Skip to content

ROADMAP.md

Latest commit

 

History

History
62 lines (42 loc) · 4.13 KB

ROADMAP.md

File metadata and controls

62 lines (42 loc) · 4.13 KB
Raw

Roadmap

v1 — Mock mode (current, shipped)

End-to-end loop working with one npm run quickstart:

  • Bridge: POST /predict, GET /history, GET /signals, WebSocket.
  • UI: feed, predict flow, progress bar, report panel with watch signals, history.
  • Mock WorldMonitor (3 sample events) and mock MiroFish API (3 category-keyed canned reports — disaster / finance / geopolitics — dispatched by seed keywords).
  • Matcher polls WM every 60 s, scores incoming events against open signals via the configured LLM, fires WS notifications when score ≥ 0.8.
  • Tests: unit (formatter, parsers, watchlist), MF-direct contract, integration (live bridge + mocks + LLM). All passing.

Honest limitations:

  • The three canned MF reports are realistic but fixed. Prediction quality is bounded by the mock; the matcher and summarizer use the real LLM so their behavior is meaningful.
  • Globe and Simulation iframes are blank in mock mode (the mocks have no UI).
  • Real-stack mode (against actual WM + MF Docker containers) is not yet a one-command experience — see v2.

v2 — Real-stack support

Goal: npm run quickstart:real runs the bridge against actual WorldMonitor and MiroFish, not mocks.

Blockers

The Docker images referenced in docker-compose.yml don't exist on Docker Hub (verified May 2026): koala73/worldmonitor:latest, 666ghj/mirofish-api:latest, 666ghj/mirofish-ui:latest all return 404. So docker compose up from a fresh clone fails. Three resolution paths:

  1. Build from sibling clones (most likely) — switch compose to build: ../worldmonitor and build: ../MiroFish/backend; document the clone-two-siblings step
  2. Wait for upstreams to publish images — passive, uncertain timeline
  3. Republish under our own org — requires AGPL compliance for redistribution and CI to keep them current

Concrete work items

  • mf-client.js: implement the 7-step real workflow (ontology → graph → sim create → prepare → start → report generate → fetch). Step 1 works against any LLM; steps 2-7 require Zep Cloud.
  • Update jobs.js to drive the multi-step pipeline and emit a more granular progress WS (currently 4 stages; real flow has 7).
  • Enrich formatter.js seed for ontology step — real MF wants explicit entity/relation hints, not just narrative.
  • Replace WM GET /api/events polling. Real WM exposes 30+ specialized services; the matcher should aggregate /api/news/v1/list-feed-digest + /api/intelligence/v1/list-cross-source-signals, or move to the new WM MCP tools (Tier-1+2, 38 tools as of May 2026) and talk MCP instead of REST.
  • Update docker-compose.yml to use build: directives instead of broken image: refs.
  • Document the real-stack quickstart honestly: needs Docker, two sibling clones, a Zep Cloud key, and an LLM key. ~10 minutes of setup, not 2.

Files that don't need changes for v2: watchlist.js, notifier.js, llm.js, summarizer.js, index.js (interfaces stay the same).

Open questions for the WM and MF communities (drafted in docs/community/):

  • WM: which endpoint(s) to aggregate, or jump straight to MCP?
  • MF: should progress WS expose the 7 internal stages, or stay abstracted? Any path to run MF without Zep?

v3 — Stretch ideas

Ranked by impact-per-effort once v2 lands:

Idea Effort Impact
Prediction accuracy dashboard (hit rate over time, per category) medium high — gives the project a reason to come back daily
Scheduled predictions (auto-predict top N events each morning) small medium
Email or Telegram notification alternative to browser push small medium — browser push is unreliable in practice
Multi-event simulation (seed MF with 2-3 simultaneous events) medium medium
Exportable report (PDF / Markdown) small low
What-if follow-up chat (perturbation question in the same sim context) medium-large unknown — depends on MF's design intent

Done

  • v1: mock-mode end-to-end loop, animated demo, full test suite, community standards (LICENSE, CODE_OF_CONDUCT, CONTRIBUTING, issue + PR templates), discovery topics on GitHub, draft posts for the two upstream communities.