Skip to content

PRODUCT_VISION.md

Latest commit

 

History

History
48 lines (40 loc) · 2.29 KB

PRODUCT_VISION.md

File metadata and controls

48 lines (40 loc) · 2.29 KB
Raw

Product Vision: collaborate CLI

Goal: Transform the experimental protocols in this repository into a reusable, standalone CLI tool. User Journey: A developer runs collaborate --prompt "..." in any git repository, and a team of AI agents (Gemini, Claude, Codex) autonomously iterates to achieve the goal.

The Core Concept

The collaborate tool acts as the Orchestrator. It replaces the manual "Human in the Loop" from Phase 0.

User Experience

$ cd my-new-project
$ collaborate init
> Initialized .collab/ directory (events, tasks, reviews).

$ collaborate --prompt "Create a Python web scraper for news.ycombinator.com"
> [Orchestrator] Created Task T-001: "Create Scraper"
> [Gemini] Claimed T-001.
> [Gemini] Wrote src/scraper.py.
> [Gemini] Completed T-001. Requesting Review from Claude.
> [Claude] Reviewing T-001...
> [Claude] Review Failed: Missing requirements.txt.
> [Gemini] Re-claiming T-001. Fixing issues...
> ...
> [System] Goal Achieved.

Architecture Shift

We are moving from Project-Internal Scripts to a External Tool.

Feature Current Prototype (Phase 0) Target Product (CLI)
Orchestration Human runs collaborate status, prompts agents manually CLI runs the loop and invokes backends via API
State Storage .collab/events.jsonl .collab/events.jsonl (hidden dir)
Agent Identity Simulated via export AGENT_NAME Real API Clients (Vertex AI, Anthropic, OpenAI)
Protocol AGENTS.md (Human read) Hardcoded into the CLI logic

Roadmap

  1. Design: Define the CLI structure (Python/Go?). [DONE]
  2. Refactor: Move logic into a proper library structure. [DONE]
  3. API Integration: Connect real LLM backends (Anthropic, OpenAI, Google). [DONE]
  4. The Loop: Implement the main event loop. [DONE]
  5. Autonomous Mode: Implement continuous execution and next-task proposal. [DONE]

Next Steps (Phase 2)

  • Parallel Sub-tasks: Allowing multiple agents to work on independent parts of a project simultaneously.
  • Git Integration: Native branch-based collaboration (Option C from orchestrator design).
  • Streaming Output: Providing real-time visibility into the agent's "thought" process.
  • Context Pruning: Smarter management of project context for large codebases.