Releases ·

bitnerve v2.1.0

@decoded-cipher/bitnerve · Private repository

Overview

v2.1 adds a new operator console to the dashboard and makes the engine more reliable under real conditions: late candles, hung sockets, parallel tool calls and races between the watcher and the model. The model now runs on Claude Opus 5.5 at xhigh effort, and every cycle records its turns, duration, cost and token usage.

Trading is still simulated. Market data is live from CoinSwitch, but no order reaches the exchange.

Live dashboard → https://bitnerve.arjunkrishna.dev

What's new

Operator console

  • New pages: Deck (live book, equity, P&L attribution, watcher feed), Record (book and round trips), Replay (per-cycle outcome bars) and Model (per-cycle run table with session tiles)
  • Forensic modals: a round-trip autopsy with fee ledger, exit attribution and hold-time outcome, and a full cycle replay
  • Live polling with a header refresh control, a ladder strip showing each position's exit levels, and a risk budget and margin allocation view
  • A new /api/v2 read-only data layer with a unioned exit timeline
  • Times are stored in UTC and shown in IST on the dashboard

Cycle telemetry

  • Each cycle runs claude -p with JSON output and records turns, duration, cost and token usage on its invocation
  • The cycle passes its invocation id to the MCP server instead of adopting the latest open one
  • Model: claude-opus-5-5 at --effort xhigh

Engine and watcher hardening

  • Exit levels only fire on bars that start after the level was armed. The watcher can still act on late 1-minute bars, and it logs feed lag
  • Levels fill and are checked at a fresh ticker price
  • Model trims no longer re-arm the stop, and a rung that closes the whole position is handled cleanly
  • Each position's exit fills, rung changes, stop moves and exit events run in one transaction, so a failure rolls back and no longer blocks other positions
  • create_position checks the cut and arms the ladder inside its transaction. adjust_position applies add, stop and cut together, under the account lock, so it can't race the watcher
  • Stop and cut moves that would widen risk are rejected
  • Exchange requests time out after 8s (server-time calls after 3s), so a hung socket can't freeze the watcher or a tool call
  • Tool calls append to the invocation log inside Postgres, so parallel calls no longer overwrite each other
  • The brief flags late or stale 15-minute candles against the board and lists late exits

Symbols in the database

  • Traded symbols move from symbols.json to a symbols table. A migration seeds the basket: ZEC, ENA, PEPE, STRK, WLD, XRP
  • bun run symbols:sync refreshes exchange constraints. bun run symbols:sync ZECUSDT ENAUSDT ... replaces the basket and keeps any symbol with an open position

Build

  • Bun workspace: bun install from the repo root installs everything, and the dashboard now runs on Bun too
  • The schema is defined once in packages/db as @bitnerve/db and shared by the server and the dashboard. schema:sync is gone
  • Docker images build from the repo root: docker build -f server/Dockerfile .
  • The dashboard uses a 10-connection pool on Node and one connection per request on Workers

Upgrading from v2.0.0

  1. Run bun install from the repo root (this replaces the separate server and client installs).
  2. In server/, run bun run db:migrate. This applies three migrations:
    • 0008: cycle telemetry columns on agent_invocations
    • 0009: converts every timestamp column to timestamptz, reading existing values as UTC
    • 0010: creates the symbols table and seeds the basket
  3. If you kept a custom symbols.json, re-create that basket with bun run symbols:sync <SYMBOLS...>.
  4. Update TRADING_MODEL in server/.env to claude-opus-5-5.

Full changelog: v2.0.0...v2.1.0