Re0 State-Note & Shower-Test Guard
"You don't need long context; you need dense state notes and recursive resets." — NVIDIA Discovery (Belcak et al., 2026)
Why This Skill Is Needed & Target State
Long context windows (>50K tokens) degrade agent reasoning via attention diffusion, hallucination debt, and conversational pollution. Belcak et al. proved on SWE-bench Verified that agents with 16K-token short state notes match or exceed 200K-token runs while serving 7x faster.
The target state is a structured context-reset protocol:
- 1,500 Token Budget: Dense, zero-fluff summary of objective, proven facts, and negative corpus.
- Shower Test (Cold-Read Audit): Verifies that a subagent with zero prior conversational history can deterministically execute the next step.
- No Ungrounded Pronouns: Banning ambiguous relative references ("this file", "the bug", "it") without explicit file paths or identifiers.
- Preserved Negative Corpus: Keeping records of discarded hypotheses and dead ends so future cycles never repeat them.
Invariant Rules
- Strict Token Ceiling: State-notes must never exceed 1,500 tokens.
- Falsifiable Invariants: Every entry in
PROVEN_INVARIANTSmust cite an executable verification command and digest. - Discrete Single Action:
NEXT_SINGLE_ACTIONmust specify exactly one concrete file and verification command, not a multi-step speculative plan.
CLI Usage
# Compile a new Re0 State-Note
bunx anti-slop-engineering re0-handoff compile \
--goal "Upgrade database layer to fail closed" \
--action "Remove sqlite fallback" \
--target "src/db.ts" \
--verify-cmd "bun test tests/db.test.ts"
# Run the Cold-Read Shower Test
bunx anti-slop-engineering re0-handoff shower-test --file .anti-slop/state-note.json