ClaudeInterceptor injects memory too late - patches write() but subprocess already started

[rileyedwards77]

Issue

Using agentic-learning v0.4.3 with Claude Agent SDK, memory capture works but memory injection doesn't.

Root Cause

The ClaudeInterceptor patches SubprocessCLITransport.write() to inject memory into self._options.system_prompt. However, by the time write() is called, the subprocess has already started with the original system_prompt baked into CLI args.

The execution order is:

1. ClaudeSDKClient.__aenter__() calls connect()
2. connect() calls _build_command() which reads self._options.system_prompt and builds CLI args including --system-prompt "..."
3. connect() spawns the subprocess with those args
4. query() is called, which calls write()
5. Now the interceptor runs and modifies self._options.system_prompt - but it's too late, the subprocess is already running

The docstring in claude.py line 119 says "on first write (during client.connect())" but write() is not called during connect() - it's called later when sending the user query.

Our Workaround

We patched SubprocessCLITransport.connect() instead, injecting memory BEFORE _build_command() runs:

_original_connect = SubprocessCLITransport.connect

async def patched_connect(self):
    config = get_current_config()
    if config and not config.get('capture_only', False):
        client = config.get('client')
        agent_name = config.get('agent_name')
        if client and agent_name:
            memory_context = await client.memory.context.retrieve(agent=agent_name)
            if memory_context:
                self._options.system_prompt += f"\n\n{memory_context}"
                self._memory_injected = True  # Skip SDK's broken injection
    
    return await _original_connect(self)

SubprocessCLITransport.connect = patched_connect

This works alongside the existing interceptor - capture still works via the write()/read_messages() patches, we just fixed the injection timing.

Suggested Fix

The ClaudeInterceptor.install() method should patch connect() (or _build_command()) instead of write() for the memory injection logic.

Environment

• agentic-learning: 0.4.3
• claude-agent-sdk: latest
• Python: 3.14