Agent Platform MVP

This directory contains a Go MVP implementation of the Agent Platform design derived from the qodercli analysis.

Implemented

• Cobra-based CLI entry points
• Runtime loop with provider abstraction
• OpenAI-compatible chat completions provider with function tool-calling
• Typed tool registry with permission checks
• Built-in tools: read_file, write_file, bash, ask_user
• Project/global resource loading for agents and skills
• Session persistence under .agent-platform/sessions
• Session memory summaries under .agent-platform/sessions/<id>/session_memory.md
• MCP server registry plus mcp-go tool bridge
• Local job registry
• Spec engine that emits staged workflow artifacts aligned to the qodercli-style process
• Interactive permission approval flow

Commands

go run .
go run . start -p "read README.md"
go run . start -p 'write notes.txt::hello world' --yes
go run . start -p 'bash pwd' --yes
go run . start -p 'Summarize README.md' --agent reviewer
go run . status
go run . jobs
go run . mcp add demo http://localhost:8000/sse -t sse
go run . mcp list
go run . spec -p 'Design a tool-calling CLI agent'
go run . spec -p 'Design a tool-calling CLI agent' --provider local --review-rounds 2
go run . spec -p 'Design a tool-calling CLI agent' --provider local --checkpoint-mode stop

Spec Workflow

Generated specs are stored under .agent-platform/specs/<spec-id> and include:

• 01-requirement.md
• 01-verify-requirement.md
• 02-high-level-design.md
• 03-high-level-design-review-round-N.md
• 04-low-level-design/overview.md
• 04-low-level-design/task-NN-*.md
• 05-low-level-design-review-round-N.md
• checkpoint-01-requirements-to-design.json
• checkpoint-02-hld-review-to-lld.json
• checkpoint-03-lld-review-to-implementation.json
• spec-state.json

In auto checkpoint mode the engine produces the full artifact set and validates the result before returning. In stop mode it stops after the first checkpoint so a human can review requirements before design continues.

OpenAI-Compatible Provider

Set these environment variables before using a remote model:

export AGENT_PLATFORM_OPENAI_BASE_URL=http://127.0.0.1:8000/v1
export AGENT_PLATFORM_OPENAI_API_KEY=your-api-key
export AGENT_PLATFORM_OPENAI_MODEL=glm-4.7

If an agent model is not local, the CLI will default to the openai provider automatically.

Resource Directories

Global:

• ~/.agent-platform/agents
• ~/.agent-platform/skills

Project:

• .agent-platform/agents
• .agent-platform/skills

Example Agent

name: builder
description: Local build agent
model: local
tools:
  - read_file
  - write_file
  - bash
systemPrompt: You are a local build agent.

Remote model example:

name: reviewer
description: Remote review agent
model: glm-4.7
tools:
  - read_file
systemPrompt: You are a precise code review agent.