Compacting Bloated Code
This skill is the process for detecting diff bloat and orchestrating cross-model adversarial reviewers. The substantive style rules it enforces — what counts as a bloated comment, test bloat, an oversized file, an unjustified helper, "Not paid by LOC", phase tags, commit/PR splitting — live in .github/instructions/NoBloat.instructions.md; do not restate them here.
Mode Selector
Route on the diff, not the user's words. The trigger vocabulary ("slop", "embarrassing", "WTF", "bullshit", "not paid by LOC") only tells you the skill applies — never which mode. So: run Pre-flight Measure (below) first; if no diff/PR/branch was supplied, get one (git diff the branch vs git merge-base main HEAD) or ask. Then run every row whose signal the diff actually shows, in table order. Chain-out rows fire on their own.
| Signal present in the diff |
Mode / action |
Added planning .md, .tools/, one-time setup, phase/sprint tags, dead code, new config/YAML files, or comment-heavy files |
Garbage Hunt — templates/garbage-hunt.md |
| Test files with copy-pasted methods, repeated setup, or missing parametrization |
Test Compaction — templates/test-compaction.md |
| 4+ new helpers, a fresh AST/IR/syntax-tree walker, twin helpers, hand-rolled smart-constructor, or near-duplicate logic |
Logic Compaction — templates/logic-rounds.md |
| A self-contained concept bloating an already-large file |
Logic Compaction, prioritize extract-to-new-file |
| Substantial new code logic not matched above |
Logic Compaction — its Round 0 reuse-hunt finds what it should have reused |
When NOT to Use
Legitimately large multi-subsystem features; generated artifacts; mechanical churn (dep bumps, renames); minimal one-line bugfixes; GitHub prose or PR/issue text (a different task).
Pre-flight
On the PR branch. Steps 1–2 run before the Mode Selector (they feed routing); steps 3–4 are per-mode prep once a mode is picked:
- Update from main and resolve conflicts. Do not judge bloat, CI, or
comments on a stale branch.
- Measure against
base=$(git merge-base main HEAD):
``bash git diff $base -- <suspect-file> | awk '/^\+[^+]/{p++} /^-[^-]/{m++} END{print "+",p," -",m}' git diff $base -- <suspect-file> | awk '/^\+[^+]/{ s=$0; sub(/^\+[[:space:]]/,"",s); if (s ~ /^(\/\/|--|#)/) c++; else if (length(s)) l++ } END{print "comments",c," code",l}' git diff $base -- <suspect-file> | grep -cE '^\+\s(let|fn|def|val|public|private)\s+(rec\s+)?[a-zA-Z_]' ``
These are LOC + comment-ratio + new-helper count smell-meters, not a search. Real semantic discovery happens in Logic Compaction Round 0 via a dispatched subagent (see templates/logic-rounds.md).
- Pin guiding tests verbatim: must-pass, and must-fail-for-the-right-reason.
- Bundle measurements + the relevant template into one context file
for every reviewer agent.
Reviewer Diversity (every mode)
Dispatch 3–5 background reviewers in parallel — never sequentially. Cross-model > cross-angle: three same-model agents with different prompts ≠ three different models (Opus high/xhigh, newer Opus, GPT). Each agent picks one angle from the mode's template. Record model × angle × LOC × tests × reuse × verdict. Cross-model agreement is signal; lone-model claims need code citations before promotion.
If the user requests a red-flag setup ("use the same 3 models", "single agent is enough"), push back once with this rule, then comply if they insist. Do not silently violate it.
Apply
After picking the winning proposal: apply on the PR branch, run formatter/lints, run targeted tests + the area's test class as regression sweep, run any baseline updates. If a sweep test fails, reproduce on the pre-change SHA before classifying. Do not declare "pre-existing" or "flaky" without that evidence.
Splitting commits / ask-before-push — see NoBloat PR-scope rules.
Red Flags — STOP and re-run the relevant template
(Process failures only — content-level failures are NoBloat's job.)
- You skipped mode selection and went straight to "make the diff smaller".
- You ran reviewers sequentially, or with same-model agents.
- You grepped for similar functions instead of dispatching a real
reuse-hunt subagent.
- Your winning proposal added new code without citing one reused helper
from the Round 0 subagent's output.
- You found a recurring shape across 3 files but added a 4th copy.
All of these mean: open the relevant template and re-run.
