Agentic Evaluation Baselines

Stage 00 of the Version 2 initiative adds repository evaluation utilities for chat, large tool observations, research/source retention, and a mock external workflow with cancellation and checkpoint continuation. The offline example exercises the managed Agent loop through strict replay and checks actual tool evidence rather than accepting plausible final prose.

From a source checkout with development dependencies installed:

bash

npm run example:agentic-baseline
npm run example:agentic-baseline -- --held-out
npm run build
npm run test:agentic:compat

The evaluation guide documents expected reports, fixture extension, provider setup, and the opt-in suite that calls actual LLMs using .env credentials. Its recorded live pilot validates the OpenAI, Anthropic, and Google paths in two separately timestamped batches. A single sample per model/workload/split does not establish comparative quality, latency, or cost savings.

These utilities require the repository; they are not new package exports. Existing applications need no configuration changes. Agent defaults, algorithmic compaction, the shared ToolManager, tool payloads, and saved-session format are unchanged. See the Version 2 Guide for published-beta availability, historical validation, and deferred stages.