story-kit: create 343_refactor_abstract_agent_runtime_to_support_non_claude_code_backends

.story_kit/work/1_backlog/343_refactor_abstract_agent_runtime_to_support_non_claude_code_backends.md

@@ -0,0 +1,40 @@
+---
+name: "Abstract agent runtime to support non-Claude-Code backends"
+---
+
+# Refactor 343: Abstract agent runtime to support non-Claude-Code backends
+
+## Current State
+
+- TBD
+
+## Desired State
+
+Currently agent spawning is tightly coupled to Claude Code CLI — agents are spawned as PTY processes running the `claude` binary. To support ChatGPT and Gemini as agent backends, we need to abstract the agent runtime.
+
+The agent pool currently does:
+1. Spawn `claude` CLI process via portable-pty
+2. Stream JSON events from stdout
+3. Parse tool calls, text output, thinking traces
+4. Wait for process exit, run gates
+
+This needs to become a trait so different backends can be plugged in:
+- Claude Code (existing) — spawns `claude` CLI, parses JSON stream
+- OpenAI API — calls ChatGPT via API with tool definitions, manages conversation loop
+- Gemini API — calls Gemini via API with tool definitions, manages conversation loop
+
+The key abstraction is: an agent runtime takes a prompt + tools and produces a stream of events (text output, tool calls, completion). The existing PTY/Claude Code logic becomes one implementation of this trait.
+
+## Acceptance Criteria
+
+- [ ] Define an AgentRuntime trait with methods for: start, stream_events, stop, get_status
+- [ ] ClaudeCodeRuntime implements the trait using existing PTY spawning logic
+- [ ] Agent pool uses the trait instead of directly spawning Claude Code
+- [ ] Runtime selection is configurable per agent in project.toml (e.g. runtime = 'claude-code')
+- [ ] All existing Claude Code agent functionality preserved
+- [ ] Event stream format is runtime-agnostic (text, tool_call, thinking, done)
+- [ ] Token usage tracking works across runtimes
+
+## Out of Scope
+
+- TBD