Skip to content

ROADMAP.md

Latest commit

 

History

History
58 lines (46 loc) · 3.43 KB

ROADMAP.md

File metadata and controls

58 lines (46 loc) · 3.43 KB
Raw

Roadmap

Planned work across the template collection. This file tracks what's coming; each template's README.md describes what it does today. Keep the two in sync — when a feature ships, move it from here into the template and its README. If it's still a scaffold, it belongs here, not in present-tense marketing copy.

AI

Several templates lead with AI in their pitch. The honest current state is mixed, so it's tracked here as a plan rather than implied as delivered.

Current state

Template AI surface Status today
contracts extract_terms — AI metadata extraction Real — calls api.openai.com directly (extract_terms.action.ts). Dark without an OPENAI_API_KEY.
content draft_outline_from_topic, suggest_cta, summarize_competitor_signal Real — same raw OpenAI calls. Dark without a key.
helpdesk ticket triage (ai_summary / ai_category / ai_sentiment / ai_suggested_reply / ai_suggested_kb_ids) Scaffold — deterministic baseline (ai_summary = description.slice(0, 280)); ai_triage_on_create.flow.ts is an "insert your LLM here" stub.
project risk / delay / budget forecasting (ai_* fields) Scaffold — the flow prediction node is labelled STUB; seed values are illustrative, not computed.
todo, expense, hr, compliance, procurement — No AI; deterministic by design.

Plan

  1. Route AI through the platform model registry, not raw api.openai.com. Replace the direct fetch() calls (contracts, content) with ObjectStack's AI / model-registry services (see the objectstack-ai skill). Payoff: provider-agnostic (OpenAI, Anthropic / Claude, Bedrock, local), central key management, and it actually demonstrates the "AI-native" platform instead of bypassing it.

  2. Degrade honestly. With no model configured, AI fields/actions should show an explicit "configure AI to enable" state — never a silent stub output (helpdesk's substring) or a failing network call (contracts/content without a key). A user must never mistake a placeholder for a prediction.

  3. Pick one AI flagship and finish it. helpdesk is the best candidate: real triage → ai_summary / ai_category / ai_sentiment / ai_suggested_reply, plus embedding-based KB recall for ai_suggested_kb_ids, shipped end-to-end through (1) and (2). Let the others stay honest deterministic starters until each earns the same treatment.

  4. Stop shipping seeded numbers that look computed. project's ai_completion_probability etc. are hand-authored demo values. Once (3)'s pattern exists, either compute them or keep them clearly marked as samples (the field group is now labelled "AI Predictions (scaffolded)").

Non-AI

  • project — Gantt rendering (schema ready, UI pending platform); resolve the scope overlap with todo (task-level vs portfolio-level).
  • content — trim toward the starter charter; at 9 objects / ~5,200 src LOC it is over the ≤6 objects / ≤2,500 LOC budget in TEMPLATE_GUIDE.md.
  • Marketplace category facets — labels render inconsistently (Title Case for some categories, lowercase for others). The fix belongs in the marketplace UI's category → label map, not in the template manifests (which all use lowercase slugs).