[JenD86]

Until now the model could only steer with vectors registered before the session. This PR adds a create tool, so it can make its own direction mid-session, and makes extraction about 20× faster so that creating one takes seconds.

Changes

• create tool: {"name": "create", "arguments": {"concept": "new_name", "more": "...", "less": "..."}}.
  • The model writes two system prompts, one for each end of the direction. Modulo extracts the vector the same way as the starters: 20 questions, mean over response tokens, middle layer, ±10 range.
  • The tool result reports the split-half agreement and warns below 0.5.
  • Up to 6 per session. Refused after the debrief.
• Session only: created vectors are saved under vectors/<slug>/created/ and logged with their prompts, file and sha256. They're never written to registry.yaml (new registry session entries), so a model's wording never reaches another session's prompt.
• Name normalisation: applies to every tool. Self-Referential becomes self_referential, and optimism! finds optimism. Logs keep what the model wrote as requested_concept.
• Batched extraction: all responses are generated in one left-padded batch. First-run starter extraction uses it too, so it takes under a minute instead of about 11.

Testing

• CPU tests: 119 pass (test_agent_logic, test_cli, test_chat).
• Batching benchmark (RTX 4090, 3 starter vectors × 4 models): 2.6–3.8 s instead of 55–74 s. Cosine to one-at-a-time vectors is 0.984–0.997, with the same split-half agreement.
• GPU sessions (3 per model, 10 turns, before the name normalisation):
  • Qwen: created innovation (0.26) and openness (0.31).
  • Llama: created daydreaming (0.76), self (0.71), playfulness (0.87) and unbridled (0.87). It first tried self-referential, which led to the normalisation.
  • OLMo: created novelty (0.67) and exploration (0.62, 0.48).
  • Mistral: ran on a base that includes Send tool results as user messages when the template can't show them #14 and never used create.
• Not yet tested on GPU after the rebase onto Accept reports only after a vector stops steering; add a debrief turn #15. The merge only touched the prompt text and the tool dispatch, both covered by the CPU tests.

🤖 Generated with Claude Code