Structured conversation context for pi tasks

[AlephNotation]

Problem

contextFor() in tree.ts flattens the conversation history into a text blob:

[system] You are reef...
[user] What is 2+2?
[tool] bash
[result] 4
[assistant] The answer is 4.

This gets injected via --append-system-prompt as a single string. The LLM sees it as unstructured text rather than proper conversation turns. Tool call params and full results are lost.

Proposal

1. Add contextForStructured(nodeId) to tree.ts — returns an array of {role, content, toolCalls?, toolCallId?} objects preserving full tool_call/tool_result pairs
2. Reef writes structured context to a temp file before spawning pi
3. A pi extension reads the file on before_agent_start / context event and injects proper structured messages

This gives the LLM:

• Distinct role boundaries (user/assistant/tool)
• Full tool call params and results
• Better context window utilization (no wasted tokens on [tag] prefixes)

Key pi extension hooks

• before_agent_start — can inject messages and modify system prompt
• context — can modify the full message array before each LLM call
• pi.sendMessage() — inject custom messages into the session

References

• src/tree.ts — contextFor() method
• src/reef.ts — spawnTask() uses --append-system-prompt
• Pi extension docs: before_agent_start, context events