Decouple voice subsystem from Claude Code conversation files

[jpelaez-23blocks]

Problem

The Cerebellum voice subsystem hardcodes ~/.claude/projects/ to read recent conversation context for generating voice commentary. This blocks voice for any non-Claude agent.

Hardcoded location

File	Line	Usage
lib/cerebellum/voice-subsystem.ts	~256	readRecentConversation() reads raw JSONL from ~/.claude/projects/

Current behavior

readRecentConversation() directly reads Claude Code .jsonl files from disk to build context for the LLM-powered voice commentary. This means:

• OpenClaw agents get no voice commentary (no Claude files to read)
• Voice subsystem duplicates work already done by the memory subsystem
• Tight coupling between voice and Claude Code's internal file format

Proposal

Replace direct JSONL reading with one of:

1. CozoDB indexed data — already available via MemorySubsystem. The conversation entries are already indexed and searchable. Voice subsystem should query the existing database instead of re-reading raw files.

2. ConversationSource interface — from Decouple conversation indexing from Claude Code file format #236. Once conversation indexing is abstracted, voice can use the same interface.

Option 1 is simpler and doesn't depend on #236.

Impact

• Unblocks voice commentary for OpenClaw agents
• Removes duplicate file reading (voice + memory both read same files)
• Simplifies voice subsystem (query DB vs. parse raw JSONL)

Related

• Depends conceptually on Decouple conversation indexing from Claude Code file format #236 (ConversationSource interface)
• Part of OpenClaw first-class citizen effort (onnxruntime-node v1.24.1+ fails on x64 macOS (missing darwin/x64 binary) #233)

Labels

enhancement, cerebellum, voice