feat(tools): agent_commit — version control for AI reasoning

[benclawbot]

Summary

Add a new tool agent_commit that records every LLM turn + tool call as a content-addressed commit with a stable, queryable ID — like git commits, but for agent cognition.

What it does

Operations: commit / log / diff / branch / reference / get

• Content hashing: sha256 of canonical (messages + tool_calls + tool_results + reasoning + metadata)
• Auto-summary: derived from last message content or dominant tool calls
• Thread-safe SQLite backend at ~/.hermes/agent_commit.db, WAL mode
• Full parent chain for reasoning traceability

Motivation

Multi-agent workflows pay massive redundant reasoning costs — Agent B re-derives context Agent A already figured out. This gives agents a shared, queryable reasoning ledger they can reference, branch, and diff.

Files

• tools/agent_commit_tool.py — implementation
• tests/tools/test_agent_commit_tool.py — 23 tests, all passing
• plans/agent-commit.md — full spec