow"
/ai-simplify src/auth/login.py
Dispatches the agent scoped to the file. It inverts the outer if user is not None into an early-return guard, converts the inner if user.active to a guard, and drops the now-unreachable trailing else: return None. After each edit it runs ruff check + ruff format --check. Report: 1 file simplified, 3 complexity points reduced, 8 lines removed, all tests passing.
Integration
Called by: operators directly via /ai-simplify (single-file or diff-scoped) — not auto-invoked by any other skill. Calls: the ai-simplify agent (.claude/agents/ai-simplify.md) via the Agent tool with the operator-chosen scope; validates after each edit, rolls back on test failure. See also: .claude/skills/ai-simplify-sweep/SKILL.md (scheduled draft-PR wrapper); .ai-engineering/manifest.yml quality thresholds (cyclomatic ≤ 10, cognitive ≤ 15, nesting ≤ 3, method length ≤ 50). Anchors: §10.1 KISS, §10.7 Clean Code; D-134-07 (agents need a discoverable slash-skill).
Inline fallback — Agent-tool dispatch is the primary path. On a harness without a subagent/Agent-tool primitive, execute this skill by reading the specialist agent file (.claude/agents/ai-simplify.md) inline and running its steps in-context, sequentially, applying the same per-edit validation and self-check; inline-sequential execution is the floor, not an alternate behaviour.
$ARGUMENTS