0.2.0

Added

• feat: add sandbox support (docker) (#11)
• feat: add license (#10)
• Docker sandbox runtime for the Claude adapter: agent runs can now
  execute inside a Docker container with a configurable workspace
  mount, additional bind mounts, env forwarding, network/user/shm
  overrides, setup commands, and cleanup policy. A bundled
  in-container bridge installs @anthropic-ai/claude-agent-sdk on the
  fly and streams SDK messages back to the host runner. The bridge's
  host-side command runner supports AbortSignal cancellation so
  early consumer termination (break, iterator.return(), or an
  onMessage handler error) tears the container down promptly.

0.1.0

Added

• feat: add llm as judge abstraction (#4)
• feat: add agent messages formatter util
• Initial release of @metamask/agent-runner
• createAgentRunner() factory with runAgent(), flush(), and shutdown() lifecycle API
• ProviderAdapter interface decoupling the runner from the Claude SDK — built-in createClaudeAdapter() translates raw SDK messages into a discriminated AgentMessage union with typed variants (init, generation, tool_result, result, system, tool_progress, tool_use_summary, rate_limit)
• formatMessage() utility for human-readable agent message output
• LLM-as-a-judge evaluation abstraction (judge()) with structured scoring, prompt-injection defence, and optional Langfuse score posting
• Optional Langfuse/OpenTelemetry telemetry with per-run span trees, reference-counted infrastructure, and sensitive-key redaction
• LiteLLM proxy support via ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY environment variables
• Dual CJS/ESM package output via @ts-bridge/cli

Fixed

• fix: cleanup unused publish docs (#8)