feat: research Amp SDK API surface and map to agent abstraction

[rubenmarcus]

Summary

Research the Amp SDK (by Sourcegraph) API surface and map it to ralph-starter's existing agent abstraction layer.

Linear: ENG-1461

What to Research

1. Amp CLI invocation: How to spawn Amp from CLI, what parameters it accepts
2. Output format: How Amp streams output (JSON stream, text, etc.)
3. Tool-use protocol: How Amp's native tool invocation works
4. Agent modes: Amp has 3 modes — smart (frontier), rush (fast), deep (extended reasoning)
5. Sub-agent support: How Amp handles sub-agents for specialized tasks
6. Prompt caching: Amp's native caching capabilities
7. SDK package: @sourcegraph/amp

Current Agent Abstraction

In src/loop/agents.ts:

• AgentType union: claude-code | cursor | codex | opencode | unknown
• AGENTS config: detection command, run command, args
• runAgent(): spawns subprocess, captures output
• Amp is listed in AGENTS.md but NOT in the code

Deliverable

Document with API mapping: Amp parameters → ralph-starter agent interface fields.

References

• Amp official
• Amp SDK Manual
• How to Build an Agent with Amp