From d-nd-core
Seed-driven autonomous decision cycle. Loads project state from multiple sources, analyzes tensions and momentum, decides ONE action, executes it, verifies the outcome, and updates the seed. Use for autonomous research, maintenance, or any recurring decision-making process.
How this skill is triggered — by the user, by Claude, or both
Slash command
/d-nd-core:autonomous-cycleThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
A project seed (`seed.json`) tracks tensions, blocked potential, variance between cycles, and direction. The cycle reads the seed, integrates multiple sources, decides what to do, does it, and updates the seed.
A project seed (seed.json) tracks tensions, blocked potential, variance between cycles, and direction. The cycle reads the seed, integrates multiple sources, decides what to do, does it, and updates the seed.
SEED (seed.json)
|
v
LOAD SOURCES (git log, data files, audit results, pipeline state)
|
v
ANALYZE (research state, artifact maturity, recent activity)
|
v
DECIDE (one action, ranked alternatives, confidence score)
|
v
EXECUTE (respecting autonomy levels: Auto / Notify / Approve / Escalate)
|
v
VERIFY (did the seed advance? did the action produce results?)
|
v
UPDATE SEED (new tensions, resolved tensions, variance, direction)
The seed is NOT a report (what happened). It IS a direction (where the potential points).
{
"timestamp": "ISO-8601",
"piano": 1,
"tensioni": [
{
"tipo": "contraddizione|confine_inesplorato|simmetria_sospetta|scoperta|tensione_aperta",
"id": "UNIQUE_ID",
"claim": "What the tension is about",
"intensita": 0.8,
"nota": "Why this matters"
}
],
"potenziale_bloccato": [
{
"tipo": "bloccato",
"id": "ID",
"claim": "What can't proceed",
"dettaglio": "What's missing"
}
],
"varianza": ["What changed since last cycle"],
"direzione": "One sentence: where the potential is highest",
"verifica": {"pass": 10, "fail": 1, "skip": 2, "total": 13}
}
The decisore ranks actions by type (highest priority first):
Every action type has an autonomy level:
| Level | Actions | Behavior |
|---|---|---|
| Auto | Lab experiments, seed update, data analysis | Execute and notify after |
| Notify | New scripts, minor fixes, state updates | Execute, notify immediately |
| Approve | Modify artifacts (papers, pages, code), new formalizations | Propose, wait for confirmation |
| Escalate | Modify axioms, contradictions in core, irreversible decisions | Ask the operator |
Use seed_cycle.py (in scripts/) as the generic engine. Configure via:
PROJECT_DIR — root of the projectDATA_DIR — where seed.json and results liveSOURCES — dict of source loaders (functions that return state dicts)ASSERTIONS — list of testable claims with test functionsEXECUTORS — dict mapping decision types to executor functions# Example: nightly at 03:00
0 3 * * * cd /path/to/project && python seed_cycle.py --execute --update 2>&1 >> logs/cycle.log
After execution, the cycle checks:
If the piano didn't advance and no new tensions emerged, the system is in stasis. Next cycle will prioritize a pivot.
$ARGUMENTS
npx claudepluginhub grazianoguiducci/d-nd-seed --plugin d-nd-coreRuns Karpathy-inspired autonomous iteration loops on any task: modify, verify, keep/discard, repeat. Subcommands for planning, debugging, fixing, security audits, shipping.
Autonomous technical lead that drives projects from commander's intent to completion using an OODA loop. Invokes other skills for implementation, refactoring, review, and bug-hunting.