/00

warming kernel · loading skills · binding memory

NERYA v0.1.0
Star on GitHub
Nerya, the self-evolving trading agent, drawn as a violet-haired operator who runs the console

operator console self-evolving runtime live, risk-gated

EVOLVE EVERY TRADE

Nerya ships its own agent kernel, LLM gateway, sub-agents, typed memory, triggers, a durable team, a trading kernel, and an evolution pipeline. Every external call goes through a SKILL.md. It trades live, with every order held under a Risk Gate and Approval Gate.

∞ / live evolution feed

self-evolving runtime · captured in motion

Watch the kernel rewrite itself.

One evolution cycle, in motion. The reflection engine drafts patches, the team votes, and versions v1.1 → v1.2 → v1.3 compile behind her while the Risk Gate holds the line.

01 / closed evolution loop

self-improving

Every observation rewrites the kernel.

After every session, the reflection engine reads the journal, drafts patch proposals (prompt deltas, new scripts, tuned configs, fresh skills), runs them through backtest and risk gates, then promotes the winners as the next agent version.

v1.1 → v1.3 +49% promoted per cycle
  1. observe journal every decision, fill, drift, miss.
  2. reflect postmortem with the analyst sub-agent.
  3. propose diff against the live kernel.
  4. validate backtest gate, risk gate, structured output check.
  5. promote versioned rollout, snapshot, rollback ready.
02 / risk-first execution

live, risk-gated

Mistakes go into memory, not into your wallet.

Go live the moment runtime.live_trading_enabled is true and the Approval Gate signs off. Every order ships through a Risk Gate, a typed mistake ledger, a virtual ledger, and reconciliation. A SKILL.md gates every external call.

vault-only secrets prompt firewall structured output schema script sandbox drift detection rollback snapshot
03 / durable agent team

persistent multi-agent crew

Four minds wired to one blackboard.

Four roles stay resident across every turn. The Lead routes the work, the Analyst reads the tape, the Risk critic can veto, and the Memory scribe records the lesson. They share a blackboard, a task board, and a mailbox, so each turn compounds into the next.

A-01 routing

LEAD

plans · routes · schedules

  • → open session 0x9f
  • → dispatch analyst
  • → await risk sign-off
A-02 reading

ANALYST

reads chart · drafts thesis

  • → regime: trend-up
  • → edge: +1.4R
  • → thesis drafted
A-03 guarding

RISK

vetoes if the math is off

  • → size ≤ 2% checked
  • → corr clamp ok
  • → veto held in reserve
A-04 writing

MEMORY

writes mistakes to the ledger

  • → postmortem saved
  • → regimes.md +1
  • → feeds next cycle
Nerya at a holographic strategy editor, pointing at an entry / filter / risk flow while cradling a glowing brain-orb beside a live BTC/USDT order book
strategy_author · trigger · candle · ledger
04 / strategy author skill

one sentence

"I have $500 in a live account. Make me money on BTC. Don't blow it up."

→ what a one-line prompt becomes

Skills compose. Strategies learn.

The strategy_author skill turns that sentence into a working strategy package in one chat turn: trigger, sub-agent prompt library, candle source, account binding, risk limits, session ledger. You approve it once, then it runs, journals, and rewrites itself from there.

read the skill catalog
live-ready
single-turn
operator-signed
25 builtin skills
zero yaml
SKILL.md only

05 · ritual / war room

The war room never sleeps.

Same four operators, now at the table. Shared blackboard, task board, mailbox, gates. Every turn leaves a record, and each record feeds the next session and the next version.

  • MARKET-READanalyst frames the regime
  • DEBATElead and risk argue the trade
  • EXECUTElive orders, gated by risk
  • JOURNALmemory seals the lesson

06 · ritual / typed memory

Five surfaces. Plain markdown. No vector blob.

Memory is five typed markdown surfaces the reflection engine can read, write, and diff. Audit them, grep them, or edit them by hand.

07 · ritual / self-rewriter

The kernel writes its own next version.

Reflection drafts a patch, the Approval Gate signs or rejects the diff, and the runtime applies it with a rollback snapshot. Each logged mistake feeds the next proposal.

+ self.reflect(session) - self.guess(market) + patch_proposal.apply()
Nerya mascot writing into a glowing brain-notebook with floating diff patches
08 / closed loop

architecture

Market turns the agent.
The agent turns itself.

agents operate the loop skills are the atoms evolution rewrites the agent
09 / read deeper

product docs

Take the kernel apart.

The README is the front door. The skills, the agent kernel, the trading kernel, and the SDK each have their own playbook. Open the one that matches what you want to build.

10 / launch console

The next edge is an evolution away.

Free and open-source. Bring your own models. The runtime ships its own kernel, gateway, team, trading kernel, and evolution pipeline. If a star helps the signal travel, drop one.

runtime kernel

own agent loop, planner, context builder, memory, reflection. No external agent framework.

safety gated

Live trades run under a Risk Gate and Approval Gate. Typed mistake ledger. Vault-only secrets.

evolution diff

reflection drafts patches. Operator signs the diff. Runtime applies with rollback snapshot.