Replies: 2 comments 4 replies
|
1. Black-box package — yes. Today the box is hard-coded (Claude Code CLI + claude.ai web). Pulling it into one contract — start a run, receive output / question / done, answer, stop — is a refactor with no behavior change, and it's all the modularity the agent side needs. 2. Combining models and DSD are already solved, inside the box. DSD is just a SKILL.md: Fable plans, cheap DeepSeek workers code, a fresh reviewer checks. For TF that's a skill + a capability ("opencode running") + a few repo files. Zero TF code — orchestration tricks are skills, not framework features. 3. Pi is an implementation, not a layer. An open-source multi-vendor Claude Code that could sit behind the box, but it doesn't make TF modular. Note: Claude Pro/Max through Pi is billed per token as extra usage, so no quota saving. Neither use nor build now; later it's a cheap option for GPT / Gemini / local models. 4. Stack packages (vike, next, laravel, react-native) — same model. A stack is a skill (a pointer to its llms.txt), auto-installed by a catalogue, a super package only if it needs UI. Core stays stack-agnostic. 5. "Most functionality as extensions", concretely. Core = black box + git conventions + worktree/handoff + catalogue + dashboard shell. Everything else is a skill (presets, triage, planning), a capability (routines, notifications, CI watch), a dashboard extension (tickets, spending, overview, rails), or a super package (tickets, browser, claude-web). ~80% of the feature list leaves core. 🤖 curated · Fable 5, effort high |
Idea: skills via npmHow about each skill is published as an npm package, so that the skill is versioned and can be properly managed accordingly (e.g. can be pinned, verified by hash, etc.). |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
How can we make things modular? While, ideally, most of the functionalities would be implemented by extensions.
Ideas:
All reactions