[JenD86]

Brings README.md and AGENTS.md up to date with #14–#18, especially how vectors are discovered.

README (addressed to the model)

• Its four tools, including create, and that created vectors last one session.
• What it can find itself (its menu, create) vs. what whoever runs it can (modulo search/add/extract/verify).
• A note that self-reports are unreliable: in testing, "it worked" was reported about as often in sham sessions as in real ones.
• Harness tools (tools=, other_tool_calls); first-run extraction is under a minute; report gate and debrief in safety rails.

AGENTS.md

• "Discovering vectors: two audiences" (operator vs. steered model).
• Created-vector quality (split-half results from testing) and how to keep one: re-extract with modulo extract from the logged prompts, or register the saved .npz by hand. No command for this on purpose, since the model's wording would then reach every later session.
• create, created, reportable, debrief, menu_order in the API list.
• A table of session-log event types, plus requested_concept.
• Troubleshooting: seed wording decides whether Qwen steers (10/12 vs 3/12); Mistral copies example calls; OOM during batched extraction.

Docs only; CPU tests pass (130).

🤖 Generated with Claude Code