Menu
Deterministic RSI pass — findings store signals → loop state mutations, refinement hints, improvement ledger.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Practitioner-elevated asset tracing patterns from SlowMist Crypto-Asset-Tracing-Handbook. Provides canonical property tables for obfuscation patterns (peel chains, mixers, bridge hops), address behavior clustering/risk profiling, cross-chain reconstruction workflows. Hard-first on behavioral + flow invariants. Integrates with codegraph-x-ray, deep-dive-handoff, and bounty forensics. Includes reference implementation for mixer deposit/withdrawal clustering. Trigger: alpha-miner on tracing handbook or similar.
Drop-in skill adapted from Glider query engine for static analysis, taint tracking, CFG/DFG traversal in bug bounty. Supports lazy reference loading, creativity layer for audited targets, integrates with bounty-loop. EVM priority with Solana compatibility.
Use for mining validator/runtime/cache/storage-key coherence bugs in blockchain clients, VMs, program loaders, bridge runtimes, and execution harnesses. Hard-first on stale cache, recycled identifier, partial flush, epoch/generation drift, mempool/order-dependent runtime state, and storage-key confusion. Trigger on runtime cache invariant, stale cache bug, VM storage confusion, client cache coherence, Aptos-style hijack, validator-local state, epoch-boundary bug, or when alpha-miner extracts runtime/client exploit engineering.
Cross-references a graphified target codebase (codegraph-x-ray output) against the local AuditVault corpus to surface structurally analogous historical vulnerabilities. Run after codegraph-x-ray has completed and produced invariants.md and property_candidates.md. Produces ranked pattern-match hits with per-finding graph anchor evidence. Pattern matches are advisory signals only — never evidence of exploitability and never a substitute for live reproduction or submission-gate validation.
Combines structured codegraph intelligence with rigorous invariant synthesis. Enforces high-quality usage of codegraph to identify the Primary Target Subsystem, then applies ordered structural invariant discovery with strict verification gates. Produces categorized invariants and high-quality property candidates ready for ultrafuzz-discovery. Trigger on codegraph-x-ray, x-ray with codegraph, invariant synthesis, primary subsystem invariants.
Optional accelerator and scaffolder for ultrafuzz-discovery. Brings parallel specialized invariant discovery agents, automated harness/handler scaffolding, coverage-driven refinement loops, property tagging, and reproducible test generation helpers from high-quality EVM fuzz patterns. Designed to speed up property fan-in and executable attempts phases while preserving NSS rigor, fresh-context pass@k, adjudication, and Crucible usage. Trigger on fuzz-scaffolder, accelerate invariant discovery, generate fuzz harness, scaffold handlers.
| name | recursive-improvement |
| description | Deterministic RSI pass — findings store signals → loop state mutations, refinement hints, improvement ledger. |
No LLM in this layer. Reads findings store + loop outcomes; writes bounded state updates and improvement_ledger.jsonl.
Runs automatically at end of each bounty loop tick. Use this skill for standalone analysis or post-cron triage.
| Action | Trigger | Effect |
|---|---|---|
repeat_fingerprint | Same top-findings hash as prior run on slug | Log + extend cooldown |
extend_cooldown | Repeat fingerprint | +12h per repeat (max 72h) |
queue_refinement | Grade 1–2 survivors, survival ≥ 0.4 | refinement_queue + refinement_hints.json |
plateau_template | Catalogue analogue grade ≥ 4 | template_plateaus[slug] |
boost_scan_priority | Refinement candidates in store | scan_boost_slugs |
config_fallback | Fork catalogue-only | config_hints[slug] → semantic recon or novel config |
failure_trace | Repeated PoC/probe failure fingerprint | failure_signatures.jsonl + semantic recon task |
.venv/bin/python -m night_shift_security.cli.main improve
cat data/security_results/loop/refinement_hints.json
tail -5 data/security_results/knowledge/improvement_ledger.jsonl
jq '.refinement_queue, .cooldown_overrides' data/security_results/loop/state.json
Summarize execution failures when trace files exist:
.venv/bin/python -m night_shift_security.cli.main traces summarize --slug <slug>
If refinement_hints.json has a top entry:
# Kamino campaign path (coordinator state):
.venv/bin/python hermes/scripts/nss-write-proposals.py
# Or scoped cross-target:
.venv/bin/python hermes/scripts/nss-write-scan-proposals.py --slug <slug>
Then re-run bounty loop with --proposals data/security_results/hermes_proposals/latest.json.
Log: which RSI actions fired, cooldown deltas, refinement queue changes. Same vs different vs prior tick.
validate_hypothesis() or evidence grading.config_hints are advisory — Day Shift or next loop tick decides config swap.refinement_seeds_from_store() — same grade 1–2 band.