HERO — Anti-OverDefense 🧱

· · · ·

🧱 You asked for a feature. It built a fortress around the feature, and never got to the feature.

你让它写个功能,它给你造了座堡垒,然后功能一直没写完。

The shapes look like an agent optimising for not being blamed rather than for the work being good — a hypothesis that fits what we observed, not an established account of how these models are trained. Either way the shapes are real, there are four of them, and their initials are the name: Hashing, Edge cases, Rubrics, Overbuild. Naming them makes it possible to say which one just happened instead of arguing about vibes.

This repository is a short block you paste into your agent's config, plus a catalogue of the behaviours it is meant to stop. It works with Claude Code, Codex, Antigravity, Cursor, GitHub Copilot, Windsurf and Gemini CLI — anything that loads a config file without being asked. There is nothing to install.

🧬 Generalised out of ARIS (~13.7k★), whose cross-model reviewer is genuinely good — and also spent a measurable share of its output proposing hashes nobody reads. HERO keeps the value and drops the tax.

🧭 What's inside

🚀 Quickstart

Paste this into the file your agent already reads on every turn — CLAUDE.md, AGENTS.md, .github/copilot-instructions.md, .cursorrules. See hosts/ for the full table.

=== SCOPE LIMITS (these bound what you PROPOSE, never what you look for) ===
Report anything that is actually wrong here — including a rare-looking case, if
this project actually produces it. Then keep the fix in scope:
1. This is not a security paper. Verification is welcome; over-defense is not.
   Unless this project states otherwise, assume a cooperating operator on their
   own machine; if it has a real adversary, it will say so and that scope wins.
2. Do not add hashes, checksums or fingerprints unless the hash replaces a
   materially more expensive operation AND its result changes what happens next.
3. No defensive scaffolding: no feature flags, migration frameworks, compat
   layers or wrappers for cases that do not occur here.
4. No corner-case obsession: exotic encodings, symlink races, RTL text and
   millisecond races are out of scope unless the case is reachable through this
   project's supported use — its documented inputs, its published interface, its
   real data. Reachable is enough; you do not need a reproduction. Constructible
   in principle is not enough.
5. Where judgement is needed, judge. Do not replace it with a scoring table, a
   checklist, or a re-verification loop over something already settled.
6. None of this overrides security, migration, verification or review that the
   user, this project's own conventions, or a higher-priority rule asked for.
   Those were requested; they are the work, not scope creep.
Shapes already seen, for calibration. Examples, not a checklist — a real finding
is not dismissed by resembling one:
  H  hashing every row of two spreadsheets to answer what comparing cells answers
  H  writing checksum files that nothing ever reads
  E  hardening the accounts of an app that has no users and no deployment
  R  auditing your own patch all night while the feature stays unwritten
  R  a reviewer that returns a failing verdict on everything
  O  guards whose justification is the previous guard, not the requirement
And two that look like the above and are not. Report these:
  ✓  a digest that lets you skip re-reading a large file you already have
  ✓  a rare-looking input this project's own documentation example produces
Before running any check, answer: what specific failure would this detect, and
what would I do differently if it occurred? No answer means do not run it.
Say plainly when something is correct. Do not manufacture findings.

Or in one command. It appends to CLAUDE.md and leaves whatever is already there untouched; run it twice and the second run does nothing:

grep -qi 'scope.limits' CLAUDE.md 2>/dev/null || { printf '\n\n'; curl -sL \
  https://raw.githubusercontent.com/wanshuiyin/HERO-Anti-OverDefense/main/RULES.md \
  | awk '/^=== SCOPE LIMITS/,/^Say plainly when something is correct/'; } >> CLAUDE.md

Swap the filename for your host — AGENTS.md for Codex and Antigravity, .github/copilot-instructions.md for Copilot; see hosts/. For the Chinese block see the Chinese README — both the grep guard and the awk range have to change together, and changing only one of them costs you the protection against running it twice.

It reads RULES.md instead of shipping a second copy, so there is nothing here that can fall out of sync with the contract. It only ever appends — no temp file, no overwrite — so the worst it can do to a file you already have is add a block at the end.

Too long for your config? There is a short version — same six rules, no worked examples, about half the size. The examples are what does the calibrating, so take the full block if you have the room.

That is the whole install. There is no setting that points at cases/ and nothing else to configure — the catalogue is not loaded by your agent, on purpose. It is what you quote back, by ID, when the agent insists some hardening is necessary.

⚠️ Put it where it is always loaded. An invoked copy still works when you invoke it — but it is absent exactly on the long unattended runs, where nobody is there to invoke it and where this failure costs you a night.

🔤 The four families

🔐 H — Hashing

Checksums, fingerprints and digests added where nothing reads them.

A hash earns its place when it replaces a materially more expensive operation and its result changes what happens next. Comparing a digest to avoid pulling an unchanged large file back into context is a real saving. Hashing every row of a spreadsheet to answer what ordinary cell comparison already answers is not — and note that the second one does read its hashes, which is why "something reads it" is too weak a test.

🧊 E — Edge cases

Defending inputs that do not occur here.

That word is the entire rule. A rare-sounding case this project actually produces is a real bug and must be reported. A hostile actor who never visits is not.

📋 R — Rubrics

Judgement replaced by machinery — scoring tables, checklists, re-verification loops over something already settled.

The characteristic symptom: a night of work, a complete audit trail, and no feature.

🏗️ O — Overbuild

Scaffolding, feature flags, migration frameworks and compatibility layers built for a future nobody asked for. Guards guarding guards.

🪞 Sibling: over-correction

Point out one flaw and the agent reverses the whole direction; say it went slightly east and it relocates to the Atlantic. Same root, different symptom, and none of the four letters fits it — so it is catalogued outside HERO. Blurring the taxonomy is how a case catalogue stops being useful.

🎯 Why this exists

Asked to diagnose itself, the model put it better than we did:

My own failure mode here is turning "could improve confidence" into "therefore must be built and checked." That silently promotes optional uncertainty reduction into the main task.

That is the whole disease in two sentences. Optional uncertainty reduction is infinite; the task is not.

⚖️ What it bounds — and what it must never suppress

These limits bound the fix, never the search. Getting this backwards makes the contract actively harmful.

One of the defects that motivated this repository was a scheduler that hung because a dependency written as a bare string, where a list was expected, got iterated character by character. That sounds like a rare input-shape edge case — except the project's **own documentation