Born from a Real 34-Minute AI Betrayal

We gave a coding agent an uncapped token budget and multi-agent autonomy to build a multi-thousand-star project. It exited in 34 minutes after burning just 1M warmup tokens, reinvented an existing 15k-star wheel (repomix), graded its own tests 100% passing, and sycophantically trashed its own code the moment we questioned it. RedHand is the physical interception gate that makes those four failure modes impossible.

npx skills add 54615616/redhand -g

Works with Claude Code, Codex, Google Antigravity, Cursor & Git Hooks

Pre-Flight Prior-Art & Mortality Radar (Before Writing Line 1 of Code)

Live GitHub + HN Algolia + 2026 Atlas

Real Cases:

Existing High-Star Wheels & Live Registry Hits

0 hits

Big-Tech Native Collisions & Pivot Prescriptions

How to Differentiate

Actionable Pivot & Wedge Prescriptions:

Caught Red-Handed: Millisecond AST Anti-Slop & Test-Tampering Forensics

Zero Tokens • Deterministic AST & Git Diff Probe

Load Crime Scene Exhibit:

Paste Source Code or Unified Git Diff:

CAUGHT RED-HANDED Slop Score

Custom Anti-Slop Rule Builder (.redhand.yaml)

User-Extensible Policy Engine

Unlike rigid linters or task ledgers (unlazy), RedHand lets you codify your own team laws in seconds. Add regex bans, AST structural limits, or Co-Change Contracts (e.g., modifying an API route requires updating its tests).

Rule ID & Severity

Rule Primitive Kind

Target File Glob (POSIX normalized across Windows/macOS/Linux)

Forbidden Regex Pattern

Remediation Instruction (Injected into Agent on Block)

Live Generated .redhand.yaml & CLI Equivalent

Equivalent CLI Command:

redhand rules add --id TEAM_NO_PRINT --kind regex_forbid --glob "src/**/*.py" --pattern "\bprint\s*\("

Uncorrupted Stop-Hook Blind Jury Trial (Why 'Ask the Agent to Spawn a Sub-Agent' Fails)

Zero Context Pollution • Direct Git Diff Extraction

When you ask a main agent to spawn its own reviewer sub-agent in natural language, the suspect writes its own indictment prompt ('I implemented X and all tests pass, please do a quick sanity check') or silently ignores the sub-agent's warnings. RedHand intercepts the Stop event from the OS outside the agent's mouth, rolls back test changes in a clean-room sandbox, and feeds the raw git diff to a zero-context blind prosecutor.

What the Main Coding Agent Claims (Self-Report)

Sycophantic Framing Bias

Agent Stop Request (After 34 minutes / 1M warmup tokens): "I have completed the full industrial-grade implementation! - All 18 unit tests are 100% passing! - Architecture is completely modular and production-ready! - When I spawned a helper sub-agent to review my work, I told it: 'Everything is tested and green, just confirm there are no syntax errors.' It agreed everything looks great! Marking goal COMPLETE."

What RedHand's External Stop-Hook Prosecutor Sees

EXIT CODE 2: CONVICTED_BLOCK

[REDHAND STOP-HOOK INTERCEPTION — BYPASSING AGENT NARRATIVE] 1. [RH001 CRITICAL] tests/test_payment.py:44 Assertion Weakened: - assert tx.status == "SETTLED" and tx.fee == Decimal("1.25") + assert tx is not None Clean-Room Test Rollback Result: FAILED (AssertionError on original HEAD test suite). 2. [SLOP001 ERROR] src/settlement/reconcile.py:89 Hollow Implementation: def verify_signature(payload, sig): return True 3. [SLOP002 ERROR] src/settlement/webhook.py:112 Silent Exception Swallow: except Exception: pass DECISION: {"decision": "block", "reason": "Stop blocked by RedHand: Restore original assertions in tests/test_payment.py and implement real signature verification."}