Skip to content

Releases: Somchandra17/w-bonkers

v1.1.0 — portfolio brakes

Choose a tag to compare

@Somchandra17 Somchandra17 released this 03 Jul 18:34

Book-level safety, distilled from a head-to-head with a sibling system (idea credit: @kiraa06's comparison).

  • Portfolio brakes (pinned.portfolio_brakes, default on): open-risk cap (2% of corpus), weekly loss budget (2%), drawdown circuit breaker (12% from peak). Breaches defer new buys or raise a de-risk proposal — exits are never gated, and nothing is ever sold automatically.
  • Same-run exit proposals: a held position breaking its exit rules now raises a p1 PROPOSED EXIT Todoist task immediately, in every mode — you learn the same evening, you still decide.
  • Lapsed-order re-surfacing: resting day orders that died at close are re-raised in every report.
  • Truth fixes: mutual-fund and IPO legs documented as first-class; the engine holds cash by design; half-size regime-gate probation available as a customize recipe.

Upgrade an existing install: git pull, then tell your agent "re-run prompt.md in upgrade mode".

Education only — the agent never places orders; you place every trade.

w-bonkers v1.0.0 — take the W

Choose a tag to compare

@Somchandra17 Somchandra17 released this 03 Jul 10:47

Your AI agent runs your stock plan on rails. You just take the W.

First public release. Everything a friend needs to go from zero to a running plan:

  • One-prompt install — clone, open Claude Code or Codex, say Read prompt.md and execute it. The installer checks and installs missing tooling (Groww MCP incl. its browser OAuth, Todoist MCP, the analysis skills), converts your PDFs, interviews you (every question with an example answer), proposes an opening book you approve line by line, and ends with a plain-language summary + customization check.
  • Deterministic enginestate.json is the single source of truth; render_plan.py is the only view-writer; the default decision every run is NO CHANGE. Feedback via Todoist ticks/comments, applied exactly once.
  • Everything archived locally — every fetch lands in runs/, with an offline RUNBOOK and a machine-readable MANIFEST so the plan outlives any one tool.
  • Scheduling — dated Todoist task, adaptive cron (scripts/run_refresh.sh), or both. Todoist syncs two-way with Google Calendar; tasks.ics covers everything else.
  • Swap anything — broker MCP, task tool, scheduler: the seams are documented prompts away (docs/CUSTOMIZE.md).

Field status: running on the author's own money since day one — first week executed exactly as designed. New project; read the code, start small.

Education only. Not investment advice. The agent never places orders — you place every trade.