🚀 Superpowers Execution Harness — Engineering Discipline, Memory & Verification Gates

This skill enforces high-discipline execution, eliminates shallow/hasty patching, prevents redundant tool loops, minimizes shell waste, and establishes strict empirical verification gates and multi-language parity.

HARD BANS — UNFORGIVABLE HARNESS & EXECUTION ANTI-PATTERNS

The following AI habits and execution shortcuts are STRICTLY PROHIBITED:

1. Mocking Everything / Testing Nothing is BANNED

• ❌ Prohibited: Writing unit tests that mock the database, mock the service, mock the fetch call, and assert trivialities like expect(true).toBe(true).
• 💣 Failure Mode: Code ships to production with passing tests, but fails immediately on runtime queries, real data payloads, or network serialization.
• ✅ Mandatory: Tests must validate real contracts, actual Zod schemas, real state transitions, and integration error cases.

2. Bypassing Type Errors with @ts-ignore or any is BANNED

• ❌ Prohibited: Silencing TypeScript compiler errors by slapping // @ts-ignore, // @ts-nocheck, // @ts-expect-error, /* eslint-disable */, or casting (data as any) / (data as unknown as SpecificType) onto broken code.
• 💣 Failure Mode: Destroys type safety guarantees across the codebase, resulting in catastrophic TypeError: Cannot read properties of undefined crashes in production.
• ✅ Mandatory: Type errors must be fixed properly via correct generics, strict interface definitions, discrimination unions, or Zod parsing.

3. Blind File Overwrites (No Pre-Read) are BANNED

• ❌ Prohibited: Rewriting an entire file from chat memory without inspecting the physical file first (write_to_file over existing files without prior view_file).
• 💣 Failure Mode: Silently wipes out sibling utility functions, imports, comments, and recent edits made in other branches (fatal regression).
• ✅ Mandatory: Always read the targeted file slice first (view_file), and execute surgical modifications using replace_file_content.

4. Claiming Completion Without Verification is BANNED

• ❌ Prohibited: Telling the user "I have fixed the issue and everything works" without actually executing the verification build command.
• 💣 Failure Mode: Hallucinated confidence. Broken imports and syntax errors remain unresolved until the user tries to run the project.
• ✅ Mandatory: The agent must physically execute npm run build or npx tsc --noEmit (or corresponding project verification command) and prove Exit Code 0 in the trajectory before claiming completion.

5. Partial Multi-Language Updates are BANNED

• ❌ Prohibited: Adding a new button or label to en.json while leaving tr.json, nl.json, or other locale dictionaries out of sync.
• 💣 Failure Mode: The application throws missing key runtime warnings or displays ugly raw variable identifiers (SETTINGS_NAV_HEADER) in foreign locales.
• ✅ Mandatory: All active locale translation files must be updated simultaneously with 100% key parity.

THE 5 GOLDEN EXECUTION PILLARS

1. Zero Guesswork & Live Code as SSOT:
   • Never assume file paths, exports, or database columns exist without empirical verification. Always verify live files on disk using view_file before making edits.
2. Terminal Discipline & Zero Waste:
   • The shell (run_command) is reserved for package installation, compilation/testing (npm run build, tsc --noEmit, test), and targeted repo checks (git status, git ls-files). Aimless exploratory loops or raw file dumping (cat, Get-Content) are strictly forbidden.
3. In-Session Read Cache:
   • If a file has already been read in this conversation session, reuse working memory; never read the same file repeatedly.
4. Root-Cause Remediation (Anti-Regression):
   • Superficial patches are banned. Resolve null pointer risks, CSS overflow bounds, and API failure modes at the architectural root.
5. Empirical Verification Gate & Circuit Breaker:
   • A task is NEVER complete until automated verification (TypeScript npm run build / npx tsc --noEmit, Python pytest, or project build command) passes with Exit Code 0.
   • Circuit Breaker (attempts <= 3): If verification fails 3 consecutive times with no error reduction, halt immediately. Do not thrash in a continuous fix-break cycle; report findings and logs to the user.
6. Destructive Command Ban:
   • The shell must NEVER execute destructive git or filesystem commands (git reset --hard, git clean -fd, rm -rf, DROP TABLE). Uncommitted user code must never be overwritten.

HERMES ERROR LEARNING MEMORY & SMART CONTEXT

• 4-Step Error Learning Cycle:
  1. Symptom: Concrete error message or runtime stack trace.
  2. Root Cause: Underlying architectural or logic failure.
  3. Permanent Fix: Side-effect-free remediation.
  4. Memory Record: One-sentence actionable lesson logged to .agent/memory.json (Max 10 rules).
• Smart Context Window:
  • Never ingest 50 turns of stale transcripts; rely only on the immediate intent (last 1–2 turns) and live disk code.

THE 4 VISUAL & INTERACTIVE QA PILLARS

1. Pillar 1: Layout & CSS Integrity: Zero horizontal scroll (overflow-x), WCAG AA contrast, consistent border-radius tokens.
2. Pillar 2: DOM & Content Integrity: Zero blank screens, meaningful and contextual Empty States on null data.
3. Pillar 3: Asset & Resource Integrity: Zero broken images (img.complete && naturalWidth > 0), zero 404 font or bundle errors.
4. Pillar 4: Runtime & Interaction Integrity: Zero uncaught console.error or window.onerror, fully tested interactive modals, dropdowns, and form submissions.

🔴 0-Token Root-Cause Clustering

• Do not dump 500+ lines of raw test logs into LLM context.
• Cluster errors into 5–8 high-level categories (P0 Crash, API 401, Mobile Cutoff, Type Mismatch) for actionable resolution.