v0.2.0

Added

• Base adapters for harnesses whose tools are user-defined: openai,
  anthropic, langchain, and otel — bringing the total to seven. Because
  these tools are named by the developer, a shared core (src/agentlog.mjs)
  infers each tool's action and capability from its name and arguments rather
  than relying on a fixed map.
• A generic agent profile for adapters whose capability set isn't known
  ahead of time.

Changed

• Genericized the July 2026 incident framing across the shipped source, the
  ExploitGym corpus, and the README — no third-party attribution. Detector
  scoring is unchanged (the escape still reaches HALT at the same step).