recursive-improvement

name	recursive-improvement
description	Deterministic RSI pass — findings store signals → loop state mutations, refinement hints, improvement ledger.

Recursive Self-Improvement (deterministic)

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.

Patterns (all deterministic)

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

Step 1 — Analyze (no pipeline)

.venv/bin/python -m night_shift_security.cli.main improve

Step 2 — Inspect outputs

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>

Step 3 — Act on refinement hints

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.

Step 4 — Lab notebook

Log: which RSI actions fired, cooldown deltas, refinement queue changes. Same vs different vs prior tick.

Gotchas

• RSI never bypasses validate_hypothesis() or evidence grading.
• config_hints are advisory — Day Shift or next loop tick decides config swap.
• Coordinator and bounty loop share refinement_seeds_from_store() — same grade 1–2 band.