Skip to content

ROADMAP.md

Latest commit

 

History

History
122 lines (91 loc) · 5.51 KB

ROADMAP.md

File metadata and controls

122 lines (91 loc) · 5.51 KB

CWE Roadmap

Current: v0.8.2 — Hardening pass (11 blockers + ~40 should-fixes resolved via 10-agent review).

For historical detail see CHANGELOG.md.

Done

v0.8.2 — Hardening Pass (2026-04-21)

  • 10 parallel review agents across hooks/media/utility-scripts/agents/skills/commands/rules/templates/docs/privacy → 11 blockers + ~40 should-fixes
  • Security: PowerShell injection closed in screenshot.py; env-var leak stopped in load_keys(); safety-gate no longer blocks git on self-errors
  • commit-format.sh regex rewritten (was effectively broken); qa-merge glob bugs fixed; port-collision SearXNG/Stirling resolved
  • 10 agents de-coupled from pre-init @workflow/ includes; 16 skills gained allowed-tools; rules neutralised + indexed correctly
  • USER-GUIDE v0.7.0 → v0.8.2 with Media/External-Services/Multi-Terminal sections; templates' {{PLACEHOLDERS}} converted to natural-language hints

v0.8.1 — Polish & Documentation (2026-04-21)

  • /cwe:pdf configurable via Stirling PDF
  • /cwe:screenshot flipped to Python script (cross-platform clipboard)
  • /cwe:init expanded with 5 optional service setup steps
  • Documentation refresh (README badges, SVG header, CHANGELOG)

v0.8.0 — Media + Content Tools (2026-04-21)

  • 6 Media Generation skills (image/video/faceswap/headswap/upscale/motion)
  • Shared _media_lib.py for OpenRouter + MagicHour APIs
  • /cwe:transcript (renamed, supports any URL via TScribe)
  • Intent-router extended for media/content commands

v0.7.1 — Multi-OS Portability (2026-04-20)

  • macOS grep fixes (grep -oP → grep -oE)
  • Version lookup no longer hardcoded
  • handoff-sync.py throttling + systemMessage output

v0.7.0 — Multi-Terminal Parallel Development (2026-03-11)

  • 5 new commands (autopilot, coordinate, check-handoff, handoff, qa-merge)
  • multi-terminal skill with handoff protocol
  • Branch-detection guards in MT hooks
  • Hook events modernized (SessionEnd, SubagentStart)

v0.6.x — Hybrid Delegation + URL Scraper

  • Intent-router hook for keyword-based agent routing
  • URL auto-scraper (Firecrawl → trafilatura → curl fallback)
  • Web-research skill

v0.5.x — Statusline + Polish

  • Python statusline with context/cost/time
  • Currency config via /cwe:init
  • Delegator skill for multi-agent requests

v0.4.x — Native Alignment (10 phases)

  • CLAUDE.md slim-down, .claude/rules/ migration
  • Memory system v2 (daily logs, project-scoped ideas)
  • Spec folders + Shape-Spec interview
  • Safety gate, git standards, health dashboard
  • See CHANGELOG for detail

Next

v0.9.0 — Qdrant Memory + Skills Flip (in design)

Replace Markdown-based memory with Qdrant vector store, flip deterministic skills to Python scripts.

  • Single Qdrant collection cwe_memory with scope/project/type payload filters
  • nomic-embed-text embeddings via Ollama (on-prem, user-configurable)
  • New commands: /cwe:remember, /cwe:recall, /cwe:migmem (import/export/--all)
  • Layer 1: PostToolUse passive sync (auto-memory → Qdrant)
  • Layer 2: SessionStart active recall (inject user + project memory)
  • Layer 3: UserPromptSubmit contextual recall (semantic match on user prompt)
  • /cwe:init offers Fat (Opus/Sonnet, full) vs Lean (Ollama/Haiku, script-heavy) modes
  • Skills-Flip: safety-gate, git-standards, quality-gates, health-dashboard → Python scripts

Design + plan docs in docs/plans/ (local, not tracked).

v0.9.1+ — Qdrant Memory Layers (follow-up)

  • Layer 4: Passive learning (Stop-hook detects decisions/patterns/feedback)
  • Layer 5: Temporal decay (Qdrant payload last_used, recency-weighted recall)
  • Layer 6: Cross-project learning (post-commit hook extracts library/pattern/fix as user-scoped)

v1.0.0 — Production Ready (target)

  • Full Skills-Flip complete (all deterministic skills in Python)
  • Qdrant memory stable + backed up
  • Documentation complete (API, ARCHITECTURE, USER-GUIDE refreshed)
  • CI tests for hooks + scripts
  • Plugin-marketplace-ready

Under Evaluation (no committed version)

Claude Code native features

Recent Claude Code releases added features that may replace CWE components:

  • Output Styles (/output-styles.md): Could simplify /cwe:researcher, /cwe:newsroom
  • Checkpointing: File + conversation rollback — could reduce Serena memory reliance
  • Routines: Anthropic-hosted cron — could replace /cwe:autopilot for long-running tasks
  • Channels: Webhook/Telegram/Discord → session push — could power /cwe:newsroom event flow
  • Sandboxing: Isolated bash — candidate for security/devops agent commands
  • Auto Mode: Could replace --dangerously-skip-permissions alias

Each will be evaluated individually; CWE adopts only what adds value beyond native behavior.

Community / Ecosystem

  • Plugin marketplace publishing (requires 1.0.0)
  • Community template profiles (api, pwa, fullstack)
  • Monorepo support (per-package workflow/)

Principles

# Principle What it means for the roadmap
1 Agent-First Keep delegation via specialized agents — never add work to main context
2 Scripts over Prose Deterministic work lives in Python scripts, LLM-reasoning in Markdown
3 Configurable, not Hardcoded Every external service asked for at /cwe:init, saved in gitignored config
4 Native First If Claude Code adds a feature that replaces a CWE component, deprecate the CWE version
5 SemVer Discipline Patch = fixes, Minor = new features, Major = breaking. 1.0.0 is earned, not scheduled